Machine

Learning for
Physicists
University of Erlangen-Nuremberg
Florian Marquardt
Florian.Marquardt@fau.de
http://machine-learning-for-physicists.org

Part Three (Image generated by a net with 20 hidden layers)

Neural networks applied to
scientific tasks

Some examples
 Neural networks applied to scientific tasks:
General questions

• scientific goal and challenges (benefits of applying

machine learning?)
• input format
• output format
• number of training examples (how obtained?)
• accuracy of predictions (better than other machine
learning techniques? better than humans?)
• speedup vs. direct acquisition of outputs (from
experiments or computation)

Input Output
(format?) (format?)
databases?
experiments?
simulations? Assessment of
generated on the fly? accuracy?
 Quantum Chemistry

• Predict binding energy of molecules, orbital energies,

structure, etc.
• Success depends on input representation: “molecular
fingerprints”. One version used has been: 1/r for the
distances r between the atoms in the molecule.
• Use training data, e.g. from density functional theory (DFT)
or Hartree Fock calculations.
• Errors achieved are comparable to error of QM numerics
used to construct the training data.
• “Delta Learning”: Use results from less expensive methods
(DFT) as input to predict results that would be obtained
using more expensive methods.
• Typical errors about 0.1 eV
(Image: Wikipedia)

https://arxiv.org/pdf/1701.04503.pdf
 Computer-aided drug design

• drug discovery (activity with respect to some target

molecule, toxicity)

• Databases exist containing the properties of 100 million

compounds (PubChem) and 100 thousand protein
structures (Protein Data Bank)
• Databases list the experimentally determined activity of
millions of compounds in 100s (or more) of different
experimental tests [some are proprietary, held by
pharmaceutical companies]
• A typical task may be to predict whether there will be a
certain kind of reaction (with some target), given an
encoding of the molecule as input
• Deep neural networks recently won challenges for
predicting activity and toxicity of chemical compounds
• “multi-task” nets: trained on predicted multiple properties;
work better because they seem to build a more useful
internal representation https://arxiv.org/pdf/1502.02072.pdf
https://arxiv.org/pdf/1701.04503.pdf
 Computational structural biology

• Protein folding: given the structure, how does the final

shape look like? Extremely challenging (molecular dynamics
possible, but too slow for larger structures)
• One particular task: “Given two parts of the sequence that
are apart, will they be close to each other in the folded
structure?” (“protein contact prediction”)

(Image: Wikipedia)
https://arxiv.org/pdf/1701.04503.pdf
 Materials research

• Find better batteries, better solar cells, materials for

storing molecules (like absorbing CO2)
• in general: predict crystal structures and properties
(electronic, melting point, heat conductvity, magnetic, ...), of
novel/untested materials, or under extreme conditions
• “Materials Genome Project” and similar large-scale efforts
to collect various databases (of real materials that have
been synthesized at some time, and also of computed
materials, and of extrapolated/predicted properties)
• Functional materials (e.g. solar cells) vs. structural materials
(e.g. steel). The latter depend on how they are processed:
very difficult to model!

https://arxiv.org/pdf/1701.04503.pdf
 Statistical physics: phase transitions
(finding the transition point? discovering new phases?)

disordered phase ordered phase

e.g. use image recognition to learn to

distinguish between the two phases
(supervised learning)
 Statistical physics: phase transitions
• Classifying phases of matter, e.g. by looking at images (of
the magnetization, of wave functions), or other extracted
quantities
• supervised and unsupervised approaches

• Example: Ising model. Predicting the temperature by

looking at magnetization patterns in an Ising model will
yield network that has different weights associated with T
below and above the phase transition. Note however that
almost any statistics of the magnetization pattern would be
able to distinguish T below and above! https://arxiv.org/pdf/1609.09087.pdf
• Example: Disordered systems. Transitions between
types of quantum states in disordered systems (localized
to delocalized, and Chern insulator to Anderson insulator).
Teach the network to distinguish localized from delocalized
wave functions. Can be used to search the transition
automatically at other parameter values (e.g. other
energies), but simple measures for localization exist
already. https://arxiv.org/pdf/1610.00462.pdf
How to find the transition point?

Teach the network to distinguish between two phases from

the high- and low-temperature cases, then check output
near the transition point
true phase: true phase A true phase B
prescribed type: “type A” “type B” control parameter
(e.g. temperature)

Carrasquilla, Melko, Nature Physics 2017

How to find the transition point?

Teach the network to distinguish between two phases at an

arbitrary artificial cut: works best if the cut actually
coincides with the true phase transition point!
true phase: true phase A true phase B
prescribed type: “type A” “type B” control parameter
(e.g. temperature)
artificial cut
fraction of samples quantified correctly

position of arbitrary cut imposed by hand

Nieuwenburg, Liu, Huber, Nature Physics 2017
 Surface code quantum error correction
Surface code: encode two logical qubits into a lattice of physical qubits
Given a ‘syndrome’ (detected anomalies), what might be the
underlying chain of errors in the qubit lattice? [for correcting them]

Train restricted Boltzmann machine

to turn syndrome S into error chain e

Torlai, Melko Physical Review Letters 2017

 Artificial Intelligence for
scientific discovery?

“Can the computer think like a scientist

and discover laws of nature?”
Let’s consider an instructive but
still restricted setting in physics,
which I will call:
“Kepler’s World”
• N particles, attracting each
other via the gravitational force
• Computer (or some super-
smart and very patient student)
is allowed to do numerical
experiments and try to learn
from them
Energy

1 2 3 4 Particles
 “All trajectories are linear!?”
Energy

1 2 3 4 Particles
 “All trajectories are unbounded!?”
Energy
“All trajectories have E>0!?”

1 2 3 4 Particles
 “All E<0 trajectories are elliptic!?”
Energy

1 2 3 4 Particles
 “All E<0 trajectories are bound!?”
Energy

1 2 3 4 Particles
Energy “For E<0, at least two particles will
end up bound together!?”

1 2 3 4 Particles
 “Kepler’s World”
...one might discover:
• Conserved quantities (total energy, momentum,
angular momentum)
• Two particles at negative energy: bound, on elliptical
orbits, relation between distance and period
• Equations of motion with inverse square force
• Approximate elliptical orbits if many light-mass
particles travel around one heavy mass
• “Moons”
• Properties of scattering: energy and momentum
exchange
• 3-particle scattering can lead to expulsion
• Effects of “resonances” between orbital periods
• Rings
• Chaotic behaviour
• Structure formation in clouds of many particles
• Local Boltzmann distribution
• Evaporation
• ...things as yet unknown...?!
 Hypotheses

hypotheses
 Hypotheses

After first observation:

 Hypotheses

After second observation:

 Hypotheses

After third observation:

 Hypotheses

Observations act as
filters (sieves)
 Hypotheses

Problem: there are infinitely many hypotheses, and we

cannot enumerate all of them (unlike in the picture). We
have to generate them and select them for consideration.
Also, keep even disproven hypotheses in order to modify
them later, to turn them into viable hypotheses again (e.g.
by restricting the suggested domain of validity, or by
slightly changing the claim).
 Hypotheses
(falsifiable claims about
relations between quantities)
Quantities about which statements can be made:
• originally accessible quantities (trajectory,
parameters such as masses and number of particles)
[in principle, all hypotheses can be formulated
entirely in terms of those, but it may be
cumbersome]
• derived quantities:
• continuous numbers and fields (distances, energy,
force fields, average velocity squared, ...)
• integer quantities (e.g. number of particle pairs
with negative pair-energy, number of local maxima
of x3(t), winding number of a particle pair
connection vector evolving in time, ...)
• Boolean quantities (e.g. “does x3(t) rise
monotonically?”, “is the energy below zero?”, “does
the trajectory lie on an ellipse within numerical
accuracy?”, ...)
 Hypotheses
(falsifiable claims about
relations between quantities)

Relations may be:

• precise statements, e.g.
• E(t)=E(0)
• “trajectory lies on an ellipse if N=2 and E<0”
• statements about limiting cases:
• e.g. Taylor-expansion: “x3(t+dt)-x3(t)=dt*vx3(t) in
the limit of small dt”
• statistical hypotheses: “for uniformly distributed
initial coordinates within some small interval, the
velocity distribution at the final time is a Gaussian,
with a spread given in the following way in terms of
the initial energy: ...”
• quantitative statements where one may quantify the
deviation?
 Hypotheses
(falsifiable claims about
relations between quantities)
• Hypothesis becomes more plausible with every test that
it could have failed (but did not)
• Generate primarily hypotheses that are not obviously
wrong [do not contradict known facts] but also not
obviously right [do not follow logically from some very
well-tested hypothesis]!
• It may still be useful to generate (by logic/math) new
statements out of well-tested ones, if the new ones are
simpler and easier to apply [e.g. energy conservation out
of eqs. of motion]
• sometimes it is useful to consider hypotheses that are
found false overall, but approx. true in some domains
• Look for cases where the hypothesis can be easily tested
(maybe far away from all those cases where it was
successful)
• Consider testing a hypothesis especially if it were to have
many consequences
• Prefer ‘simple’ hypotheses (“Occam’s razor”)
Learning
(discovering new, ever more
general hypotheses)

observations
(experiments)
Learning
(discovering new, ever more
general hypotheses)

current
hypotheses fail
here
 Learning
(discovering new, ever more
general hypotheses)

unknown
(untested)

known, but not

understood
(current
hypotheses fail)

known and
understood (so far)
 Learning
(discovering new, ever more
general hypotheses)

where to perform the next observation?

(conduct the next experiment?) ?
?

?
 Learning
(discovering new, ever more
general hypotheses)
Active learning: learner/scientist selects
observations to be made (selects “training data”)
• at the border “understood” vs. “not understood”
• or, more precisely: where one might learn the most
• i.e.: parameter values where the uncertainty of
predictions is very large (e.g. existing hypotheses
leave a large freedom in outcome), and where
refining them would have significant consequences
in many areas of parameter space
• also: parameter values where many hypotheses
may be falsified (based on experience? based on
“distance” to known examples?)
• try to falsify (or make more plausible) especially
those hypotheses that are most useful, because
they are:
• most simple and general
• can be applied usefully if true (e.g. to obtain a
new very sensitive measurement method, or to
simplify predictions)
 Learning
(discovering new, ever more
general hypotheses)

Select (numerical) experiments not only according to

possible reward (new insights), but also according to
effort! [have to learn to predict effort]

Experiments: Time [1nsec - several months], costs,

energy consumption, component requirements (for
fabrication, setup, measurement time, calibration)
Numerical experiments: Time [1 µsec - a week],
number of cores, energy consumption
Hypothesis generation steps: e.g. effort needed for
logical/mathematical deductions, for running a neural
network etc.
Distilling analytical laws from experimental data
M. Schmidt and H. Lipson, Science 2009
“Adam”, the Robot Scientist King et al., Science 2009
• Generates and tests hypotheses about the role of various
genes in yeast (“functional genomics”)
• time-consuming tests: select strains of microbes (with
certain genes modified) and let them grow on selected
media over several days
• Robot can initiate about 1000 strain-media combinations
a day
• “Adam” consists of three liquid-handling robots, three
robotic arms, three incubators, a centrifuge, a freezer, and
more equipment; plus a computer to control everything
• model=collection of facts about genes, enzymes and
metabolites (small molecules)
• hypothesis forming: which gene might encode an enzyme?
[take clues from other organisms]; e.g. “2-aminoadipate
transaminase” might be encoded by either YER152c,
YJL060w or YJL202w genes
• assign probabilities to hypotheses, select experiments that
minimize the expected cost for eliminating all but one
hypothesis
• found some new gene-enzyme relations
 “Physics 101”: Learning physical properties
from unlabeled videos
Wu et al., “Physics 101 – Learning Physical Properties”, BMVC 2016

Inclined plane

spring

fall

drop into
water

>17000 video clips

“Physics 101”: Learning physical properties
from unlabeled videos
Wu et al., “Physics 101 – Learning Physical Properties”, BMVC 2016
How might science look like in the future?

• Right now: Machine learning / neural networks as a new

tool, for making computationally cheap approximate
predictions based on large training sets, or to discover
hidden underlying structures (e.g. based on the image
recognition power of neural networks)
• Later, for true scientific discovery AI: form human-
computer teams, where the human could focus on
selecting goals
• Important additional task for AI: Train the computer to
explain conclusions in a way that is optimized for human
understanding
• Very long-term, once we have superhuman scientific AI: ?
 Fu
re nct
pr io
es n/
en Im
tat ag
i o e
[H Im n
an ag
dw e
rit clas
ing si
Co re ficat
nv co io
olu gn n
Au t i on
itio
a n]
to ln
en et
dim V c od s
er
en isua
sio iza l s
na tio
lr n
Re ed by
cu uc
tio
rre
nt n
ne
W tw
or or
dv ks
Re ec
inf to
or rs
ce
me
Co nt
nn lea
ec rn
Ar t ion ing
tifi st
cia o
l In ph
ys
te i cs
llig
en
ce
http://machine-learning-for-physicists.org