Kenneth G. Wilson Award colloquium 2020

"Physics & Astronomy"
21 followers
Very good, the meeting's being recorded. This award is really an absolute honor and it really is very,
very humbling. It also gives me a nice opportunity to say a quick thank you to everyone in our
community who makes this such a fun area to work in, and especially to my mentors and my
collaborators over the years as I found my way into this community, starting from the Adelaide
group, Ross, James and Tony, to my wonderful collaborators in NPL QCD, and especially Will,
who is just so much fun to share an overlapping group with at MIT, and of course John and my
group of absolutely awesome students and postdocs whose faces you'll see on some of my slides.
Okay, so before I start, the brief for this book was a colloquium to a broad demographic,
but with a reality of a mostly Lattice audience. So I have to start by apologizing to the Lattice
crowd for the extended introduction, and apologizing to everyone else for a few
extra technical details that I let slip into some of the slides for the experts in the room.
So the topic of this talk is, broadly, the structure of matter, and in particular the
structure of the protons, neutrons and nuclei that we know make up the majority of the visible
matter in the universe. Our modern understanding is, of course, that the protons and the neutrons
in an atomic nucleus have substructure. They're made of matter particles called quarks, held,
glued, bound tightly together by the strong force manifesting as force-carrying particles known as
gluons. And to the best of our current understanding, quarks and gluons are it,
they are fundamental particles, and they can't be divided any further. The quarks and the gluons and
their interactions are encoded in the standard model of elementary particles. Now this framework,
which was put together in the 1970s, has just 17, you can see them here, just 17 fundamental
particles, and it describes three of the four fundamental forces, so notably excluding gravity,
in this very elegant and beautiful, very symmetric, very predictive framework.
And in fact, a number of the standard model particles were predicted before they were
seen experimentally. Most recently, I think many of us will remember the Higgs boson in 2012.
Now this framework is an extraordinarily successful description of the matter in our
universe, which has withstood almost every experimental test. And just to give some
sense for that, here's an example. The magnetic moment of the electron, that's how much torque
an electron fuels in a magnetic field. What I'm showing here is the anomalous magnetic moment,
the deviation from the tree level or classical result. This is the most accurately verified
prediction in the history of physics. And for any graduate students in the room, you'll understand
the heroic effort that's needed for up to five loops of Feynman diagrams. That's 30 years of
to calculate these 12,000 Feynman diagrams to give this result to 12 significant figures that
agrees between theory and experiment. But it's not just that the standard model is successful
at sort of certain key predictions, it's that it's successful over such a wide range of scales.
So here's just one other example in deep inelastic scattering. So that's scattering
where you take an electron and fire it at a proton so hard that the proton breaks apart.
This is encoded in what are called structure functions, shown here in this chart.
On the horizontal axis, we have energy on a logarithmic scale. So this covers many orders
of magnitude here, this figure. You can see a consistency between the experimental results,
the points here, and the theory of predictions, which are the yellow band.
But of course, we also know that this extraordinarily successful and predictive
theory doesn't describe everything we observe in our universe. Just one example that we'll
get back to later in the talk is dark matter. We in fact know that the matter described by the
standard model is just this tiny sliver of the energy density pi, and that most of the matter
in the universe is something else. We also of course know that neutrinos have masses,
but they don't in the minimal standard model. There's an asymmetry between matter and antimatter
that we can't explain. How does gravity fit in? There are naturalness questions,
why is there this big discrepancy between the weak and gravitational scales?
So clearly there is physics beyond the standard model, and we're always seeking to
constrain that better to understand the physics that we don't know.
And just one class of experiments searching for physics beyond the standard model is particularly
relevant to the discussion of this talk, and that's experiments at what we call the intensity
frontier. That's experiments where we're looking to probe the rarest standard model interactions
and look for beyond standard model effects. So things like dark matter direct detection
fall into this category, neutrino physics, many other examples. And what these experiments have
in common is that because we're looking for these really rare interactions for signs of new physics,
you want to increase the cross section as much as possible. So they use nuclear targets.
And so to understand physics beyond the standard model to constrain it in these experiments,
we also need to understand the standard model physics of the protons and the nuclei
in the targets in those experiments. So when I'm discussing understanding the structure of matter,
of protons and of nuclei from quarks and gluons, there's really several reasons. The first is for
its own sake to understand the emergence of all of the complexity in nature from this beautiful,
elegant theory of the standard model. And the second is that since we know the standard model
isn't everything, we need to understand it better to constrain searches for new physics.
So this really lines up with the award citation and also forms the outline for this talk. I'll
give some examples of first principles theory studies to study the structure of the proton
and of nuclei, and also how advances in computing and algorithms are accelerating and enabling this
So a little bit more introduction. The hard part of this is, of course,
the study of the strong interactions described by the theory of quantum chromodynamics.
The strong interactions are the strongest of the four forces in nature. And of course,
they're responsible for binding quarks and gluons into protons and neutrons and binding protons and
neutrons into nuclei. So if we're interested in the structure of those protons and nuclei,
the strong force is key. But studying it is complicated by the fact that at low energies,
so the energy scales relevant to the structure of, say, a nucleus at rest,
the strong interactions are non-perturbative. So here we have the strong coupling constant
on the horizontal axis is energy. And as we come to the left here, to low energies,
the coupling strength grows. And what that means, of course, is that you can't expand
in powers of that coupling strength. You can't do a perturbative solution.
So we need to do something non-perturbative. And that's where lattice field theory comes in
as the only known systematically improvable way to study the strong interactions from first
principles. And the idea is quite straightforward to say, less straightforward in practice,
that we take the equations of QCD and write them for a discrete spacetime grid. And then the path
integral becomes essentially integrals over the values of the quark and gluon fields at each site
or link on the lattice. So this is a very high dimensional integral, something like 10 to the 12
variables. And we evaluate that integral by importance sampling, which is, again, easy to
understand. You can just think about quantum mechanics, if you like, that we know that
paths near the classical action dominate. And so we can sample those proportionately more
than the less important paths. So to put it in just a little bit more technical detail,
we have a Euclidean spacetime lattice with, of course, some non-zero lattice spacing and some
finite volume. And you have to take the limits. As the spacing becomes small, the volume becomes
large to match to continuum QCD. And some calculations, and some of the ones I'll talk
about, use larger than physical quark masses simply because it's computationally more efficient.
You have to invert a lot of matrices that get close to being singular as the quark mass gets
wider. And then here, if we just write out the expectation value of some observable,
here's the QCD path integral with our integrals over quark and gluon fields.
In this important sampling approach, we sample the field configurations with this weighting,
the exponential of the action built in. And then computing observables becomes just taking,
essentially, averages and standard deviations over that samples. And some of the challenges
in this sampling are what questions we'll come back to at the end of the talk.
So just to summarize, really, in lattice QCD, the input parameters, the free parameters,
are the same as the free parameters in QCD. That's the quark masses and the strong coupling.
And once those are fixed by matching to some of the measured hadron masses,
and with one additional experimental input to fix the lattice spacing of a strong coupling,
then everything else, from the details of the structure of the proton to the structure and
interactions of nuclei, are predictions from the theory. There's no additional model.
And these calculations are really computationally expensive, as you might imagine,
from the very large dimensionality of the integrals. They use many millions of CPU and
GPU hours just to give some scale. Some of the calculations I'll describe would have taken
something like 30,000 years on my laptop. Here's SUMMIT, currently the world's largest supercomputer.
And if every person on Earth did one computation every second for a year,
that's about what SUMMIT can do in a second. So just for a sense of scale.
And of course, as the lattice field theorists in this room very well know, lattice QCD works.
And we're really in a precision era of lattice field theory for simple quantities,
for properties of single hadrons, for example. Over the last five years or so,
we've really seen examples of the ground state hadron spectrum being reproduced,
a nice calculation of the proton-neutron mass difference, including both QCD and QED.
And so you can really see in a theory calculation how the competing effects
of the strong interaction pushing that mass difference one day,
and the electromagnetic effects pushing it the other way, cancel to give us this
small but critically important mass difference in nature.
Predictions for new states with controlled uncertainties. For example, here, all of the
blue points are states predicted from lattice QCD. The red lines are from experiment. And this
additional red line, state found by LHCb after it had been predicted. And of course, when we talk
about more complicated quantities, such as quantities for nuclei, we're not in this precision
era yet. And that's because nuclear physics in this framework is really just computationally
very hard. The noise, the statistical uncertainty grows exponentially with the number of nucleons
in the calculation. The complexity grows factorially. So we have these compounding
exponential and factorial challenges. And so calculations at this stage are possible for
atomic number A less than about five. So the nuclear physics calculations I'll talk about
are done with the NPO QCD collaboration. Of course, there are also other collaborations
around the world also working on lattice QCD studies of nuclei. And I'll only have time to
briefly highlight a couple of latest results. Okay, so after the extended introduction,
coming back to the first topic, the emergence of the complex structure in nature,
that is understanding the quark and gluon structure of matter. Now, of course, I'm discussing this from
a theory perspective, but there's a great synergy between experiment and theory, which we'll see
some examples of. And understanding the quark and gluon structure of matter really is a defining
effort in experiments, with experiments around the world having the same goal.
So one place to start is talking about the three-dimensional quarkonic, that is the quark
and gluon structure of the proton. And there's a really nice community overview this year of
lattice QCD contributions to this area. And I'll give just a few examples, primarily from my own
recent work. So if we're thinking about the quark and gluon structure of the proton, we encode that
structure in a number of distributions. So one of the simplest we can think about is the parton
distribution functions here. These Fs, these are the number densities of quarks or gluons of some
flavor F in the hadron H, say the proton, carrying some longitudinal momentum fraction x, so some
fraction of the longitudinal momentum of that proton. If we also consider the transverse structure,
so add in our additional transverse scale, we call them transverse momentum-dependent parton
distribution functions. And we can also allow a momentum transfer, so this Xi is a skewness,
it's a longitudinal momentum transfer, we get generalized parton distribution functions.
And I'll talk about some aspects of each of these. So starting with the PDFs, the parton
distribution functions sound complicated, but the moments in particular encode some very key and
intuitive aspects of hadron structure or proton structure. So the moments are these x-weighted
x intervals, where again x is the longitudinal momentum fraction carried by that parton. And
the lowest moment describes the contribution of each parton to the momentum of the hadron.
And there's this nice highlight this year of the proton momentum decomposition by Dean
Alexander and collaborators where they've calculated each contribution up, down, strange,
charm, and gluon contributions to the proton momentum and see that they add up to one as
of course they must. And LatticeQCD can cleanly access low moments of parton distribution functions
and it would be remiss of me to mention PDFs without saying that of course there's massive
community efforts to also study the x dependence, that momentum dependence, from LatticeQCD with
various approaches. But coming back to the moments for the moment, we can also use
LatticeQCD results for the moments and include those constraints to global fits to experimental
data. So again the synergy between theory and experiment to better understand the parton
distribution functions themselves. And there's this nice example from a few years ago showing
this exactly for the transversity parton distribution functions which are less well
constrained from experiment especially at large x, so large longitudinal momentum and
the flavor separation is not complete. And just by including the lowest moment of the
parton distribution functions calculated in LatticeQCD, you go from these yellow bands
to these smaller blue and red bands. So adding that one number gives you a much better constraint
on these parton distribution functions and here the fragmentation function,
which you can think of as being the final state analog of the initial state final parton
distribution function. So the probability that the hard parton called i here will fragment
into some hadron carrying some fraction z of parton momentum. So one of my interests is extending this
great success of LatticeQCD to nuclei. So of course we can also ask this about nuclei and
it's one of the simplest and most intuitive questions you might ask about a nucleus
is how is the partonic structure, that's the distribution of quarks and gluons and nuclei,
different from that of a collection of nucleons. So of course we know that a nucleus is not just
a collection of nucleons stuck together. We can call these sorts of effects, these nuclear effects,
EMC type effects in the honor of the European Muon collaboration who saw in 1983 that deep
inelastic scattering of a nucleus doesn't just look like the sum of deep inelastic
scattering of the constituent nucleons. So here's the DIS cross-section ratio per nucleon for iron
to deuterium versus this longitudinal momentum fraction x and you see that this is clearly not one.
So evidence of nuclear effect. So we can start to get at the EMC type effects from LatticeQCD
by looking again at these simple quantities where we have such success in calculating them
precisely for protons in the lattice field theory approach. So we can look for example
at the momentum fraction carried by quarks of different flavors in light nuclei. So this is
recent work in the NPOQCD collaboration where we've calculated the up minus down
on quark contribution to the momentum fraction in helium-3. So this calculation was done at
larger than physical quark masses corresponding to a pion mass of about 800 MeV then we had to
extrapolate by constraining the low energy constants of effective field theory and extrapolating
physical quark masses but after you do this you still have a result here with the red band
which is a constraint on this momentum fraction. The best experimental, the global fits to
experimental data from the NNPDF collaboration is this blue, light blue ellipse here. So you can
include the LatticeQCD constraint just like we saw for the proton into this global experimental
fit and improve the constraints and go to this purple ellipse here. And so even though we're
really in the very beginning of these sorts of studies for nuclei we can already see that
there's great potential for the sort of success we've seen in the study of single hadrons to
extend. We can then also do the same trick of course what about including this constraint
on the moments to see what you can do for the PDFs themselves and you can see that even this
constraint gives us an improved path on distribution function, this is actually the
ratio of the path on distribution function of helium free to that for the proton, an improved
constraint by including the theory result into the global fits to experimental data. We can come
back to this discussion of proton structure and talk not just about the longitudinal structure
but what if we also think about the transverse structure and of course there's been a lot of
work in the community particularly by Michael Engelhardt and collaborators to compute TMDs
themselves from lattice QTD. I'll talk just about one calculation done with Mike Wagman
and Yong Zhao of the Collin-Soper evolution kernel. So this is a quantity that governs the
rapidity evolution of these TMDs. It's governed for dependence on this hard scale which is related
to the momentum of the hadron or equivalently to the hard scale of the scattering process.
And this evolution is of course needed to match TMDs measured at different scales. It's also
needed if you're trying to match quasi-TMDs so that's the framework we can use for accessing
TMDs, one of the frameworks we can use for accessing them on the lattice to physical TMDs.
This quantity is perturbative at short distances but it becomes non-perturbative
at longer distances and so of course that makes it a key target for the non-perturbative tools we
have in lattice field theory. So it's been shown that you can access this Collin-Soper evolution
kernel via a ratio of non-local matrix elements and this is where I'm getting a bit technical
for a physics colloquium. And some examples of these non-local matrix elements, these quasi
beam functions are shown here on the right of the slide. And so we followed this calculation
through in the first quenched calculation this year. Since this kernel is independent of state
we could study a pion and since it's a quenched calculation we could even study an unphysically
heavy pion with no systematic effect to calculate this evolution kernel. And that's what I'm showing
here and here's this transverse scale on the horizontal axis. These lines here are the
perturbative results. They become singular about here because they reach the Landau pole
and what you can see is that we have reasonably precise results up much further than the
non-perturbative results can reach. And if you do a comparison between these results
to phenomenological models, and of course this deserves a significant word of caution, this is
just a photoshop overlay of phenomenological models which are of course not quenched, so not
zero flavor, with the zero flavor lattice calculations. What you can see is that systematically
controlled results at this position would already impact phenomenology. So this is preliminary work
with again lots of promise in this area. And then we can come back and say, well what about the
generalized parton distribution functions? And again I can give a nice example here.
If we think again about, like we thought about the moments of the parton distribution functions,
we can think about the moments of the generalized parton distribution functions.
And what these describe are generalized form factors which encode for example matrix elements
of the energy momentum tensor. So the moments of the generalized parton distribution functions
give us these things here, these generalized form factors. What I'm showing here are matrix elements
of the gluon traceless energy momentum tensor in a nucleon state. And if you decompose those
matrix elements just into its Lorentz structures, you have these a priori unknown coefficients which
you can relate to the moments of the generalized parton distribution functions.
To give some physical intuition, if we look in the forward limit, so where we have no momentum
transfer, so when p and p prime here, your initial state proton and your final state proton, have the
same momentum, then these moments reduce just to the moments of the parton distribution functions
that we saw before. So this a here with a zero momentum transfer, which is p, p equals zero,
just looks like the momentum function. So how much momentum is carried by the quarks and gluons in
the proton. The a and b are related to the spin. And then we also have the d, which are called the
d terms, which are largely unknown but equally fundamental. And then if we include this
generalized part, the momentum transfer part, and look at the t dependence, then these d term
generalized form factors, you can think of those in a nice way. You can think of them as encoding
pressure and shear distributions in the proton. And so this is an example where we can have a
nice physical picture. So quarks are of course confined inside protons. So if you think you
have a pressure distribution, then any expanding forces have to be cancelled by compressive forces
for stability. So we can draw a cartoon. What I'm showing here is the radial distance out from the
center of the proton, that distance squared times the pressure. And here on the horizontal axis,
that radial distance out from the center of the proton. And you can see that this is the quantity
that has to integrate to zero to stability. So if we make a cartoon and say, there's a repulsive
pressure at the core of the proton, balanced by a confining pressure, as you move out further along
the radius, we can ask some questions about this distribution, such as, how do quarks and gluons
contribute? What's the radial extent of the pressure distribution? What's the peak? Is it
more like a bike tire that's 10 to the 5 pascal? Or is it more like a neutron star
at 10 to the 34 pascal? What's the pressure inside the proton? And so this has been particularly
interesting in just the last couple of years, since there was a first experimental extraction
of the pressure distribution in the proton. And what they found in that study was that the peak
pressure near the center of the proton was, in fact, greater than that, an order of magnitude
greater than that at the interior of a neutron star, so the densest known object in the universe.
But the experiment, through deeply vertical confidence scattering, was only sensitive to
the quark contributions to the pressure and not to the gluons. So to make these conclusions about
the proton's pressure distribution, one had to assume that quarks and gluons contribute to the
same way. So what we were able to do was compute the gluon generalized form factors, and hence the
gluon contribution to the pressure distribution, and see that gluons change this picture, and also
to really explore what will be needed in future experiments, like an electron-ion collider,
to measure the gluon contribution to the pressure from experiment. So what you see here is the same
as the cartoon I just showed. This is that radial distance squared times the pressure distribution
on the horizontal axis, the radial distance from the center of the proton. The blue here is what
you get from the experimental result, assuming quarks and gluons contribute the same way.
And in red, you have the portal taking the quark contributions from experiment and the gluon
contributions from lattice energy and putting them together for a first complete pressure
determination. And you see that gluons change the picture. The peak is shifted, we have the tail is
extended by including gluons. Gluons are, of course, important in the pressure in the proton.
So we can also ask, well, how does this pressure distribution compare to that in other hadrons?
Can we get some intuition on what this means? So these pictures look a bit nicer because I'm
showing just lattice results with a heavier than physical quark mass here, but it's the same idea.
This is the pressure in the proton. Here we have the pressure in the pion. And what you see is that
although if you look just at the simplest static quantities, like the momentum carried by quarks
and gluons in the proton versus the pion, in each case they're quite similar. Quarks and gluons carry
about 50% each of the momentum of the proton and also of the pion. If you go add the extra dimension
and look not just at the single quantity but at this distribution, you see they have quite
different pressure distributions. The one for the pion has a higher peak and a much shorter
radial extent. And so we can also ask, well, what about other hadrons? And so new studies
of the rho and delta at heavy quark masses and also the nucleon and pion at physical quark masses
will be coming soon, led by grad student Dimitri Pefkou and Dan Hackett, who's a postdoc at MIT.
We can also ask, of course, well, what about light nuclei? And maybe you might think that
calculations of pressure inside nuclei could be connected to appraisal of the states of neutron
stars in some way. And this is, of course, work that's also on the horizon. Okay, so we can also
come to this question, well, what about more directly the impact of these sorts of calculations
for searches for new physics? Of course, I should have said that all of the other results also have
implications for searches for new physics, for example, the parton distributions and the
colon superkernel at the large hadron collider. But let's talk more explicitly about the example
of dark matter. So, of course, there is abundant evidence for dark matter, from galactic rotation
curves to gravitational lensing to structure formation to evidence in the cosmic microwave
background. And we know that about five times as much of the energy density of the universe is
carried by dark matter as compared to normal matter that we can describe so successfully
with the standard model. There are many, many models for dark matter, but relatively few
constraints at this stage. I like to show this nice figure from Tim Tate, sort of trying to map
the various different theories and models for dark matter from axions to sterile neutrinos,
extra dimension, supersymmetry, little Higgs, and it goes on, right? And what we need to really
understand dark matter better is constraints in this space. Of course, there are many ways that
we try to constrain the properties of dark matter, and I'll mention just one, and that is in direct
detection experiments. And so this is where, of course, dark matter is dark. It doesn't interact
with light. It does interact through gravity. That's how we know it's there. But maybe it interacts
very weakly through the weak interactions. And then we might think that if we have a detector
deep underground, so it's isolated from everything else, we might be able to see
direct interactions of dark matter with our terrestrial detectors. So this is the dark
matter direct detection program looking for, say, weakly interacting massive particles, for example.
And then the detection rate, of course, depends on the properties of the dark matter, which is
what we're trying to constrain, and also the probability that the dark matter interacts with
nucleons. Okay, so these sorts of experiments, we don't have evidence for dark matter yet,
but you can set limits on the interaction strength of dark matter candidates with nuclei.
And what these limits look like is something like this in a figure, where on the horizontal
axis we have the mass of our proposed weakly interacting massive particle, and on the
vertical axis we have the limits on the cross-section for these weakly interacting massive
particles to interact with the proton or neutron, with the nucleon. And the way to read these plots
is that everything above these lines, each of these lines comes from some experiment. All of that
phase space above the lines is ruled out, and here below is the still available phase space
for dark matter to live. What I want to emphasize here is that these experiments all use large
nuclei. For example, here xenon 1 ton uses xenon, the nucleus with 130, 140 protons and neutrons.
But what we're showing here is the spin-independent, that's what SI means, WIMP nucleon cross-section.
That's the cross-section for interaction with a proton or neutron, which is of course what we'd
like to constrain. So what's happened implicitly here is there's been a translation between not
seeing dark matter hit xenon to what that means for its interaction strength with a single proton
and the simplest way to do that translation is of course to use the Born approximation,
and that's just the approximation that when your dark matter comes in and it hits the nucleus,
it hits a single proton or neutron at a time. So you just scale your interaction of your dark
matter candidate with a proton or neutron, this is something you can calculate precisely in
lattice QCD, you just scale that by the atomic number and then that gives you the cross-section.
But of course, well, what if the dark matter interacts non-trivially with multiple nucleons,
then you don't just want the interaction with a single proton or neutron, you also want this
matrix element for its interaction with multiple protons or neutrons, and the second term could
be significant and it could scale differently with Z and with A, and this can lead to significant
uncertainty. This is again an NPO QCD collaboration result where we studied the
spin-independent scattering of these WIMP candidates, of course we don't know the
microscopic theory of WIMPs, but the low energy limit of a generic spin-independent interaction
is scalar coupling to any quark flavour, so what's relevant are the scalar matrix elements,
and so we calculated in a calculation with a larger than physical value of the pion mass,
these scalar matrix elements in light nuclei, and what we saw was really larger than anticipated
nuclear effects. So what's being shown in this figure here is this quantity, the ratio
of the matrix element in a nucleus to A times the matrix element in the nucleon minus one,
so this is just zero if you have the Born approximation holding, so deviations from
zero here are what are revealing nuclear effects, and each of these columns is a different combination
of quark flavours, so here up minus down, here up plus down plus strange, and so on. And what you
see here for these green points, this is helium-3, you see nuclear effects around the 10% level.
So to put this in context, this is large, you can look at say the axial matrix elements
where we can measure them in nuclei, and we know that in medium mass nuclei that the nuclear
effects in the axial matrix element are tens of percent, but in helium-3 they're very very small,
they're the fraction of a percent, whereas here we have the scalar matrix elements having 10%
effects even in helium-3, and of course these calculations are not systematically controlled,
larger than physical value of the pion mass, but what this says is that these unexpectedly large
multi-body effects really shouldn't be neglected in the interpretation of dark matter direct
detection experiments, and it's really crucial to do a systematically controlled study, and so we
have a calculation at almost physical quark masses in progress on this one. So of course, I talked
about direct detection experiments with xenon, of course we would also love to know the matrix
elements and the QCD structure of larger nuclei, which are in many cases phenomenologically relevant.
So the way to constrain those is via nuclear effective theory, where one-body currents are
dominant, two-body currents are sub-leading but non-negligible, you'll note that I said
a two-body effect of 10% was sort of unexpectedly large, so that you still have a hierarchy,
and so the program is really to determine the one-body contributions from single nucleon
calculations, few-body contributions from calculations with two-body, three-body, four-body
states, and then match effective field theory and many-body methods to lattice results to make
predictions for larger nuclei. But of course we can also ask, well, what are the real barriers
to studying larger nuclei? And a key aspect is just computational, that we have exponentially
harder problems computationally, and so you'd like exponentially improved algorithms. And just
to give some examples of the real pace for trying to study and constrain matrix elements of large
nuclei, the DUNE long baseline neutrino experiment, which is looking to measure the neutrino masses,
the neutrino mass hierarchy really, and the oscillation parameters, is doing neutrino
scattering on argon, that's A equals 40, so you really want to know the actual form factors of
argon and how big the nuclear effects are on those. Looking for double beta decay, so for
telltale signs of whether the neutrino is its own antiparticle or not, the lightest nucleus that
undergoes double beta decay is A equals 48. I gave the example of the scalar matrix elements seen on
A equals 131, we'd love to constrain that. So there's real physics cases for trying to study
light nuclei and heavy nuclei. But we face computational challenges. So this brings me to the
last point, so how advances in algorithms can of course enable physics calculations that wouldn't
otherwise be possible. And I'll give just one example of something that my group's been focusing
on recently, is trying to accelerate the generation of QCD gauge field configurations. So the problem
is quite simple to state, that we want to generate these field configurations with a known probability
distribution given by the exponential of the action here. And what we typically do is we use
Hamiltonian or hybrid Monte Carlo. It's a scheme where we update from one configuration to the next,
it's a coupled walk in position and momentum space through Hamiltonian dynamics. And because it's an
update, right, from one configuration to the next, you have some correlations between configurations
in your chain. So you have some amount of burn-in time and then you're still generating correlated
configurations, so you want to sample every configuration where n is dictated by the Markov
chain autocorrelation time. And having a shorter autocorrelation time means less computational
cost to generate independent configurations. So a particular challenge with this is that these
updates are in a sense diffusive or close to local. So you can think that if you're trying
to update physics to generate independent configurations on some physical length scale,
say the size of a proton, then as the ladder spacing becomes fine, the number of steps you
need to take to decorrelate physics on that fixed scale diverges. So you need to do more and more
updates to affect the correlations on that scale. So this is a manifestation of critical slowing
down in this context, and this is a particular computational challenge. So if we come back to
looking at this problem, what we'd like to do is generate these field configurations with some
known probability distribution. So let's visualize this. In a scalar field theory, so just a number
at each ladder site, we can visualize this very nicely. We would like configurations that look
like this and not like this. So this is just random noise, this is very unlikely. These are
likely configurations, they have characteristic correlations on a characteristic length scale.
I should thank grad student Ted Canlar for some of these nice figures.
Generating configurations like that has some parallels with an image generation problem
that's received a lot of attention from the machine learning community and a lot of great success.
So there's been a lot of success in generating images like this. So these are not photos,
they are machine generated images of faces. You can go to www.thispersondoesnotexist.com
and refresh and it will give you a new fake person every time. And so there have been nice
algorithms developed that will generate this, but not this. But of course the key difference is that
we do want this configuration, just very rarely, with the probability exactly specified by the
exponential of the actin. Whereas here you never want random noise to be generated if you're trying
to generate pictures. So one way of trying to address this generation problem using machine
learning tools, like has been done for generating these images of faces, is via what are known as
flow-based models. And the idea is very, very simple. You're trying to optimize a change of
variables. If you have some distribution that's easy to sample from and you have a function,
a change of variables that takes you from that to your target distribution, then you've solved
your sampling problem. You just take samples from the easy to sample distribution, you put them
through your function, and then out pop the samples you're interested in. So if you also
design this function so that it's invertible and has attractable Jacobian, then even if this
change of variables is not exact, you can still guarantee exactness. You can show that you can
create an ergodic and balanced Markov chain out of these independent samples and guarantee
convergence to the desired probability distribution in the limit of a long Markov chain, just like
you do with HMC. And so you can create these very general changes of variables by composing
many simple transformations. In particular, if you compose simple layers acting only on parts
of the field each time, you can have a nice tractable Jacobian that's lower triangular
and easy to invert. And so then you can optimize this function, that optimization is the machine
learning part, to do your sampling. So it's worth noting that this is, applying this to
QTDs, unfortunately not at all an out-of-the-box application of these machine learning tools
developed in industry, because of course lattice gauge fields have lots of features that you don't
have in, say, images, in particular symmetries. So the physics is invariant under specific field
transformations like rotations, translations with our four-dimensional boundary conditions.
We have gauge transformations, so where you're changing each link differently based on your
gauge transformation. And so if we compare the properties of an ensemble of lattice QTD gauge
fields to, say, a benchmark image set for machine learning, you see a number of really significant
differences. The first is that a state-of-the-art lattice gauge field is described by something like
10 to the 9 numbers, and we want thousands of configurations, compared to, say, an image
described by 3,000 numbers with 60,000 samples available for training. So we have this extreme
inverted data hierarchy, which means a lot of machine learning approaches based on training,
using training data, won't work for us. It's the ensemble of gauge fields that has meaning,
long-distance correlations are important, as opposed to short-distance correlations having
meaning like a nose or an eye in an image, and we have these complicated symmetries.
So addressing this problem is part of a great collaboration from my group, as well as
collaborators at Google DeepMind. Michael L. Berger, who's now a grad student at NYU, started
with this as an undergrad. Julian Urban at the University of Heidelberg. So great collaboration,
lots of fantastic young people, where we've been pursuing this program of how can we use
generative machine learning models for gauge field generation. And as I sort of indicated,
one of the big challenges is incorporating symmetries. So our field configurations are
represented not as in the scalar field example I gave, with a number at each site of a ladder,
but by links here on our lattice encoded as matrices. So for QCD, we have SU3 matrices,
that's 3 by 3 complex matrices with unit determinant. And so the very first thing is
that group-valued fields don't live on the real line, but they live on compact manifolds. So we
first had to go back and design these sorts of flow architectures on compact connected manifolds,
even just the angles on a circle. And then, even once you have flows on compact manifolds,
well, the action is invariant under group transformations on gauge fields,
so gauge transformations. So we had to go back and incorporate symmetries
for gauge equivariant flow architectures. And I can't go into too much detail here,
just because of the length of this talk, but you can see a recorded talk by Tej Kanwar
on the virtual lattice colloquium series, which is fantastic. And I can show you an artist's
impression of the sort of architectures that we're developing, which are really
gauge-specific to incorporate these symmetries. And I can show you some of the results.
So in an application to U1, pure gauge, pure theory, we see that training these sorts of
flow models really gives you a significant reduction of your problem of critical slowing
down, of course, at a cost, at the cost of upfront training of the model. So here I'm showing
sampling of the topological charge, which is the winding number of the gauge field. So the
ITS Singer index theorem tells us it's quantized, it's a global quantity. And what we see here,
this is on the horizontal axis, the Markov chain step, and on the vertical axis, the topological
charge. You can see HMC, hybrid Monte Carlo. So one of the conventional approaches gets very stuck
in topological sectors. Similarly, heat bath, here another conventional algorithm shown by the gray
line, better, but still very stuck. But the flow-based approach samples very efficiently
across this space. So we can turn this into a measure of cost. So here's the integrated
autocorrelation time, which you can think of as being a measure of the cost per independent sample
in the sampling phase of this algorithm. And here on the horizontal axis, we have beta. So you can
think of this as moving along a critical line. You can see that here, HMC and heat bath, this is a
log axis, so they become exponentially more expensive as you get to larger beta. And the
flow algorithm here is reasonably constant. It does increase at larger beta, and that's due to
it getting harder to train. But you still have several orders of magnitude more efficient
sampling. And I should say that this is a fair comparison in sampling. The cost of one sample
from HMC and the flow model is about the same. Heat bath is a bit cheaper, maybe a factor of two,
but you still have genuine orders of magnitude more efficient sampling. So this is really a
success. This is a proof of principle of an efficient, exact machine learning algorithm
for sampling in lattice field theory. We've extended this to SUN. You can see in this paper
here, fermions coming soon. And the big question, of course, where significant work is required is,
can this be scaled to the size of state-of-the-art lattice calculations? So we have an early science
project for Aurora to try to do this and make this a genuine tool for lattice field theory
calculations. And it's promising, but there's a lot of work to do. And of course, I should say
that although we still face a great scaling challenge, these sorts of techniques really
have found interdisciplinary applications. So for example, developing these flow architectures on
compact connected manifolds, or something else that lives on a compact connected manifold is the
joints of a robot arm described by angles. And so that work has interdisciplinary applications.
These same sorts of ideas have been applied in molecular genetics and drug design problems
where the scaling challenges aren't so severe. And so there's been significant advances made
for using these sorts of algorithms. So to summarize, I've given really a sort of broad
discussion of how we can understand the structure of hydrons and nuclei from the standard model with
several very specific examples pulled from some of my recent work. But the upshot is really that
first principles QCD calculations are revealing, and precisely, the structure of protons, neutrons,
and even nuclei, both giving us new understanding of the structure of matter in first principles,
but also giving us new understanding that we can use to find the physics beyond the standard model,
so to constrain beyond standard model physics searches. And then I talked also about how we
can try to use new computational tools to push the barriers of what's possible with the limited
computation we have available, particularly through machine learning. Thank you very much.
Thank you very much Fiala for this very nice talk.
Thanks Carsten.
It's a strange setup. Okay, I'm sure there are questions to Fiala, so please raise your hand
and then I should be able to allow you to talk. Okay Manuel, you should unmute your microphone.
I have a question regarding this interaction of dark matter with nuclei. This effect that you are
talking about, this is distinct from form factors? I mean, form factors have been included for a long
time, of course, to go beyond the very simplest point like nucleus approximation. But I guess
your effect is a bit more complicated than that. Yeah, so this is just literally nuclear effects
in the scalar matrix element. So not the scalar form factor, just as the scalar
matrix element itself, which would affect the whole form factor as well.
Okay, that's a question by Daniel Jenkins. Yeah, again, when you were discussing the dark matter,
especially for the helium case, there were multiple different quark combinations and it
looked like there was much more deviation in the helium case when we expect disconnected
contributions, like for when we talk about individual nucleons we have connected and
disconnected. Also, there was much more deviation for when there is disconnected contributions. Does
that connected-disconnected thing carry over into the helium case? Let me go back to that slide.
Oops, I have a spinny wheel. Okay, my slides aren't changing. So your question
was for the helium-3, we saw that there were significant nuclear effects in the disconnected
case. That's the case. And of course, which of those different columns of different quark
combinations is relevant depends on the exact microscopic model of dark matter that you're
interested in. So the simplest thing, of course, is that you would couple to all of the
different quark flavors, so strange quarks and disconnected contributions would be important.
But yeah, so this was the flavor separation. You see that the strange quark contributions
were much smaller themselves, and so the reason this looks so large is that this is showing the
percentage nuclear effect. So you can see the disconnected part itself is only a couple of
standard deviations from zero, but it looks, it's much smaller and much noisier. So there's,
so basically that's not all that's driving these differences here, because it's a small
contribution. This is many more standard deviations from zero, which is not just driven by the
disconnecteds. All right, thank you. No problem. Okay, Chris. Yeah, thanks y'all for the nice talk.
With regards to the machine learning stuff, I mean obviously this helps you generate a lot
of configurations very quickly. I was wondering if you felt like that this would be enough
to kind of solve the problem of the exponential growth of the noise of like many nuclear systems,
or if you think anything else would be done to solve that kind of problem?
Yeah, so it's, I mean, it's the first, it's one aspect of that problem. If you can generate a lot
of configurations quickly, you still have to do calculations on all of those configurations,
and you still have to have the factorial growth in contractions on all of those configurations.
So it's by no means a solution to that problem, but it's, you know, a chunk of the solution to
that problem. And the other things we find with these machine learning models is that once you've
trained them for one set of parameters, you can quite efficiently move to other places in parameter
space. So you want to think about the sort of setup cost of giving you not just one ensemble,
but all sorts of ensembles in some region of parameter space. And so you can think about,
you know, efficiently trying to do things like interpolate or extrapolate in quark masses by
this approach. But no, it certainly doesn't completely solve the problem, because you still
need to, on all of those configurations, do all of the work, which is much more expensive for nuclei
than it is for neutrons. Are there further questions? Currently I don't see any. Then
let's thank Tiala again for this very nice talk, and let me take the opportunity to also
congratulate you again for this award. Things were deserved. Thank you very much. I thank all
of you for attending this virtual colloquium. I hope you all stay safe and healthy in these times,
and I hope very much that we see each other next year, not next year, but at least in two years
from now at an in-person lattice conference, hopefully in Bonn.
So once more, thanks very much, and bye-bye. Take care. Thank you.
