Introduction to Tokamak Plasma Control*

Michael L. Walker, Peter De Vries, Federico Felici, and Eugenio Schuster

Abstract— This paper provides an introduction to the prob- now, magnetic fusion research devices have not been capable
lems of control of plasmas and plasma magnetic-confinement of hosting a plasma with the number and frequency of fusion
devices known as tokamaks. The basic science of fusion plasmas reactions sufficient to produce more output power than is
and objectives of plasma magnetic-confinement technologies
are described. In addition to a general overview of plasma consumed in confining and controlling the plasma. This is
control problems, more extensive discussions of three specific about to change with anticipated completion of the ITER
classes of problems - control of plasma magneto-hydrodynamic tokamak currently under construction in southern France
behavior, control of plasma parameter internal distributions, [1]. Early operation of ITER will focus on learning how
and methods for handling system faults or unexpected loss of to produce and control plasmas that are far more energetic
control - are provided.
than in any existing magnetic-confinement device. The initial
I. INTRODUCTION plasma control system is being designed now, including both
This paper provides an introduction to the problems of the software architecture and the algorithms that will be used
plasma control in tokamaks – plasma-confining devices for control during ITER first plasma operation starting in
used in magnetic-confinement fusion research. Significant approximately 2025.
progress has been made in the decades since controlled II. MAGNETIC-CONFINEMENT FUSION
magnetic fusion was first envisioned as a potential power A. Objectives of Fusion
source, most focused on achieving necessary scientific under-
The fusion process1 proposed for power generation com-
standing of fusion plasmas and how best to produce energy-
bines two hydrogen isotopes - deuterium (D) and tritium (T)
generating fusion reactions within those plasmas. As greater
- to produce a helium nucleus (also known as an α particle),
scientific understanding was gained, more attention began to
a neutron, and thermal energy. In this fusion reaction (and
be paid to technological issues associated with confining and
any nuclear reaction, including fission), a small amount of
controlling these energy-producing reactions. Initial active
mass is converted into energy. Twenty percent of this energy
control approaches consisted primarily of a small number
is thermal - associated with the charged products that remain
of SISO PID controllers. More recently, as the number
in the plasma - and goes to support the continued reaction.
of plasma parameters to be controlled has increased, more
The remaining 80% is in the free neutron that is captured by
sophisticated controllers have been designed, implemented,
the walls of the fusion reactor and its energy is converted to
and tested on a number of experimental fusion devices. Up to
heat. The heat from captured neutrons can be subsequently
*This material is based upon work supported by the U.S. Depart- converted to electrical energy. The fraction of mass “lost” is
ment of Energy, Office of Science, Office of Fusion Energy Sciences, tiny, just 38 parts out of 10,000. Nevertheless, fusion energy
using the DIII-D National Fusion Facility, a DOE Office of Science user
facility, under Awards DE-FC02-04ER54698, DE-SC0010661, and DE- released from just 1 gram of deuterium-tritium fuel equals
SC0010537.;Disclaimer:This report was prepared as an account of work the energy from about 2400 gallons (7.4 tons) of oil.
sponsored by an agency of the United States Government. Neither the United For the two isotopes to fuse together the nuclei must
States Government nor any agency thereof, nor any of their employees,
makes any warranty, express or implied, or assumes any legal liability collide, but their similar positive electric charges creates a
or responsibility for the accuracy, completeness, or usefulness of any repellent force between them. This force can be overcome
information, apparatus, product, or process disclosed, or represents that by causing them to collide at sufficiently high speed. These
its use would not infringe privately owned rights. Reference herein to any
specific commercial product, process, or service by trade name, trademark, high-speed interactions can be produced by heating the gases
manufacturer, or otherwise does not necessarily constitute or imply its to sufficiently high temperatures; increasing the gas density
endorsement, recommendation, or favoring by the United States Government increases the frequency of particle collisions. Creating the
or any agency thereof. The views and opinions of authors expressed herein
do not necessarily state or reflect those of the United States Government conditions for frequent high-speed collisions is a primary
or any agency thereof.; ITER is a Nuclear Facility INB-174. The views objective of controlled nuclear fusion.
and opinions expressed herein do not necessarily reflect those of the ITER Assuming that such collisions can be produced in a con-
Organization.; This work is supported in part by the Swiss National Science
Foundation.
trollable manner, a fusion power plant would use a concept
M. L. Walker ([email protected]) is with General Atom- similar to a power plant that burns coal or oil. A heat
ics, San Diego, CA 92121, USA. source boils water and produces high-pressure steam, which
P. De Vries ([email protected]) is with ITER Organi-
zation, St Paul Lez Durance, France.
turns a turbine generator to produce electricity. A power-
F. Felici ([email protected]) is with EPFL - Swiss producing fusion reactor would simply use a different heat
Plasma Center, Lausanne, Switzerland.
E. Schuster ([email protected]) is with the Department of 1 This paper expands on concepts and results previously described in
Mechanical Engineering and Mechanics, Lehigh University, Bethlehem, PA [2],[3],[4]]. Some repetition of these concepts is required in the initial
18015, USA. sections to make this work self-contained.
source (fusion). A significant difference from conventional is
that fusion requires millions of times less fuel for the same
amount of steam and electricity. In fact, there is sufficient
fuel for thousands of years. Deuterium can be extracted
from the inexhaustible supply in sea water (1 part/ 6,500
parts H2 O). Tritium can be produced from lithium inside the
reactor itself, where lithium extracted from ocean water is
enough for millions of years [5].
Fusion energy has other advantages. It does not produce
any of the compounds responsible for global warming. (In
fact, there is currently a global shortage of its primary
reaction product, helium.) There is virtually no risk of
nuclear accident, since (in contrast to fission) meltdown is
not possible because an uncontrolled energy increase leads
to self-shutdown of the reaction. A fusion power plant
would produce no high level radioactive waste (e.g., no Fig. 1. Illustration of magnetic confinement in a tokamak.
fission fragments) - only a small volume of activated reactor
components with short lived radioactivity (10 years vs 10 Consequently, immediately after plasma initiation, the ohmic
thousand years for fission). coil current, which is commanded to continue its downward
However, constructing an energy-generating fusion reactor ramp, now operates as the primary side of a transformer
is challenging, since not only must the deuterium and tritium whose secondary is the conductive plasma. This transformer
nuclei be heated to very high temperatures, but the hot action causes current to flow in the plasma by means of
gases must also be confined in a small enough volume for a the opposing flows of oppositely charged particles. This
statistically large number of D-T collisions to occur. This is effect is known as inductive current drive. Collisions of
where magnetic confinement comes in. the electrons and ions cause the plasma to be resistive,
which causes the plasma to heat (thus the origin of the term
B. Tokamaks ohmic heating). The rate of ohmic coil current decrease is
The magnetic confinement approach to constructing a used to control the plasma current I p up to a target flat-
fusion reactor is to ionize neutral hydrogen gas isotopes, top level by about 1 second after plasma initiation. At the
resulting in a plasma. This plasma, consisting of electrically same time, currents driven in poloidal field (PF) coils (blue
charged particles, can be held within a fixed volume through in Fig. 1) confine and shape the plasma, to prevent it from
use of magnetic fields created by large electrical currents. contacting the vessel wall and to produce desirable fusion
The plasma is then heated to sufficiently high temperatures performance characteristics. The combination of toroidal
to induce fusion reactions. field Bφ produced by TF coil currents and poloidal field B p
There are several variations of magnetic confinement, but (orthogonal to Bφ ) produced by PF coil and plasma currents
the concept that is the most mature uses a toroidal device results in a helical magnetic field around the torus (Fig. 3).
called a tokamak. All tokamaks use the same basic concept, Shortly after t = 0, additional gas is puffed into the
illustrated Fig. 1. All tokamaks produce plasma pulses (also chamber to increase density and pressure to desired levels.
referred to as shots or discharges) comprising approximately In most DIII-D discharges, neutral beams (uncharged atoms
the same sequence of events. These events can be illustrated of deuterium) are injected into the plasma at high velocity.
using the DIII-D experiment [6] as a typical example (Fig. 2). These particles collide with particles in the plasma, thereby
Time during the discharge is measured relative to t=0, the converting their momentum into heat and further heating
approximate time at which plasma is initiated. Current in the bulk plasma. When directed toroidally into the plasma,
the toroidal field (TF) coil (brown in Fig. 1) - known as the this momentum can also be used to provide torque for
B-coil at DIII-D - is brought up early to create a constant plasma rotation and to drive plasma current. Various forms of
toroidal magnetic field to confine the plasma when it is radio-frequency (RF) actuators (with action similar to your
initially created. Just prior to t=0, deuterium gas is puffed microwave oven) are also employed to heat and drive current
into the interior of the torus-shaped containment vessel (gray in plasma. One system (the electron-cyclotron (EC) heating
in Fig. 1), and the ohmic heating coil (magenta in Fig. 1) - and current drive system - (ECRH/ECCD) can preferentially
known as the E-coil at DIII-D - is brought to its maximum control the location and direction of heating power and
positive current in preparation for plasma initiation. driven current by use of steerable beam-injection mirrors.
At t = 0, the E-coil current is driven down quickly to NBI and RF actuators both generate what is known as non-
produce a large electric field within the torus. This electric inductive current. Another source of non-inductive current
field accelerates free electrons, which collide with and rip is the so-called bootstrap current, which is self-generated by
apart neutral gas atoms, thereby producing ionized gas or the plasma when a density gradient is present [4].
plasma (red torus in Fig. 1). The plasma consists of charged The separate time intervals in which the plasma current
particles that are free to move, and thus acts as a conductor. is increasing (0 to 1 second), constant (1 to 5 seconds),
 2
Plasma Current
1 Z Magnetic field line
MA 0
Bp P

150
E-Coil
R
kA

B-Coil Bφ
R0
–200
4 Gas Puff
(a.u.)

2 X Magnetic axis
0
Fig. 3. Illustration of (helical) total field B equal to the sum B = B p + Bφ
4
of poloidal and toroidal field vectors, and of flux Ψ at a point P = (R, Z).
Density
1013 cm–3

2
a diverted plasma separatrix strikes the vessel wall are strike
0 points. The location of the separatrix and the X point (or
8 strike points) are controlled (as discussed in Section IV)
Beam Power
4
by the poloidal field generated by currents in the PF coils
(whose cross sections are labeled F1A through F9B here).
0 The nesting of contours of constant flux Ψ in Fig. 4 is pre-
–4000 –2000 0 2000 4000 6000
dicted by ideal MHD theory and confirmed by experimental
Time (ms)
measurement. The complete representation shown cannot be
Fig. 2. Evolution of a plasma in the DIII-D tokamak measured directly, so is produced by an algorithm that fits
an ideal MHD model known as the Grad-Shafranov (GS)
or decreasing (5 to 6 seconds) are often referred to as, equation to measurements from magnetic sensors attached
respectively, the rampup, flattop, and rampdown phases. to the vessel wall or just outside of it. The GS equation is
There are currently several experimental tokamaks around given by
the world [2]. Every tokamak is different, with capabilities ∂p ∂f
designed to support a particular experimental goal. Each is −∆∗ ψ = µ0 R2 +f , (1)
∂ψ ∂ψ
designed to operate with a different maximum plasma current
level (a few hundred kilo-Amps to a few Mega-Amps), where p and f are fitted-parameter functions [8], µ0 is
toroidal field (a few Tesla), and duration of discharges. vacuum magnetic permeability, ψ = Ψ/(2π), and the ∆∗
Nearly all use some form of active plasma control, with operator is
∂ 1 ∂ψ ∂ 2ψ
several actively conducting research to address existing and ∆∗ ψ = R ( )+ 2 . (2)
future plasma control needs. ∂R R R ∂z
Although Fig. 1 illustrates a plasma with a circular cross- There are several types of magnetic sensors, distinguished
section, obtaining high performance generally dictates a non- by whether they measure local magnetic field, magnetic flux
circular cross-section [7]. Fig. 4 illustrates the cross-section through a surface, current in coils driven by power supplies,
of an experimental plasma in the DIII-D tokamak, repre- or current induced in conducting structures or plasma. For
sented by a set of contours representing points of constant most tokamaks, there are a few hundred such measurements.
poloidal flux Ψ. To understand this figure, it is necessary Other sensors make additional measurements that can also be
to understand the notion of flux coordinates. The standard used to constrain the fit by characterizing internal parameters
definition of magnetic flux is the integral
R
of magnetic field such as temperature, pressure, or internal magnetic field. As
normal to a specified surface Ψ = S B · dS. For fusion with magnetic sensors, these internal-parameter sensors must
plasmas, this definition is extended to define flux at a point by necessity be remote from the plasma because of high
P = (R, Z) in a poloidal cross-section as the flux through temperatures, radiation effects, or both.
an imaginary disk (Fig. 3) whose boundary passes through When applied external magnetic field forces (magnetic
(R, Z). Contours in Fig. 4 represent curves of constant flux pressure) balance kinetic pressure within the plasma so that
according to this extended definition. the plasma does not move or deform (Fig. 4), the plasma is in
Flux has largest absolute value at the center of these nested equilibrium. This equilibrium constitutes part of the overall
flux contours at the magnetic axis. The closed flux contour operating point of the controlled system.
farthest from the magnetic axis (also called last closed flux
surface or separatrix) defines the edge of the plasma. In a III. TOKAMAK CONTROL NEEDS
diverted plasma, the separatrix forms an X point, in Fig. 4 at Control needs for tokamaks are driven largely by the long-
the bottom of the plasma. (It is also possible to have a limited term goal of putting energy on the grid. Economic factors
plasma, in which the separatrix is defined as the contour dictate the need for high-fusion-performance plasmas, which
tangent to or “touching” the vessel wall.) The points at which are invariably either unstable or marginally stable. Efforts
 Z F8A F9A
F5A 90291 at 2500 ms

Poloidal field coils

(F coils)
F4A F7A

F3A kinetic magnetic pressure

pressure
F2A
F6A

F1A surfaces of
Magnetic Axis constant flux
R
F1B

F6B
F2B
Plasma boundary
(separatrix)
F3B
ZX
F4B RX F7B
Vessel wall Fig. 5. The ITER tokamak. The cryostat is a large tank filled with
F5B supercritical helium to cool the superconducting coils. Blanket modules
F8B Divertor strike points
protect the vessel walls from the intense heat of the plasma. The central
F9B solenoid (CS) coils are PF coils whose purpose includes acting as a virtual
X-point
ohmic coil. Note man for scale.
Fig. 4. Cross section of the DIII-D tokamak showing a flux contour
representation of the plasma. Energy-producing fusion power plants (and even ITER)
require control performance and robustness comparable to
to reduce plasma instability while maintaining or increas- high performance aircraft (e.g., jet fighters). Similar to high
ing performance lead to two separate but interdependent performance aircraft, fusion plasmas are intrinsically unsta-
approaches - active stabilization by feedback control and ble (although closed loop stable), operate near the edge of
search for system operating points with improved passive technologies’ performance envelope, and yet must operate
stability. Research into operating point modification is gen- with high robustness to disturbances and even system faults.
erally conducted by plasma physicists, since it requires deep However, in contrast to aircraft, the tokamak and plasma are
understanding of plasma physics. However, it also requires completely under autonomous control by a real-time Plasma
active control to gain access to and regulate around a chosen Control System (PCS). Controlling automatically all aspects
operating point. At the same time, minimization of plant of a tokamak discharge is a challenging task. Actuator and
cost (economics again) leads to small control margins, with sensor capabilities are limited and can compromise control
a concomitant increased potential for loss of control. This performance when operating the tokamak at its highest
non-negligible risk implies the need for robust mechanisms performance. The control must also avoid a number of
to detect or predict incipient loss of control and respond in a operational limits related to stability of the plasma itself.
way that minimizes negative impact of that loss of control. Uncontrolled growth of instabilities can lead to deterioration
The previously-mentioned ITER tokamak (Fig. 5) is the of the plasma confinement, yielding a disruptive end of the
near-term target of a significant part of ongoing plasma discharge [9]. During such a discharge disruption the thermal
control research since its operation will depend critically on and magnetic energy is dissipated onto the device in a very
control. ITER already has several of the characteristics of short time (ms to hundreds of ms in ITER), yielding very
envisioned fusion power reactors. Most importantly, it will be high thermal and electro-magnetic loads. Such events should
the first device to demonstrate a self-sustained fusion reaction to be prevented and, if this is not possible, mitigating action
(referred to as burn). It will also operate pulses with long must be taken [10].
duration characteristic of a real reactor. ITER plasmas are Energy content of plasmas produced during the ”first
projected to generate approximately 10 times more output plasma” phase of operation are very small when compared
fusion power than input power (fusion gain Q ≥ 10) when with plasmas planned for later operation phases, which
fully operational. However, ITER is not actually electricity- means that consequences of most possible control failures
producing and will not connect its output power to the are similarly small. However, even during ”first plasma”
electrical grid. Instead, it will be used to understand scientific operation there are certain control failures that can lead to
and technical issues related to sustainment of a burning many millions of dollars of device damage. During later
plasma and, ultimately, to demonstrate feasibility of fusion phases of ITER operation, the consequences of control
as an energy source. ITER is huge (see man at bottom right failure will become more severe at the same time that
of Fig. 5 for scale) and expensive (billions of dollars), so is requirements for control grow more complex. Thus low
being built by a multi-country consortium that includes the disruptivity operation is key to maintaining the integrity of
EU, Japan, USA, China, India, South Korea, and Russia. the tokamak and is fundamentally a plasma control problem.
The PCS must actively regulate the plasma state to remain as self-inductance and R p is the plasma resistance. Ic , I p are the
(passively) stable as possible. Any instabilities that remain coil and plasma currents, respectively.
must be actively controlled and stabilized. In addition to This equation is coupled to the circuit equation for current
continuously-acting control, the PCS must be able to detect in the coil
and respond asynchronously to hardware faults and “off- Lc I˙c + Rc Ic + Mcp I˙p = Vc (4)
normal” plasma conditions [11].
In the following we provide a more detailed discussion Where Lc is the coil self-inductance, Rc is the coil resistivity,
of three categories of plasma control. Axisymmetric control and M pc = Mcp as previously defined.
is discussed in Section IV, magnetic instability control in From equation (3) one can observe that a constant I p >
Section V, control of internal plasma quantities, including 0 requires a time-varying Ic , i.e., I˙c < 0. Indeed, the SISO
fusion burn, is discussed in Section VI. Finally, Section VII transfer function from Vc to I p is written as:
describes the emerging field of off-normal event handling. I p (s) M pc s
= 2 2 (5)
IV. AXISYMMETRIC MAGNETIC CONTROL Vc (s) M pc s − (Lc s + Rc )(L p s + R p )

Axisymmetric magnetic control refers to control of the Since this transfer function features a zero at the origin, it
plasma position and shape in the 2D poloidal plane, and requires at least two integrators in the controller to control
of plasma current. Since the first tokamaks in the 1960s and the plasma current with zero steady-state error.
1970s, this type of control has been extensively studied, since 2) Radial position control: Since the plasma carries an
it is necessary to keep the plasma in the desired position in electrical current, magnetic fields (self-generated or applied
the reactor vessel. Additionally it was soon found essential externally through magnetic coils) result in a Lorentz force
to avoid significant contact with the (cold) reactor walls in on the plasma, following the equation fL = j × B, where fL is
order to reach the high temperatures desired for fusion. This the force density, j the plasma current density, B the magnetic
is to keep the plasma pure, mostly containing light nuclei field vector. This force has to balance the pressure gradient
(Deuterium and Tritium), and avoiding contamination with in the plasma everywhere in space, a condition known as the
heavier elements that may be present in the material of the MHD force balance.
wall. This field of research is relatively mature, and several The main effect of the plasma pressure is to push the
tutorial papers and books have been published, e.g. [3], [12]. plasma radially outward, in a similar fashion as the outward
Here, we present a brief tutorial introduction to the problem force on a pressurized tire. Another outward force, known
and refer to references for more detail. as ‘hoop force’, occurs due to compression of the magnetic
field on the inside part of the torus. To counteract this force,
A. Plasma current and position control a vertical magnetic field is applied by a set of ‘vertical
As explained in the introduction, a plasma equilibrium field coils’, which, in combination with the toroidal plasma
features a toroidal electrical current (I p ) in the conducting current, generates an inward radial force that compensates the
plasma. Since many properties of the plasma (stability, plasmas outward radial force (see left half of Fig. 6). Since
confinement) are strongly affected by the value of the plasma the required radial force will vary depending on the plasma
current, it is typically regulated. It is ramped from zero (at pressure, another feedback loop known as radial position
the beginning of the discharge), maintained at a desired set- control is used to keep the plasma at the desired radial
point value for the duration of the plasma, then ramped location. A simple model for the dynamics of the radial
back down to a low value before the plasma is extinguished. position can be written as:
Additionally, the position of the plasma in the (R, Z) plane
d 2 R µ0 I p2
must be controlled to avoid wall contact. Together, the mp = Γ(R, β p , li ) + 2πRI p BZ (R, Z, Ir , Iv ), (6)
control of (R, Z, I p ) requires a basic set of 3 feedback loops dt 2 2
that any tokamak must have in order to function effectively. Where m p is the plasma mass, R is the plasma radial position
1) Plasma current control: In most cases, the plasma (Fig. 3), Γ is a factor that depends on internal quantities of
current is driven by induction (Section II-B). The so-called the plasma (mainly the normalized pressure β p = B2<p> ,
p /(2µ0 )
central solenoid (Fig. 5), a set of magnetic coils close to the where <> indicates an average over the plasma volume,
vertical axis of the device, is used for this purpose, as it is and the normalized internal inductance per unit length `i =
Li R B2
designed to change the poloidal magnetic flux (which drives ( 2πR µ0
)/( 4π ), where Li = 2 P 2µp0 dr/I p2 and P is the plasma
0
current inductively), while having a low magnetic field in volume), and Bz is vertical field generating the counteracting
the plasma region (which would affect the plasma position inward Lorenz force [8]. Since the plasma mass is very small,
- see Radial Position Control, below). A simplified equation one can consider the limit m p = 0 and assume the plasma
for the dynamics of the plasma current, in case of a single instantaneously satisfies this force balance equation, with the
solenoid coil, can be written as: position determined by the (slower) time-evolution of the
0 = M pc I˙c + L p I˙p + R p I p , (3) two forcing terms on the right-hand-side. Bz depends on the
spatial position of the plasma, as well as on the current in
Where M pc is the mutual inductance between the plasma the radial position control coils Ir and any other currents in
current distribution and central solenoid coil, L p is the plasma the surrounding vessel structure (Iv ).
 to the moving plasma current column). The growth rate of
resulting instability can be rather fast, with a real pole up to
B FL
1000s−1 for some tokamaks. The instability can be stabilized
by acting on a combination of coils that generate a radial
magnetic field (as illustrated in right half of Fig. 6), which
FL Fhoop + Ftyre induces a vertical force on the plasma. In many tokamaks
j B with high elongation (hence having high growth rates of the
j
vertical instability), specialized ‘fast coils’ are placed close to
the plasma to react more rapidly to the plasma displacement.
Since there are no net forces on a plasma equilibrium
Z in the vertical direction, a linearized model of the vertical
R position dynamics around an equilibrium can be used [12].
This model must include the coupling to the ’eddy currents’
Fig. 6. Illustration of vertical field used to balance the radial outward
forces of a tokamak plasma (left), and radial field used to stabilize the
induced in the vacuum vessel as a reaction to changing
vertical position (right). Typical coil current arrangements to generate such magnetic fluxes.
fields are shown as well, with top and bottom coils at the same radius on
the right connected in anti-series. Crosses indicate currents pointing into the 2R0 ∂ M pv 2R0 ∂ M pz
δz− δ Iv − δ Iz = 0
page, and points indicate currents pointing out of the page. µ0 I p0 Γn ∂ Z µ0 I p0 Γn ∂ Z
∂ Mvp (7)
@BR @BR Mvv I˙v + Rv Iv + I p0 δ ż + Mvz I˙z = 0
Ip = 0,  = 1, n = 0 Ip < 0,  > 1, n < 0 ∂Z
@z @z
∂ Mvp
Lz I˙z + Rz Iz + I p0 δ ż + Mzv I˙v = Vz
∂Z
FL Here δ indicates the variation of a quantity with respect to
its equilibrium value, n is the curvature index defined as
BR j
n = − BRz00 ∂∂BzR , (with Bz0 the vertical magnetic field generated
j by the coils at the plasma position, and R0 the radial
FL B
coordinate of the center of the vacuum region). Iz is the
z
current in the anti-series coil combination used for vertical
j BR position control, and Iv is a (vector) of vessel currents in the
surrounding structure. Substituting δ z from the first equation
FL into the last two, a model is obtained with two real poles,
of which one is unstable, and an unstable zero. This system
can be stabilized by a PD controller but, as is well-known
Fig. 7. A purely vertical magnetic field yields a circular plasma (κ = 1). from classical control theory, an unstable system with a zero
Adding curvature to the magnetic field results in elongated plasma (κ > 1), imposes limitations on the control performance. Additionally,
which is vertically unstable. FL indicates the Lorentz force produced by the presence of delays will lead to further reduction of the
current density jφ (into the page) and magnetic field B
maximum attainable control performance. [13].
Since the current in the coils is, in steady state, propor- In practice, the (R, Z, I p ) control loops are often analyzed
tional to the applied voltage, a PID controller is sufficient to simultaneously by combining the equations of all the coils,
control the radial position to a desired set-point value. the radial force balance, the vertical force balance and the
3) Vertical position control: The third control loop is the plasma current, as well as measurement equations for all the
so-called vertical stability loop. This is the most complicated magnetic field, flux and current measurements.
position control loop since the open-loop plant is inherently The resulting model, which represents the dynamics of
unstable in the vertical direction. For reasons related to the tokamak modeling the plasma as a rigidly displacing
detailed physics of magnetic confinement, plasmas with conductor, is referred to as the Rigid Plasma Displacement
a vertically elongated shape (such as in Figure 4) have Model. Despite its simplicity, it covers the most important
higher pressure than circular plasmas. To obtain an elongated dynamics required for a design of a position and current
plasma, a curved field must be applied (Figure 7, right). controller. The simplifications in the modeling result in
Intuitively, one can think that this curved field has the effect controller gains that are not directly applicable to the real
of applying an upward force on the top half of the plasma, tokamak. For this reason, PID-type controllers are usually
and a downward force on the lower half of the plasma, designed so that gains of each loop have effects that can
resulting in an elongating effect. Unfortunately, this same easily be interpreted, so that they can be manually tuned
field makes the plasma position vertically unstable: a small as required based on experimental findings. Alternatively,
vertical displacement will increase the vertical force in the more accurate models can be used to directly synthesize
direction of the displacement, leading to an instability. This controllers, as discussed in the next section.
instability is partially counteracted by the eddy currents 4) Plasma shape control: As mentioned before, the shape
(currents induced in surrounding conducting structures due of the plasma in the 2D poloidal plane has a significant
 Shape control loop
effect on the plasma performance. By shape we mean the (1kHz)
Magnetic stabilization & control loop
location of the Last Closed Flux Surface (LCFS) as well as Pre-shot
(10kHz) R,Z,Ip
calculations
estimation
the configuration of the divertor region (see discussion for R,Z,Ip
Magnetic
R,Z,I + - errors
Fig. 4). Since the flux distribution, and hence the location p,ref
+
KRZIP
V
measurements
MIMO a
of the LCFS, depends on the magnetic fields, the shape of Coil current
Coil current
PI TCV
feedforward
errors
Tokamak
the externally applied magnetic fields affect the shape of the Kcoils
Vfast
+ - Kfast(s)
plasma and of the divertor region. Fast δz estimate
Fast stabilization loop (50kHz)
Coil current
A first step for controlling the shape is to be able to Coil current
measurements

directly control the current of any remaining magnetic coil corrections R,Z,Ip corrections
Plasma Shape
Kshape estimation
combinations (after having used three coil combinations for Shape errors

the three control loops for R,Z, and I p ). This is usually done
with another set of PID control loops for each circuit. Then, Fig. 8. Schematic diagram illustrating magnetic control as implemented in
the TCV tokamak [16]. Note that three control loops are used, each running
the plasma shape can be controlled using pre-calculated at different frequencies. A fast loop (yellow) controls the fast internal coil
feedforward references for these coil currents. Indeed, the for vertical position stabilization. A slower set of loops (blue) control the
coil currents required to obtain a desired plasma shape can plasma current, plasma position, and coil currents. The slowest, outer loop
(red) controls the plasma shape and other quantities. Similar hierarchical
be quite accurately determined by solving the MHD force control loops are used on various tokamaks.
balance, coupled to time-varying equations for the coil and
vessel currents. Ad-hoc corrections to these coil currents supply voltages), input delays (due to digital control or power
are often employed in a shot-to-shot, trial-and-error (and supply dynamics), state constraints, including nonlinear ones
time-consuming) fashion to compensate for unmodeled dis- (maximum current, maximum heat dissipated in coils, forces
turbances. To avoid these manual corrections, many existing between coils). To treat these operational limits, more ad-
tokamaks employ feedback control to change the coil current vanced approaches such as MPC [18] have been proposed.
references in order to control plasma shape. This shape Also, there is a wide literature on use of model-based modern
control loop is often an additional, external loop closed multivariable control techniques on the tokamak control
after the (R,Z,Ip) and coil current control loops have been problem, including H-infinity control [19], [20], Kalman
closed, though other approaches also exist. This hierarchy of filtering for plasma state estimation [21], and more. However,
controllers is schematically shown in Fig. 8. perhaps surprisingly, these solutions are not routinely used
Two main approaches exist for shape control: control of for operating real-life tokamaks, to date, where individual
Gaps or Isoflux control. In the first approach, the distance PID controllers (with some decoupling scheme) are the norm
between the plasma LCFS and the first wall is controlled [15], [22], [23], [16], [24]. One principal reason cited by
to a desired value. In the second approach, the differences tokamak operators is the lack of interpretability of results.
between poloidal flux values at various control points on the When the controller does not perform as expected, one needs
boundary are controlled to zero. To determine the control detailed knowledge of control theory/control mathematics to
errors in real-time, one can either extrapolate magnetic resolve the problem. Here, it is important to realize that the
measurements close to the plasma (which may not work well reality of tokamak operations is that control is a tool required
in the divertor region where the magnetic fields are low) to achieve plasma discharges that are then studied (for
or perform a real-time equilibrium reconstruction (solving physics reasons, e.g. physics model validation, investigation
the GS equation (1) in real-time, using constraints from of a physical effect, etc.). Any perceived delay in tokamak
measurements to determine the internal current distribution). operations due to introduction of more sophisticated, and not
To obtain a model for designing a shape controller, the easily interpretable controllers, often results in reverting to
simplest approach is to neglect the plasma dynamics entirely more primitive (e.g. PID) controllers that are easier to tune,
and use the static relation between the currents in the coils even if they take some trial-and-error attempts to obtain the
and the fields to determine the expected change in plasma desired result. At the same time, automated methods to tune
shape to changing currents. A more accurate (dynamical) the control gains for individual loops have proven very useful
model can be obtained by linearizing (1) around an equilib- in practice [25], [23]. For next-generation tokamaks such as
rium (e.g. [14]). Since the number of poloidal field coils may ITER, and future fusion reactors, controller design relying
be different (sometimes smaller) than the number of shape more accurately on models will become more important
errors, a non-square model of the plant is obtained. One then owing to the vastly higher cost of individual discharges, but
takes an SVD of the static gain of the plant to determine the this may not completely supplant the need for some degree
error directions which can most easily be controlled [15], of manual tuning. This will require a new generation of
[16], [17], and design a diagonal PID controlling the errors engineers trained in both control theory and tokamak physics,
projected on the principal singular vectors. Still, there often is to be able to support the use of advanced controllers.
an important component of feedforward control action, either Though position and shape control are relatively mature
computed in real-time or off-line to determine the currents fields, further challenges remain: a recurring issue when
required to obtain a desired shape. designing individual control loops is the effective decoupling
To make these controllers of practical use, they must of the shape and position control loops. In the scheme shown
also take into account input constraints (saturation of power in Figure 8, this is achieved by sending position reference
corrections from the shape controller to the inner position energy from the inner to the outer regions of the confined
control loop, while ensuring that the shape control does not plasma. This has the beneficial effect of removing unwanted
attempt to modify the position or current. impurities from the plasma, in particular the ‘Helium ash’
Another emerging challenge is to adapt, in real-time, the that is the product of fusion reactions. However, these saw-
references for the plasma shape and position to respond to tooth crashes generate a perturbation of the plasma magnetic
changes by supervisory control decisions (refer to exception fields, which may trigger other unwanted instabilities.
handling, section VII). The amplitude and period of the sawtooth crashes can be
controlled in several ways. Highly localized current drive
V. 3D INSTABILITY CONTROL (typically by ECCD) has an effect of stabilizing (leading to
In the previous section we discussed axisymmetric control, more infrequent and larger sawtooth crashes) or destabilizing
i.e., control of the plasma properties averaged toroidally sawteeth (leading to more frequent and smaller crashes),
and represented in the 2D poloidal (R,Z) plane. However, depending on the localization of the driven current. This
there are important 3D effects that affect time-evolution of method of control has been investigated in detail (e.g. [30]),
a tokamak plasma and its confinement qualities, that must where it was shown that applying said current drive has
be taken into account and, in many cases, controlled. While the expected physical effect. Feedback control approaches
a comprehensive review is outside the scope of this paper, have also been demonstrated e.g. [31] where the steering
we give a flavor of the issues here. For further reference, the mirror angles were controlled in feedback to achieve a
reader is referred to [2] for a control-oriented overview and desired sawtooth period. It has also been shown that, since
to [26] for a physics introduction. the sawtooth instability can be described by a simplified
nonlinear model generating a limit cycle, synchronization
A. Error field, RWM control techniques can be applied. By periodically varying the power
Due to small misalignments in the mechanical assembly of the ECCD source, the sawtooth frequency can be made to
of a tokamak, the confining magnetic fields may not be lock to the frequency of the power perturbation [32]. This is
completely axisymmetric. The non-axisymmetric component of practical interest, since it is highly beneficial to be able to
of the field is known as the error field. The toroidal plasma control the sawtooth period precisely in order to regulate the
will have a tendency to deform (in 3 dimensions) aligning mixing between core and peripheral plasma region, as well
itself with this non-symmetric magnetic field. When the as to preempt any negative effects that the sawtooth crash
plasma pressure increases, these deformations will tend to has on the global plasma stability.
grow, in practice limiting the maximum pressure that can Another important class of instability is the so-called
be achieved by the plasma. For this purpose, external non- Neoclassical Tearing Mode (NTM). This is a 3D helical
axisymmetric magnetic coils are used to compensate for deformation of the plasma around certain flux surfaces where
these imperfections and try to restore a perfectly axisymmet- the magnetic field closes on itself in a rational number
ric field. Recently, extremum-seeking methods were used to of turns around the torus. A cross-section view of this
optimize this field in real-time [27]. deformation is illustrated in Fig. 9, where ‘magnetic island’
Even with a perfectly axisymmetric field, at high enough can be seen on the poloidal plane. Since the island X-points
pressure the plasma can helically deform, due to an insta- connect two flux surfaces that would otherwise be separated,
bility known as the Resistive Wall Mode (RWM) which an NTM leads to a local increase of thermal transport, and
leads to a loss of confinement followed by a disruption. This therefore to decrease of the local plasma pressure gradient.
deformation can be detected by magnetic probes surrounding This local reduction of the pressure gradient causes a de-
the plasma, and controlled by so-called non-axisymmetric crease of the local bootstrap current, which in turn causes
coils surrounding the plasma. These coils significantly differ a change in the magnetic field that increases the size of
from the poloidal field coils discussed in the previous part, the island. Appearance of an NTM can be triggered by the
since they generate a field that is not axisymmetric but can previously mentioned sawtooth crashes, or they may appear
have a 3D structure. Modern control methods have been used spontaneously, with a higher chance of being triggered at
to control such instabilities, for example [28], [29], and these higher pressure. NTMs are metastable, meaning that often
are expected to play a role in ITER. they are linearly stable, but if a perturbation generates a
large enough seed NTM, its growth rate may be positive.
B. Sawtooth and NTM control If an NTM’s size can be reduced below a given threshold,
Next to global responses of the plasma, described above, the growth rate will become negative and the mode will
localized 3D deformations of the plasma magnetic structure self-stabilize. Without suppression, an NTM might grow
have also been widely observed. Here we discuss only to a size large enough to couple to one of the previously
those most relevant for real-time control: those that must mentioned external modes, degrade the confinement and
be controlled to maintain the plasma at good performance cause a disruption.
(high pressure) and that can be affected by actuators. NTMs can be controlled by providing localized current
One is the so-called sawtooth instability, which manifests drive, again with ECCD, on the island location (Fig. 9).
itself as a periodic collape of the pressure in the innermost This localized current compensates for the loss of bootstrap
part of the plasma, leading to a redistribution of current and current. In present-day experiments, current drive is often
 Fast steerable mirror
Deposition on tearing mode Magnetic axis
to stabilize the ‘island’
ρb
High-power (~1MW)
microwave beam
z
Z
r
R ɸ P ρ
P Sɸ

Magnetic flux surfaces

Deposition closer
to center for Fig. 10. Magnetic configuration in a tokamak.
sawtooth control
which reduces the requirement on externally driven non-
Fig. 9. Schematic illustration of simultaneous control of several MHD
modes. An NTM is visible as a rotating ’island’ in the poloidal plane. inductive current for steady-state operation. The dynamics
Driving current with EC actuators on the location of the mode can stabilize of these core-plasma properties depend on both time and
the island and restore the nested flux surface topology. Simultaneously, space. Their high dimensionality (infinite-dimensional PDE
another EC actuator might be necessary to control the sawtooth instability,
which occurs closer to the center of the plasma. system), together with the nonlinearity exhibited by the
dynamics of these properties and the limited actuation capa-
applied on the island location by trial-and-error feedforward bilities available in tokamaks, makes the regulation of both
positioning of the launcher steering mirrors. Significant the magnitudes and the spatial profiles of these core-plasma
research effort has been put into achieving this alignment properties one of the most challenging control problems in
using feedback control. Also here, automatic optimization tokamaks.
algorithms based on extremum-seeking [33] as well as ap-
proaches requiring accurate calculations of the island loca- A. Control of Plasma Profiles
tion have been investigated [34]. Since accurate positioning Under ideal MHD conditions, the magnetic-flux sur-
of the microwave beam on the island is important, this places faces in a tokamak (see discussion in (Section II-B)) form
stringent requirements on the sensor accuracy for real-time toroidally nested surfaces around the magnetic axis as shown
estimation of the island position. Placing sensors in the same in Fig. 10. The axisymmetry provided by the toroidal geome-
viewing line as the injected microwave beams may alleviate try, which is indeed an ideal assumption, together with the se-
this problem [35]. lection of a spatial coordinate indexing the nested magnetic-
One difficulty in MHD control for reactor-grade devices, is flux surfaces reduces the three-dimensional problem to just
that the same (ECCD) actuators must be shared between dif- one dimension. The mean effective minor radius, ρ, which is
ferent control tasks. They may be used for other control tasks related to the toroidal magnetic flux, Φ, and to the vacuum
when no NTM is present, while they must be repurposed toroidal magnetic field at the geometric major radius R0
to NTM control when NTMs appear. While many examples (Fig. 3) of the tokamak, Bφ ,0 , by means of πBφ ,0 ρ 2 = Φ,
exist of dedicated experiments for NTM or sawtooth control, can be used as the indexing variable. At any point P on a
comparably few examples [36] exist of integrated control of magnetic-flux surface, the toroidal magnetic flux subtended
multiple MHD instabilities. In a future tokamak, the choice R
by that surface is defined as Φ = Sφ Bφ dSφ , where Sφ is the
of which MHD mode to control, with which actuator, must poloidal surface normal to the φ -axis depicted in Fig. 10.
be made automatically. This is the subject of intense study The mean effective minor radius is normalized as ρ̂ = ρ/ρb ,
in recent years [37], [38], [39], [40], [41], [42]. where ρb is the value of ρ at the last closed magnetic-flux
VI. CORE MAGNETIC AND KINETIC CONTROL surface as depicted in Fig. 10. This non-dimensional variable
A fundamental control problem arising in tokamaks is the ρ̂ is the spatial coordinate generally used to model the spatial
regulation of several properties of the core plasma such as dependence of the plasma dynamics in tokamaks.
density, temperature, current, and rotation since these prop- 1) 1D Nonburning Plasma Dynamics: The pressure pro-
erties are strongly linked to advanced modes of operation file, p, is given by
characterized by a high confinement state with enhanced p = ne kTe + ni kTi , (8)
magnetohydrodynamic (MHD) stability, which yields strong
improvements in plasma performance. Such improvements where ne and ni are the plasma electron and ion densities
are quantified by increases in energy confinement time, (units of # particles/volume), respectively, Te and Ti are the
plasma pressure, and fusion power density (properties for- electron and ion temperatures, respectively, and k is the
mally defined below). Moreover, in these advanced modes Boltzmann’s constant. The thermal energy, E, is given by
of operation a dominant fraction of the plasma current is E = 32 p (units of energy/volume since kTe and kTi have units
self-generated by the neoclassical bootstrap mechanism [43], of energy, usually keV). Similarly to both the poloidal and
the toroidal magnetic fluxes, both the pressure p and thermal Both the safety factor and the toroidal current density depend
energy E are constant on a magnetic-flux surface. on the gradient of the poloidal stream function, θ , ∂ ψ/∂ ρ̂.
a) Electron Density Dynamics: The electron density Therefore, it is common to speak interchangeably of the
transport equation (EDTE) is approximately written as [44] current profile, the q-profile, the ι-profile, the θ -profile, and
  the ψ-profile control. One of the main challenges associated
∂ ne 1 ∂ ∂ ne
= ρ̂ ĤDne + Se , (9) with the control of core-plasma properties such as p, q
∂t Ĥ ρ̂ ∂ ρ̂ ∂ ρ̂ and ωφ is the development of control-oriented models [46],
with boundary conditions ∂ ne /∂ ρ̂|ρ̂=0 = 0 and ne (1,t) = [47], [48] for the diffusive terms (Dne , χe , χφ , η) and
ne,bdry (t), and where t is the time, Dne is the electron the source terms (Se , Qe , tω , jni ) that are accurate enough
density diffusivity coefficient, and Se (ρ̂,t) represents any to correctly predict the plasma evolution for model-based
flux-surface averaged source or sink of electron density. control design purposes but simple enough to keep the
b) Electron Temperature Dynamics: When the heat control-design problem tractable.
diffusion is the dominant heat transport mechanism (transfer 2) Density, Temperature, and Pressure Profile Control:
via convection or particle transport can also be modeled if The control of density, temperature, and pressure profiles in
necessary), the electron heat transport equation (EHTE) can tokamak plasmas is challenging because of limited actuation
be written as [44] capabilities and a relatively poor understanding of particle
transport phenomena. In principle, these profiles can be
ĜĤ 2
  
3 ∂ 1 1 ∂ ∂ Te controlled by fueling mechanisms such as puffing of gas and
[ne Te ] = 2 ρ̂ χe ne + Qe , (10)
2 ∂t ρb Ĥ ρ̂ ∂ ρ̂ F̂ ∂ ρ̂ injection of frozen-fuel pellets and by heating mechanisms
such as neutral beam injection (NBI) and different sources of
with boundary conditions ∂ Te /∂ ρ̂|ρ̂=0 = 0 and Te (1,t) =
radio-frequency waves. It is believed that the gradient of the
Te,bdry (t), and where χe denotes the electron thermal con-
pressure profile around magnetic flux surfaces where q is a
ductivity and Qe represents the electron heat sources.
rational number can play a critical role in triggering certain
c) Toroidal Rotation Dynamics: The dynamics of the types of MHD instabilities (see Section V). Moreover, certain
toroidal angular velocity, ωφ , which denotes the rate at which density and temperature profiles are necessary to produce
the plasma rotates toroidally in the tokamak, is given by the the self-induced non-inductive bootstrap current [43], which
toroidal rotation equation (TRE) equation [44], favors steady-state operation.
2 ∂ (ni ωφ )
3) Current Profile Control: Shaping the current profile,
 
1 ∂ ∂ ωφ
mi hR i = f φ χφ n i + tω , (11) usually defined in terms of the safety factor q or rota-
∂t ρ̂ Ĥ ∂ ρ̂ ∂ ρ̂
tional transform ι profiles, has been demonstrated to be a
with boundary conditions ∂ ωφ /∂ ρ̂|ρ̂=0 = 0 and ωφ (1,t) = key condition for realization of advanced plasma scenarios
ωφ ,bdry (t), and where χφ is toroidal momentum diffusivity, characterized by MHD stability, improved confinement, and
mi is the ion mass, and tω is the ion-torque deposition source. possible steady-state operation (an operating scenario is
d) Poloidal Magnetic Flux Dynamics: The dynamics roughly defined as a target operating point and the path
of the poloidal stream function, ψ, which is related to used to reach that point). The achievement of some types
the poloidal magnetic flux as Ψ = 2πψ, is defined by the of current profiles favors generation of a so-called “internal
magnetic diffusion equation (MDE) [45], transport barrier” (a region where particle and heat transport
∂ψ η 1 ∂

∂ψ
 are strongly reduced), which improves confinement and facil-
= ρ̂D ψ + R0 Ĥη jni , (12) itates steady-state operation by enhancing bootstrap current.
∂t µ0 ρb2 F̂ 2 ρ̂ ∂ ρ̂ ∂ ρ̂
It has been demonstrated that current profile control can
with boundary conditions ∂ ψ/∂ ρ̂|ρ̂=0 = 0 and slow down, and possibly stop, temporal evolution of current
∂ ψ/∂ ρ̂|ρ̂=1 = −(µ0 R0 I p )/(2π Ĝ|ρ̂=1 Ĥ|ρ̂=1 ), and where profile peaking, which tends to develop due to the fact that
η is the space-dependent plasma resistivity, which is a the plasma temperature is higher in the core and plasma
nonlinear function of the electron temperature Te , jni is the resistivity is inversely related to temperature. Reduction of
space-dependent non-inductive current sources, µ0 is the this peaking effect also allows avoidance of some instabilities
vacuum magnetic permeability, and I p is the plasma current. or disruptive events related to presence in the plasma of
F̂, Ĝ, Ĥ, Dψ , fφ , hR2 i are spatial profiles corresponding to magnetic flux surfaces where q is a rational number, such
a particular magnetic equilibrium. The safety factor, q, and as NTMs (see Section V-B).
the rotational transform, ι , 1/q, which are measures of the As can be noted from (12), the dynamics of the poloidal
pitch of the magnetic field lines, are defined as magnetic flux can be modified through three different mecha-
nisms: i- the plasma resistivity η (diffusivity control), which
1 dΦ dΦ Bφ ,0 ρb2 ρ̂ can be “actuated” by controlling the electron temperature
, q(ρ̂,t) , =− =− . (13)
ι(ρ̂,t) dΨ 2πdψ ∂ ψ/∂ ρ̂ profile by using different radio-frequency heating mecha-
Moreover, the toroidal current density, jφ , is computed as nisms such as ECH (electron cyclotron heating); ii- the non-
  inductive sources in jni (interior control) such as neutral
1 1 ∂ ∂ψ beam injection (NBI) and radio-frequency current drives,
jφ (ρ̂,t) = − ρ̂ ĜĤ . (14)
µ0 ρb2 R0 Ĥ ρ̂ ∂ ρ̂ ∂ ρ̂ which may include ECCD (electron cyclotron current drive),
 5 dominant physics of q-profile and W responses to actuators.

EC-assisted Formation Phase

Target (q=1.8403,t=3s)
Actual 4) Rotation Profile Control: Control of the toroidal rota-
4 tion profile has recently attracted much attention because of
its connection with MHD stabilization and advanced-mode
3 access. As an alternative to active stabilization of RWMs
Physics-Study
qmin

Flattop Phase by magnetic feedback control [28], [55], briefly discussed in

2 Section V-A, experiments have shown that RWM passive
stabilization is also possible by plasma toroidal rotation
1
q-Profile+W Control
Rampup Phase coupled with an energy dissipation mechanism. It has been
#163832 (FF+FB)
shown that the critical rotation speed for RWM stabilization
0 is a function of the rotation profile shape, implying a radially
0 0.4 1 2 3 4 5
Time [sec]
8 distributed stabilizing mechanism. RWM stability therefore
7 depends on both β (a measure of plasma pressure, defined
6
below) and toroidal angular velocity profile ωφ . Unfortu-
Initial Profile (0.411 s) nately, RWM stabilizing rotation is often dissipated by drag
5
torques generated by different mechanisms, including plasma
4
q

fluid viscosity, and interaction between the plasma fluid with

3 Final Profile (2.991 s) magnetic field perturbations. Modification of the rotation
2 profile through active control [56], [57], in order to affect
1 Target Profile (2.991 s) energy confinement, can improve stability of not only RWMs
0
#163832 (FF+FB) but also NTMs (see Section V-B). Other applications that
0 0.2 0.4 0.6 0.8 1 might call for rotation profile control include: i- modification
0.8 ρ̂
Formation Phase

Actual of the NTM rotation frequency to facilitate ECCD stabiliza-

0.6 Target
Physics-Study tion of the NTM as it rotates toroidally in the plasma, ii-
W [MJ]

0.4 Flattop Phase

establishment of conditions favorable for access to advanced
0.2 q-Profile+W Control
Rampup Phase #163832 (FF+FB)
operation modes (e.g., free of certain MHD instabilities).
0
0 0.4 1 2 3 4 5 Neutral beam injection (NBI), and non-axisymetric magnetic
Time [sec]
field coils (the same coils that are used for RWM feed-
Fig. 11. Experimental test of feedforward + feedback q-profile con- back stabilization) that allow modulation of the error fields
troller based on off-line+on-line model-based optimization. Target (Section V-A) for magnetic braking (controlled drag torque),
q profile achieved at t f = 3, 000 ms while tracking a target W . provide effective torque sources in (11) for control of the
rotation profile in tokamaks.
FWCD (fast wave current drive), LHCD (lower hybrid
current drive), etc.; iii- the plasma current I p (boundary B. Control of Plasma Scalars
control), which can be “actuated” by controlling the inductive Magnetic and kinetic profiles are very strongly and nonlin-
component of the plasma current by transformer action. early coupled in tokamak plasmas (note that the EHTE (10),
Fig. 11 shows the performance of active control in DIII-D TRE (11), and MDE (12) are coupled by means of both
shot #163832. The goal was to drive both q and thermal their diffusive terms (through η, χe , and χφ ) and their source
stored energy W (defined as a volume-integrated quantity be- terms (through jni , Qe , and tω )). Moreover, a single actuator
low), from a given initial state at ti = 400 ms (end of plasma may have strong effects on more than one plasma profile.
formation phase) to a target final state at t f = 3, 000 ms (start It seems natural to try to develop an integrated magnetic
of physics-study flattop phase) by controlling individual on- and kinetic profile control algorithm. However, actuation
axis and off-axis NBI powers, total EC power, and total resources are limited by design in tokamaks. Therefore,
plasma current. This target final state is characterized by as we keep integrating control objectives, controllability
a minimum value of q, denoted as qmin , of around 1.85, issues arise. The question that follows is to what extent
and a value of q at ρ̂ = 0.95, denoted as q95 , of around several control goals can be satisfied simultaneously. Due
5.5. A primary goal for the DIII-D research program is to to actuation limitations, the “profile” control expectation for
develop the physics basis for a high q (qmin > 1.5 − 2.0), a given plasma variable may be reduced to its control at
high W , steady-state operation mode that can serve as the two or three locations at most. In some cases only control
basis for future steady-state burning plasmas. The tight of a plasma variable at a single point or integrated over the
regulation of both q and W is made possible by combining plasma volume may be feasible. On the positive side, this
off-line (feedforward control) [49], [50], [51] and on-line may be enough for proper operation. This is particularly
(feedback control) [52], [53], [54] model-based optimization true when the control objective is given in terms of these
(i.e., Model Predictive Control (MPC)). At the core of volume-averaged plasma properties. For instance, thermal
both optimizations are physics-based, control-oriented mod- stored energy, W , is defined as
∂V (ρ̂,t)
Z Z ρ̂=1
els of different complexities (off-line optimization: nonlinear
W (t) , E dV = E(ρ̂,t) d ρ̂, (15)
model, on-line optimization: linearized model) capturing the Vp ρ̂=0 ∂ ρ̂
where Vp is the plasma region enclosed within the last shown in Fig. 12. The DT reaction appears as the most
magnetic-flux surface, and V (ρ̂) is the plasma volume en- promising not only because of its relatively large release
closed by the magnetic-flux surface labeled with ρ̂. of energy per reaction, QDT , but also because of the fact
1) 0D Burning Plasma Dynamics: By also computing that the reactivity reaches the highest value at the lowest
volume-averaged values for densities of different species, it temperature when compared with other reactions. As can
is possible, for instance, to model the dynamics of a DT be noted from the figure, the reactivity increases as the
plasma by using a zero-dimensional model: temperature increases, which in turn produces more fusion
dW W power density and heating. This property of burning plasmas
=− + Pα − Prad + POhm + Paux , (16) may lead to thermal excursions (uncontrolled heating) or
dt τW
quenching (uncontrolled cooling). But even when operating
dnα nα
= − + Sα , (17) at stable equilibria, system performance during transients and
dt τα against disturbances could be undesirable without feedback
dnD nD R control.
= − + SD − Sα + SD , (18)
dt τD Neglecting ohmic power, which is usually small compared
dnT nT to the other sources in a burning plasma, and optimistically
= − + STR − Sα + ST , (19)
dt τT assuming that the radiation power is also negligible, at steady
dnI nI state (dW /dt ≡ 0) the energy balance (16) reduces to
= − + SI + SIsp , (20)
dt τI
W
where nD , nT , nα , and nI are deuterium, tritium, alpha- − + Pα + Paux = 0 (24)
τW
particle and impurity densities, respectively, Pα , Sα Qα
is the alpha-particle heating from fusion reactions, Sα = Taking into account that nα , nI << nD , nT , and assuming Te =
nD nT hσ νiDT is the source of alpha particles (i.e., 42 He Ti and nD = nT , which implies from (21) that ne = ni and
particles), hσ νiDT is the DT reactivity, Qα = 3.52 MeV is n = 2ne , we can firstly write the fusion power density in (23)
the energy deposited in the plasma by each alpha particle, as Pf ∝ β 2 B4 , where β , 2ne kTe /(B2 /(2µo )) denotes the
Prad is the radiation losses, POhm is the ohmic heating and ratio between kinetic and magnetic pressure, and secondly
Paux is the controlled auxiliary power, SD R , SR and Ssp are conclude from (24) that to operate at ignition (Paux ≡ 0 ⇔
T I
particle fluxes resulting from plasma-wall interactions, while Q , Pf /Paux = ∞) we need
SD , ST and SI are the controlled injection rates of deuterium, Te2
tritium and impurities, respectively. Electron density is given ne τW Te |DT
IGN ∝ (25)
hσ νiQDT
by the quasi-neutrality condition (# electrons = # protons)
This condition is usually referred to as the Lawson criterion
ne = nD + nT + 2nα + ZI nI , (21) (or as Triple Product if written in this form). The expression
on the right is plotted in Fig. 12. This condition indicates
where ZI is the impurity atomic number. The ion density is
that the product of confinement and pressure (density and
given by ni , nD + nT + nα + nI , whereas the total plasma
temperature) needs to be above certain levels to achieve
density is simply n = ne + ni . The energy confinement time,
ignition. When compared to other fusion reactions, the DT
τW , is modeled by scaling it with key plasma and machine
reaction appears once again as the most promising one since
parameters. The IPB98(y,2) scaling law is given by
it achieves the lowest minimum at the lowest temperature.
0.19 0.58 0.78 0.41 −0.69
τW = 0.0562HI p0.93 R1.97 B0.15
φ M ε κ ne Ptotal , (22) Unfortunately, as we increase kinetic pressure in order to
increase plasma performance (β , Pf , Q), we reach MHD
where H is a constant that depends on the quality of the stability boundaries since many of these instabilities are
plasma confinement, R is the plasma major radius, M is triggered by pressure (see Section V).
the effective mass, ε = a/R is the aspect ratio, a is the 2) Burn Control: Due to the nonlinear coupled dynamics
plasma minor radius, κ is the vertical elongation, Ptotal = of the plasma, feedback control of the burn condition will
(Pα − Prad + Pohm + Paux ) × V is the total plasma power in be necessary in ITER and future fusion reactors to avoid
MW, and V is the plasma volume. The ITER values of the undesirable transient performance and to respond to changes
parameters in (22) are listed in [58]. The particle confinement in plasma confinement, impurity content, or operation con-
times are scaled with the energy confinement time τW such ditions that could lead to thermal excursion or quenching.
that τα = kα τW , τD = kD τW , τT = kT τW and τI = kI τW , where Since the primary goal is to regulate the overall amount of fu-
kα , kD , kT and kI are constants. sion power produced by the reactor, 0D models such as (16)–
The fusion power density for the DT reaction is given by (20) are appropriate for the design of controllers for tight
Pf = 5Pα = nD nT hσ νiDT QDT , (23) regulation of plasma density and temperature. In order to
overcome the operability limits imposed by the linearization
where QDT = 5Qα = 17.6 MeV. The DT reactivity hσ νi, of the burn dynamics in prior work, nonlinear techniques for
which is the average of the cross section σ (probability burn control [59], [60] have been proposed to account for the
of fusion) over a Maxwell-Boltzmann distribution of the non-local character of the dynamics and to extend the stable
velocity field, is a function of the plasma temperature as operating region. The controllers utilize several actuators
 reactor might not be able to maintain the originally specified
1024 10-21
concentration (90% T + 10% D in the DT fueling line and
100% D in the D fueling line), particularly for long pulse
10-22 operation. This fueling limitation might severely impact burn
1023 control in ITER and prevent the achievement of desired
10-23 high-Q operating points. The steady-state plasma conditions
Triple Product

DT Reactivity
for Q = 10 operation in ITER are studied in the density-
1022 10-24 temperature space [66]. The steady-state solution of (16)–
(20) is found repeatedly over a grid of fixed ne and T0 values,
10-25
where T0 is the temperature on the magnetic axis of the
assumed temperature profile at the moment of computing the
1021
volume-averaged quantities in (16)–(20). The unknowns are
10-26
the alpha particle density fraction fα = nα /ne , the external
ne w
Te < v> fueling rate of deuterium SD , the external fueling rate of
1020 10-27 tritium ST , and the impurity density fraction fI = nI /ne .
100 101
10 2

Kinetic Temperature [KeV] The resulting contour lines are plotted in the ne -T0 space in
Fig. 13. This study assumes H = 1, γ , nT /(nT + nD ) = 0.5,
Fig. 12. Reactivity (red - right) and Triple Product (Lawson) criterion and Neon as impurity (ZI = 10). Multiple lines mark the
(blue - left) for a DT burning plasma.
saturation curves of the ITER actuators. The dashed-blue line
simultaneously, using auxiliary power modulation to prevent marks where Paux = 0. Below this line, the auxiliary power
quenching, impurity injection to stop thermal excursions by is negative and the results are physically meaningless. The
increasing radiation losses, and fueling modulation to regu- solid-blue line denotes where Paux = 73 MW, the maximum
late the density. Isotopic fuel tailoring [61], which requires power expected to be available in ITER. The yellow line
separated actuation of the deuterium and tritium fueling marks where the D injector saturates. Where the D-T injector
systems to allow for control of the isotopic mix in the saturates with a tritium concentration of 90%, 80%, and 70%
core, was exploited more recently to use the relative mix is plotted using the solid-red, dashed-dotted-red and dotted-
of tritium fuel in the plasma as a virtual control to cool red lines, respectively. On the magenta line, the total plasma
the plasma during thermal excursions. In this way, impurity power, Ptotal , equals the threshold power Pthresh (the plasma
injection, and the consequent reduction in fusion power is in H-mode (high-confinement mode - an operating mode
because of enhanced radiation, needs to be used only in cases characterized by an enhanced level of energy confinement)
where isotopic fuel tailoring is limited by severe particle above this line). The green zone marks the steady-state
recycling conditions [62]. Recycling describes the various operable space with Q = 10 for the given ITER plasma
processes that may occur when an ion exits the plasma and (tritium concentration of 90% in the DT fueling line). Here,
strikes some plasma facing component. The incident ion the actuator saturation limits are not violated and the plasma
either immediately reflects back into the plasma or implants is in H-mode. This operable space, which is already relatively
itself into the material. After implantation, the particle might small, shrinks as the tritium concentration decreases and
get trapped in the material. Alternatively, the particle could vanishes when the tritium concentration in the DT fueling
randomly diffuse out of the material, re-emit into the plasma line drops below 60%. Even if nominal concentrations could
and contribute to the recycling flux. Recent experimental be initially achieved, it might not be possible to keep them up
results in DIII-D [63] have shown the in-vessel coil system for the total duration of long pulses. After about 400s of burn,
as another effective actuator for burn control. The in-vessel the T concentration in the DT line would likely begin to drop
coils can be used to generate non-axisymmetric magnetic and could go down to 80%. To add to the challenge, neither
fields that are able to reduce the energy confinement time, the perturbation nor the drift in the concentrations would
becoming in this way an alternative actuation mechanism for be directly measurable without an expensive solution. This
reduction of the plasma stored energy [64]. In addition to be- has motivated recent work towards the design of a nonlinear
ing capable of rejecting perturbations leading to both thermal controller capable of overcoming unmeasurable variations of
excursion and quenching, when complemented with real-time the D-T concentration expected in the fueling lines during
model-based optimizing schemes [65], the controllers can long-pulse operation [67] by exploiting Lyapunov Redesign
drive the system from one point to another during operation. techniques.
Hence, the controllers can increase or decrease β , modify
the fusion power, the temperature or the density, and go VII. TOKAMAK EXCEPTION HANDLING
from a subignition (Q < ∞) to an ignition (Q = ∞) point As discussed in the introduction, low disruptivity operation
and vice versa. Moreover, the controllers are robust against is key to maintaining integrity of the tokamak. There are
model uncertainties in the confinement times. multiple disturbances that can occur asynchronously includ-
The ITER Organization (the organization responsible for ing, e.g., the failure of an actuator or sensor, which can lead
building and operating ITER) recently realized that the to plasma destabilization so that it disrupts. We label this
plant system producing the tritium needed to fuel the ITER type of disturbance an event. Some events can be handled
 added in an ad-hoc manner when deemed necessary, without
a clear architectural strategy. Exceptions were typically han-
dled on a case-by-case basis. In the simplest case, all of them
were simply assigned to a single handling policy, such as
switching off all heating and fueling actuators and ramping
down plasma current. With hundreds of input signals and
dozens of control parameters and actuators, the number of
possible exceptions is enormous. In addition, proper handling
of an exception is dependent on the system state (e.g., plasma
or coil current levels, total energy content, etc.) when it
occurs. With a multitude of potential fault sources and many
variations in plasma dynamics, the decision logic can become
very complex. Developing this logic organically leads to
an increasingly complex system with too finely detailed
states, poor choices in exception definition and even illogical
exception responses [11]. Instead, exception handling should
be designed systematically, based on knowledge of possible
fault sources and their impact.

A. Defining the set of exceptions

Fig. 13. Contours of constant ITER fusion power [MW] (black) for plasma
with kD = kT = kα = 1 and no particle recycling. Shaded area represents
The first step in defining an automated system is to identify
Q=10 operation, bounded above by the maximum 90% tritium fueling and all exceptions that must be handled by the PCS. This begins
below by the power required for enhanced confinement (H-mode) operation. by determining which events can be handled with sufficiently
robust control and which must be treated as exceptions. This
by sufficiently robust control. Others, like the loss of a process requires participation of both physics operators (re-
critical actuator, require a change in control approach such sponsible for operation of the device) and control engineers,
as switching to use of a controller that employs a different who are best able to understand what can be handled via
actuator or changing the control goal, e.g. to controllably standard disturbance rejection. Next, physics operators must
terminating the discharge (i.e., without disrupting). If a define a handling policy for each exception, i.e., how the
response requires changing the control approach, we relabel control goal should change to effectively prevent or mitigate
the event as an exception. any negative consequences of the exception.
Exception handling (EH) is a function or combination 1) Basic tokamak exception handling: In Fig. 14, an
of functions in the PCS that detects such exceptions and example is given of active instability control using a single
adapts the control approach to prevent or mitigate a control actuator. Preferred operation is in the ideal control zone,
degradation or failure. The exception breaks the normal flow where even large actuator perturbations do not affect stabiliz-
of execution and usually executes a pre-determined handling ability. Improving fusion performance often requires relying
policy. A handling policy is basically a specification of a on control robustness at a less stable operating point. An
control goal, i.e., what device operators would like to happen actuator fault could result in a transition back to a stabilized
when a particular exception appears. The plant and the recovery state, if properly handled by EH, or a loss of
nominal plasma control should be so reliable that exceptions control, requiring a hard (uncontrolled) termination of the
are rare. discharge. If in a recovery state, the EH system can decide
Controlled termination of a tokamak discharge is not easy, to recover or to terminate the discharge.
especially if initiated when the plasma is at high performance This example (and all other desired handling of tokamak
(e.g., high plasma current or during the fusion-burn-phase) exceptions) can be represented by a finite state diagram
and even more so when control robustness has been com- (Fig. 15) that consists of a number of characteristic states
promised. Not all exceptions necessarily lead to controlled - normal (ideal) control, robust control, recovery state, con-
terminations. For example, EH can be used to optimize pulse trolled termination and hard termination. Hard termination is
evolution including enabling necessary changes of control the application of measures to mitigate the impact of, rather
schemes. Nevertheless, the risk of tokamak discharge disrup- than prevent, a disruption. Mitigation measures can reduce,
tions make a well-developed exception handling capability a but usually not eliminate, the impact of the exception on
vital part of a tokamak control system. the device. Their use should be minimized by preventing,
Early versions of Plasma Control Systems were developed if possible, the disruption from occurring. Each transition
around requirements of individual control functions, such between states in Fig. 15 represents an exception handling
as the need to control the plasma current, the magnetic policy. The transition goal state depends on the reason for
confinement configuration, or the plasma density or the the transition - often an actuator or control fault - and the
heating systems. Fault or exception handling methods were initial system state.
 No control
Normal Robust Hard
Control Control Termination

Robust control

Fault: - X Controlled
Recovery
Termination
Plasma instability Y

Recover Fig. 15. A number of characteristic states can be identified from the
example shown in the previous figures, as well as the transitions between
Ideal control
Terminate them, resulting in the following basic finite state-diagram. The transitions
between are the handling policies. However, this only represent one typical
control function, and there exist many, that all again interact with each other,
creating possible a very complicate state-diagram, with numerous possible
Inherently stable transitions, or exceptions.

(or predicted in advance if possible), then an alternate control

Actuator capacity X must be activated. A number of technologies exist for auto-
mated detection including, e.g., simple threshold comparison
Fig. 14. Control of a plasma instability with growth rate defined by
parameter Y, using an actuator X. By means of sufficiently robust active
of a relevant signal, statistical classification algorithms, and
control, it is possible to operate at high Y (cross-hatched symbol) to optimize neural networks, the latter two of which are able to operate
fusion performance. An actuator fault could cause the instability to be on multiple signals or information sources. Ultimately, all
uncontrollable (red symbol) and result in disruption. Exception handling
should ensure that such events are detected and bring the plasma and plant
methods derive a ”decision variable” whose value is used to
into a more stable lower performance state (dark green symbol), after which make the decision about whether to declare an exception. A
recovery of operation or termination (reducing X and Y to zero) is possible. threshold is chosen so that the exception is declared when the
decision variable’s value is greater than the threshold. One
Finite state machines have been implemented in existing
method for choosing this threshold uses the so-called ROC
plasma control systems in efforts to improve disruption
(receiver operating characteristic) curve (Fig. 17). The key
prevention [38], [68] and also to better optimize actuator
metrics for detection/prediction processes are the probability
management (i.e. manage requests from different control
of correct detection (Pd ) and probability of false alarm (Pf a ).
functions for the same set of actuators) [40], but these efforts
The threshold is chosen so that a satisfactory trade-off is
have not yet reached the level of sophistication required for
achieved between detection of the true exceptions (Pd ) and
devices like ITER.
spurious declaration of exception when none exists (Pf a ).
2) Determining required exceptions and priorities: Using
Once an exception is declared, control must switch to
Failure Mode Effect Analysis (FMEA), a complete inventory
execution of a predetermined handling policy, i.e., a specified
is made of possible exceptions (failure modes) that could
new control goal (typically instantiated as a set of control
lead to undesirable outcomes. The FMEA also determines
reference signals, which implicitly include, of course, the
the likelihood of occurrence of the failure mode (per plasma
identity of the variable(s) chosen to be controlled). Executing
discharge), the severity or impact of the failure, and the
the policy may or may not require the use of a different set
difficulty to detect it, which in combination gives a risk
of controllers.
value due to the exception (i.e. occurrence × severity × un-
detectability). In an alternative approach, a Fault Tree Anal- The effectiveness of a handling policy is often compro-
ysis (FTA) works backward from an undesirable outcome mised by a slow response of the plasma or actuators. To over-
to identify potential failure modes. This analysis has the come this, forecasting of future events can make it possible
advantage over the FMEA of allowing the process that leads to still provide timely response. Forecasting functions (open-
to the event to be evaluated, as opposed to looking at the loop system models), whose predicted signal values are used
failure mode in isolation. Both methods allow prioritization to evaluate likelihood of future exceptions, can be included
of faults based on exception consequences. Figure 16 shows in the PCS. Methods used can be as simple as extrapolation
a basic tree of characteristic events that can lead to a using a linearized response model or can be based on faster-
plasma disruption. Many possible exception sources on the than-real-time simulation of future plasma evolution.
left (human error, computing error, transient event, sensor Obviously most attention on forecasting is focused on
or actuator fault), can lead to a sequence of events that disruptions. Simply detecting an ongoing disruption enables
eventually ends with a disruption. For a plant as complex triggering of mitigating measures for some of its conse-
as a tokamak, this collection of exception sources leads to a quences, but not all [69]. For example, mitigation of heat
large list of exceptions. loads due to the fast quench of thermal energy during dis-
ruptions (on order of a few ms) is not possible. The standard
B. Automated exception handling approach mitigation method of injecting large amounts of particles
The obvious ”algorithm” for automating exception han- into the tokamak takes at least several tens of milliseconds.
dling is that an exception, if it occurs, must be first detected Thus considerable efforts have been dedicated to providing
 C. Managing complexity
Even a concise definition of exceptions using, e.g., FMEA
and FTA, will result in a very large number of them. Together
with the many plasma and plant state-specific responses , this
can result in an explosion of possible exception handling
paths. It is important to contain the complexity of PCS EH
or it may become impossible to validate its design or to
commission it.
The only feasible method for limiting the number of
exceptions is to ignore faults that do not lead to important
consequences if not handled in real-time. On the other hand,
exception priorities defined by FMEA and FTA can be
used to manage their handling. For example, exceptions do
not generally occur in isolation, since the modified plasma
state resulting from an initial exception will often create
Fig. 16. A basic event tree analysis of a tokamak discharge disruption. For conditions in which other exceptions occur. In fact, multiple
a single disruptive outcome, there are multiple potential causes (either alone exceptions can often occur closely together in time. If one
or in combination) including plasma instabilities, operational limits, control
errors, actuator and sensor faults, and external transients. Computation or
exception follows another having lower priority, response to
human error can also lead to disruption. The many-to-many relationships the later exception may need to override the earlier triggered
result in a complex set of potential event paths leading to disruption. This handling policy. The handling policies of the two exceptions
diagram can be compared with an actual root-cause analysis for disruptions
in the JET tokamak [70]. Next to the boxes defining event types, basic
could be contradictory. The first exception could request to
techniques to mitigate the events are shown. turn-off an actuator that is required to handle the later, more
important exception. Delay in re-activation of this actuator
could affect handling of the later higher-priority exception.
advance prediction of thermal quench in plasmas [10]. Some guidance for how to manage conflicting exception re-
sponses through priorities is available. For example, [72] and
Prediction of disruptions can be achieved by detecting spe- the references therein discuss prioritized execution in hard
cific instabilities that grow in the plasma prior to disruption. real-time systems (but focus on faults in the computational
Such techniques are reasonably successful but generally not process). There is also experience with fault handling in
good enough to predict disruptions in devices such as ITER nuclear power plants that may be relevant (e.g., [73] and
which, because of limited tolerance to the impact of thermal references therein). However, further research on prioritized
quenches, requires them to have a very high success rate execution is needed for application in tokamaks.
[69]. In recent years, more advanced methods have aimed One can also attempt to limit the number of possible
to predict disruptions using neural networks. These neural handling policies. In tokamak research, the stability and
networks have been trained on a large set of example plasma characteristics of discharge terminations are studied [74]. The
discharges, to distinguish between those that disrupt and aim is often to develop a single policy robust enough to be
those that do not. The disadvantage of these methods is used for discharge termination in response to a large number
that they focus purely on the plasma stability and not, for of events. For simple plasma discharges, this may well
example, on the detection of incipient actuator and sensor be possible. However, controlled termination of a burning
faults or control errors, which are often the root cause of the plasma - the main target of ITER operation - is complex,
plasma destabilization. The PCS must monitor all of these and variations in handling policy are needed depending on
potential sources of disruption. the plasma state or control capabilities. Hence, the control
In addition to disruptions, the PCS may be able to forecast system will be required to choose from a list of termination
impending threats if continuing as scheduled. For example, policy variants.
a control scheme may be sufficiently robust in the current Complexity can also be reduced by defining intermediate
plasma state. However if, for example, plasma pressure is handling policies (goals) such that a sequence of such
increased further, the available set of sensors and actuators policies accomplishes an overall goal. For example, if the
may not be capable of supporting the required control overall goal is to terminate the plasma discharge, there is
robustness. So-called hazard-functions can be determined in likely to exist an intermediate goal to first bring the plasma
real-time based not only the stability of the plasma, but also to a more stable target state. Although this is only a one-step
the general state of the plasma control [71]. Hazard-functions example, the value in the approach is the ability to define
provide estimated future event probabilities as a function of for each exception a sequence of steps between intermediate
time, allowing the control system to act if the probability goal states, with a smaller set of potential states at each step,
of a high-risk event is large but the event has not yet taken which at the end reaches the final termination goal.
place. This approach provides more complete coverage of This approach does not actually reduce the number of
detection of the root causes of disruptions. paths that must be taken between the overall set of excep-
 1.0
make updates to control methods used during this evolution,
0.9
a significant extension of traditional MPC would be required.
0.8
Since requirements for nominal control are usually well
0.7
defined, validation of a controller’s capabilities prior to
0.6
operational use is straightforward. However, validation of the
0.5
Pd

suggested MPC algorithm’s execution of exception response

0.4
or even defining its prerequisite control requirements may be
0.3
challenging. Specification of control metrics such as response
0.2
time, settling time, tracking error, etc. are well-understood,
0.1
but a similar metric for ”goal achievement” must still be
0.0
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 defined to enable quantitative comparison with alternative
Pfa
handling policies or with requirement specifications.
Fig. 17. Illustration of an ROC curve, which plots calculated Pd and Pf a
values for many possible thresholds. A point on this curve can be used to R EFERENCES
choose a threshold for the decision variable .
[1] [Online]. Available: http://www.iter.org
[2] Pironti and Walker, “Control of tokamak plasmas: introduction to a
tions and the termination goal, but it does allow re-use of special section,” IEEE Control Systems Magazine, vol. 25, no. 5, pp.
previously programmed handling policies. This means that 24–29, Oct 2005.
construction of a handling policy for a new exception is not [3] ——, “Fusion, tokamaks, and plasma control: an introduction and
tutorial,” IEEE Control Systems Magazine, vol. 25, no. 5, pp. 30–43,
required to define the entire evolution to termination from the Oct 2005.
plasma state at the time of exception. Instead, it can choose [4] Walker et al., “Emerging applications in tokamak plasma control,”
from already defined intermediate ”handling” states the one IEEE Control Systems Magazine, vol. 26, no. 2, pp. 35–63, April
2006.
that is most easily achieved from the beginning state. In fact, [5] [Online]. Available: https://www.iter.org/sci/Fusion
as the exception handling system of the JET tokamak was [6] [Online]. Available: http://www.ga.com/diii-d
being built over time, it was found that the system naturally [7] Artsimovich and Shafranov, “Tokamak with non-round section of the
plasma loop,” Soviet Phys. - JETP Letters 13.72, 1972.
evolved into this pattern. [8] Freidberg, Ideal MHD. Cambridge University Press, 2014, pp. i–iv.
[9] Hender et al., “Chapter 3: MHD stability, operational limits and
D. Aspects of control design for EH disruptions,” Nuclear Fusion, vol. 47, no. 6, pp. S128–S202, 2007.
[10] Strait, Barr, Baruzzo et al., “Progress in disruption prevention for
Many requirements for controllers used to execute han- ITER,” Nuclear Fusion, vol. 59, no. 11, p. 112012, jun 2019.
dling policies are similar to those of nominal control, but [11] Humphreys et al., “Novel aspects of plasma control in ITER,” Phys.
of Plasmas, vol. 22, no. 2, p. 021806, Feb 2015.
there are some important additional constraints. Smoothly [12] Ariola and Pironti, Magnetic Control of Tokamak Plasmas, ser. Ad-
transitioning to the new control is critical to prevent trig- vances in Industrial Control. Springer International Publishing, 2016.
gering of a secondary exception while handling the first. [13] Skogestad and Postlethwaite, Multivariable Feedback Control. John
Wiley & Sons Ltd., 2005.
Robustness of the new control is paramount, even at the [14] Albanese, Ambrosino, and Mattei, “CREATE-NL+: A robust control-
expense of performance, since inadequate achievement of the oriented free boundary dynamic plasma equilibrium solver,” Fusion
control goal(s) may lead to device damage. Robustness can Eng. and Design, vol. 96-97, pp. 664–667, oct 2015.
[15] Albanese et al., “Design, implementation and test of the XSC extreme
also serve to reduce control system complexity (and therefore shape controller in JET,” Fusion Eng. and Design, vol. 74, no. 1-4,
risk) by enabling one handling policy to service multiple pp. 627–632, nov 2005.
exceptions. [16] Anand et al., “A novel plasma position and shape controller for
advanced configuration development on the TCV tokamak,” Nuclear
In the traditional approach, a handling policy would define Fusion, vol. 57, no. 12, p. 126026, dec 2017.
reference signals which, if followed, would transition the [17] Albanese et al., “A MIMO architecture for integrated control of plasma
system into a ”safe” state. It would also define the controllers shape and flux expansion for the EAST tokamak,” in 2016 IEEE
Conference on Control Applications (CCA). IEEE, sep 2016, pp.
to be used to force the system to track those references. 611–616.
However, the most advanced handling policy would be one [18] Gerkšič et al., “Model predictive control of ITER plasma current and
that is not prescribed in advance, but is optimally chosen by shape using singular-value decomposition,” Fusion Eng. and Design,
vol. 129, pp. 158–163, apr 2018.
the control system. Model-based predictive control (MPC)
[19] Ariola et al., “A Modern Plasma Controller Tested on the TCV
could determine, based a set of constraints and faster than Tokamak,” Fusion Technology, vol. 36, no. 2, pp. 126–138, sep 1999.
real-time predictive modeling of the system, what the op- [20] Nouailletas, Nardon, and Brémond, “Robust Vertical Plasma Stabiliza-
timum transition is between the initial system state when tion of the future tungsten divertor configuration of Tore Supra,” IFAC
Proceedings Volumes, vol. 47, no. 3, pp. 1349–1354, 2014.
the exception occurs and a specified stable target state at [21] Ferrara, Hutchinson, and Wolfe, “State Reconstruction and Noise
the end of the MPC prediction horizon. Using the example Reduction by Kalman Filter in the Vertical Position Control on Alcator
of discharge termination, instead of following a predefined C-Mod,” Fusion Science and Technology, vol. 56, no. 4, pp. 1476–
1488, nov 2009.
set of reference signals an MPC can choose the path that it [22] Hyatt et al., “Designing, constructing and using Plasma Control
follows to achieve the target end state. A mixture of hard System algorithms on DIII-D,” in 2013 IEEE 25th Symp. on Fusion
and soft state constraints on the MPC can be used to avoid Eng. IEEE, jun 2013, pp. 1–6.
[23] Hahn et al., “Progress and improvement of KSTAR plasma control
potential sources of instability in addition to dealing with the using model-based control simulators,” Fusion Eng. and Design,
usual actuator limitations. However, if it is necessary to also vol. 89, no. 5, pp. 542–547, may 2014.
[24] Nouailletas et al., “The WEST plasma control system: Integration, [51] Wehner et al., “Optimal Current Profile Control for Enhanced Repeata-
commissioning and operation on first experimental campaigns,” Fusion bility of L-mode and H-mode Discharges in DIII-D,” Fusion Eng. and
Eng. and Design, vol. 146, pp. 999–1002, sep 2019. Design, vol. 123, pp. 513–517, 2017.
[25] Kolemen et al., “Strike point control for the National Spherical Torus [52] Ou et al., “Receding-Horizon Optimal Control of the Current Profile
Experiment (NSTX),” Nuclear Fusion, vol. 50, no. 10, p. 105010, oct Evolution During the Ramp-up Phase of a Tokamak Discharge,”
2010. Control Eng. Practice, vol. 19, pp. 22–31, 2011.
[26] Zohm, Magnetohydrodynamic Stability of Tokamaks. Weinheim, [53] Maljaars et al., “Control of the tokamak safety factor profile with
Germany: Wiley-VCH Verlag GmbH & Co. KGaA, dec 2014. time-varying constraints using MPC,” Nuclear Fusion, vol. 55, no. 2,
[27] Lanctot et al., “Error field optimization in DIII-D using extremum p. 023001, 2015.
seeking control,” Nuclear Fusion, vol. 56, no. 7, p. 076003, jul 2016. [54] Wehner et al., “Predictive control of the tokamak q profile to facilitate
[28] Dalessio et al., “Model-based Robust Control of Resistive Wall Modes reproducibility of high-qmin steady-state scenarios at DIII-D,” in Proc.
via µ-synthesis,” Fusion Science and Technology, vol. 55, no. 2, pp. of IEEE Conf. on Control Appl., Buenos Aires, Argentina, Sep. 19-22
163–179, 2009. 2016, pp. 629–634.
[29] Katsuro-Hopkins, Bialek, Maurer, and Navratil, “Enhanced ITER [55] Dalessio et al., “Model-based Control of the Resistive Wall Mode in
resistive wall mode feedback performance using optimal control DIII-D: A Comparison Study,” Fusion Eng. and Design, vol. 84, pp.
techniques,” Nuclear Fusion, vol. 47, no. 9, pp. 1157–1165, sep 2007. 641–645, 2009.
[30] Chapman et al., “Sawtooth control using electron cyclotron current [56] Wehner, Barton, and Schuster, “Toroidal Rotation Profile Control for
drive in ITER demonstration plasmas in DIII-D,” Nuclear Fusion, the DIII-D Tokamak,” in Proc. of 2015 American Control Conf.,
vol. 52, no. 6, p. 63006, 2012. Chicago, IL, USA, July 1-3 2015, pp. 3664–3669.
[31] Lennholm et al., “Feedback control of the sawtooth period through [57] Wehner, Barton, and Schuster, “Combined rotation profile and plasma
real time control of the ion cyclotron resonance frequency,” Nuclear stored energy control for the DIII-D tokamak via MPC,” in Proc. of
Fusion, vol. 51, no. 7, p. 073032, jul 2011. 2017 American Control Conf., Seattle, WA, USA, May 24-26 2017,
[32] Goodman, Felici, Sauter, and Graves, “Sawtooth pacing by real-time pp. 4872–4877.
auxiliary power control in a tokamak plasma,” Physical Review Letters, [58] N. A. Uckan, “Confinement capability of ITER-EDA design,” Proc.
vol. 106, no. 24, 2011. of 15th IEEE/NPSS Symp. on Fusion Eng., vol. 1, pp. 183–186, 1993.
[33] Rapson et al., “Amplitude based feedback control for NTM stabilisa- [59] Schuster, Krstic, and Tynan, “Nonlinear Lyapunov-Based Burn Control
tion at ASDEX Upgrade,” Fusion Eng. and Design, vol. 89, no. 5, pp. in Fusion Reactors,” Fusion Eng. and Design, vol. 63-64, pp. 569–575,
568–571, 2014. 2002.
[34] Kolemen et al., “State-of-the-art neoclassical tearing mode control [60] ——, “Burn Control in Fusion Reactors via Nonlinear Stabilization
in DIII-D using real-time steerable electron cyclotron current drive Techniques,” Fusion Science and Technology, vol. 43, no. 1, pp. 18–
launchers,” Nuclear Fusion, vol. 54, no. 7, p. 073020, jul 2014. 37, January 2003.
[35] Hennen et al., “Real-time control of tearing modes using a line-of- [61] Gouge, Houlberg, Attenberger, and Milora, “Fuel source isotopic
sight electron cyclotron emission diagnostic,” Plasma Physics and tailoring and its impact on ITER design, operation and safety,” Fusion
Controlled Fusion, vol. 52, no. 10, p. 104006, 2010. Technology, vol. 28, pp. 1–18, 1995.
[36] Felici et al., “Integrated real-time control of MHD instabilities using [62] Boyer and Schuster, “Nonlinear burn condition control in tokamaks
multi-beam ECRH/ECCD systems on TCV,” Nuclear Fusion, vol. 52, using isotopic fuel tailoring,” Nuclear Fusion, vol. 55, no. 8, p.
no. 7, p. 074001, jul 2012. 083021, 2015.
[37] Maljaars and Felici, “Actuator allocation for integrated control in [63] Hawryluk, Eidietis, Grierson, and Hyatt, “Control of plasma stored
tokamaks: architectural design and a mixed-integer programming energy for burn control using DIII-D in-vessel coils,” Nuclear Fusion,
algorithm,” Fusion Eng. and Design, oct 2017. vol. 55, p. 053001 (9pp), 2015.
[38] Eidietis et al., “Implementing a finite-state off-normal and fault re- [64] Pajares and Schuster, “Nonlinear Burn Control Using In-vessel Coils
sponse system for disruption avoidance in tokamaks,” Nuclear Fusion, and Isotopic Fueling in ITER,” Fusion Eng. and Design, vol. 123, pp.
vol. 58, no. 5, p. 056023, may 2018. 607–611, 2017.
[39] Vu et al., “Tokamak-agnostic actuator management for multi-task [65] Boyer and Schuster, “Nonlinear control and online optimization of
integrated control with application to TCV and ITER,” Fusion Eng. the burn condition in ITER via heating, isotopic fueling and impurity
and Design, vol. 147, p. 111260, oct 2019. injection,” Plasma Phys. and Controlled Fusion, vol. 56, no. 10, p.
[40] Blanken et al., “Real-time plasma state monitoring and supervisory 104004, 2014.
control on TCV,” Nuclear Fusion, vol. 59, no. 2, p. 026017, jan 2019. [66] Graber and Schuster, “Tritium-concentration requirements in the fuel-
[41] Kong et al., “Control of neoclassical tearing modes and integrated ing lines for high-q operation in iter,” in Proc. 46th European Physical
multi-actuator plasma control on TCV,” Nuclear Fusion, vol. 59, no. 7, Society Conf. on Plasma Physics, Milan, Italy, 2019.
p. 076035, jul 2019. [67] Pajares and Schuster, “Robust nonlinear burn control in ITER to handle
uncertainties in the fuel-line concentrations,” Nuclear Fusion, vol. 59,
[42] Pajares et al., “Integrated current profile, normalized beta and NTM
no. 9, p. 096023, jul 2019.
control in DIII-D,” Fusion Eng. and Design, vol. 146, pp. 559–562,
[68] Maraschek et al., “Path-oriented early reaction to approaching disrup-
sep 2019.
tions in ASDEX upgrade and TCV in view of the future needs for
[43] Peeters, “The Bootstrap Current and Its Consequences,” Plasma Phys.
ITER and DEMO,” Plasma Physics and Controlled Fusion, vol. 60,
and Control. Fusion, vol. 42, pp. B231–B242, 2000.
no. 1, p. 014047, nov 2017.
[44] St. John, “Equations and associated definitions used in ONETWO.”
[69] De Vries et al., “Requirements for triggering the iter disruption
[Online]. Available: http://fusion.gat.com/THEORY/onetwo/
mitigation system,” Fusion Science and Technology, vol. 69, no. 2, pp.
[45] Hinton and Hazeltine, “Theory of Plasma Transport in Toroidal
471–484, 2016. [Online]. Available: https://doi.org/10.13182/FST15-
Confinement Systems,” Rev. of Modern Physics, vol. 48, no. 2, pp.
176
239–308, 1976.
[70] ——, “Survey of disruption causes at JET,” Nuclear Fusion, vol. 51,
[46] Ou et al., “Towards Model-based Current Profile Control at DIII-D,” no. 5, p. 053018, apr 2011.
Fusion Eng. and Design, vol. 82, pp. 1153–1160, 2007. [71] Olofsson, Humphreys, and Haye, “Event hazard function learning and
[47] Witrant et al., “A Control-oriented model of the Current Profile in survival analysis for tearing mode onset characterization,” Plasma
Tokamak Plasma,” Plasma Phys. and Controlled Fusion, vol. 49, pp. Physics and Controlled Fusion, vol. 60, no. 8, p. 084002, jun 2018.
1075–1105, 2007. [72] Punnekkat, “Schedulability Analysis for Fault Tolerant Real-Time
[48] Felici et al., “Real-time physics-model-based simulation of the current Systems,” Dissertation, University of York, June 1997.
density profile in tokamak plasmas,” Nuclear Fusion, vol. 51, no. [73] Jianping Ma, “Applications of Fault Diagnosis in Nuclear Power
083052, 2011. Plants: An Introductory Survey,” Proc. of 7th IFAC Symp. on Fault
[49] Xu et al., “Ramp-Up Phase Current Profile Control of Tokamak Detection, Supervision and Safety of Technical Processes, Barcelona,
Plasmas via Nonlinear Programming,” IEEE Trans. Plasma Science, Spain,, June 30 - July 3 2009.
vol. 38, no. 2, pp. 163–173, February 2010. [74] De Vries et al., “Multi-machine analysis of termination scenarios
[50] Felici and Sauter, “Non-linear model-based optimization of actuator with comparison to simulations of controlled shutdown of ITER
trajectories for tokamak plasma profile control,” Plasma Phys. and discharges,” Nuclear Fusion, vol. 58, no. 2, p. 026019, Dec 2017.
Controlled Fusion, vol. 54, no. 025002, 2012.