Upload
others
View
2
Download
0
Embed Size (px)
Citation preview
A proposal for solving the high-Tc superconductivity conundrum via artificial
intelligence
A research proposal by Yoon Tiem Leong
School of Physics, Universiti Sains Malaysia, Penang
March 2019
Key words: Artificial intelligence; Machine learning; Many-body physics problem; High-Tc
superconductivity; Extended Hubbard Model; Phase classification
Executive Summary of Research Proposal
The EHM [6], a USM-grown theoretical model, is a potential resolution to the ‘Holy Grail’
problem of HTSC. Latest AI-inspired algorithms provide a promising tool for overcoming the
major technical bottleneck of applying EHM to solve the HTSC conundrum which involves
the classification of the phases inherently embedded in the model Hamiltonian. AI-inspired
algorithms will be used to solve for the ground states, classify the quantum phases and the
dependence of the transition temperatures in the EHM. This is a computational condensed
matter physics project requiring heavy HPC facilities. The technical details and procedures of
calculations in this project are to be based heavily on, but not exclusively, that published in
[24], [32], [33], [36], [37]. The generic machine learning platform TensorFlow [46], will be
used for performing ML calculations, in conjunction with a list of open-source QMC codes
[38, 39, 40, 43] for generating configurations for the 2D Hubbard and EHM models to
sample data for classifying the phases. Other than deploying existing software packages, we
shall also need to develop our own codes for handling part of the calculations. The project is
strategically segregated into stages with progressive level of complexity. Classical results
and quantum results of selected prototypical many-body systems using AI-inspired
algorithms will first be reproduced, followed by the reproduction of the known results for
the prototype 2D Hubbard model. Finally, the EHM modelled will be dealt with based on the
ML techniques and computational experience picked up in the preceding stages by
numerically investigating the superconducting and the pseudo-gap sectors, and the
interplay between them. Apart from contributing to the resolution of the HTSC conundrum
[47], the know-how gained in this highly theoretical project can be translated into other
productivity applications which the industries are desperately needing.
Problem Statement
High-temperature (high-Tc) superconductivity (HTSC) in cuprates [1] is one of the most
profound physics problems since 1986. The strong electronic correlations, which is
intrinsically a quantum many-body effect, responsible for the superconducting, pseudo-gap
[45] and other measured phases in unconventional high-Tc superconducting cuprates
remain an unresolved problem in condensed matter physics to date. This is famously known
as the ‘HTSC conundrum’ [47]. Arriving at a working theoretical model for resolving the
conundrum in HTSC constitutes one of the ‘Holy Grails’ in condensed matter physics. Apart
from arriving at the right theoretical model, the other major challenge concerns with the
difficult task of solving these many-body physics models which are, generically,
computationally expensive if not practically formidable. Specifically, classification of
quantum phases in a generic quantum many-body model Hamiltonian with existing,
conventional computational approaches is known to be particularly so. We propose a
resolution to the ‘Holy Grail’ problem of HTSC based on a model, so called the Extended
Hubbard Model (EMH), published by B. S. Lee [6], B. S. Lee, T. L. Yoon and R. Abd-Shukor [7]
in 2010 and 2017 respectively (T. L. Yoon is the leader of this proposal). One of the technical
bottlenecks of this research proposal is to classify the phases inherently embedded the EMH
without having any prior knowledge of the solutions embodied in the exponentially large
state space of the many-body system. Latest AI-inspired algorithms which are well suited
with the ability to classify, identify, or interpret exponentially large data sets provide a
promising tool for overcoming this major bottleneck.
Hypothesis
The many-body physics responsible for the unconventional high-Tc superconducting and
pseudo-gap phases in cuprates are correctly captured in the renown Hubbard model and
Hubbard-type models. In principle, given the underlying Hamiltonian of a many-body
system, it is possible to derive all we need to know about the physical and macroscopic
details of the system via a spectrum of conventional computational techniques (such as
quantum Monte Carlo [2,3,9], density matrix renormalization group [4], and tensor
networks [5]). We distinguish two distinct categories of computational technique to solve
many-body physics, namely, the conventional (i.e., non-artificial intelligence) and artificial
intelligence (AI) approaches. Those AI-inspired computational approaches, which rise rapidly
only in the very recent years, could deliver a much feasible, cheap and relatively promising
way for solving quantum many-body problems that are otherwise proven extremely
expensive and often intractable via non-AI approaches. Specifically, AI-inspired algorithms
are expected to provide a powerful, computationally cheap, convenient, and accurate tool
for quantum phase identification in quantum many-body problems over existing,
conventional approaches.
Research Questions
1. How to technically classify, via the novel AI approaches, the quantum phases as well
as the parameters driving these phases as embedded in the EHM?
2. What are the undiscovered physics embedded in the EHM relevant to the pseudo-
gap and superconducting phases in the HTCS cuprates?
3. Could there be any other unexpected or serendipitous new physics embedded in the
EHM?
4. Could we finally nail down the ‘Holy Grail’ of unconventional HTCS in the EHM with
the new arsenal of AI-inspired computational approaches?
Literature Reviews
One of the main tasks in theoretical condensed matter physics is to infer the macroscopic
properties of a physical system from their microscopic description, i.e., its underlying
Hamiltonian. The most drastic changes in the macroscopic properties of a physical system
occur at phase transition. Different phases can be identified by an order parameter that is
zero in the disordered phase and nonzero in the ordered phase. Whereas in many known
models the order parameter can be determined by symmetry considerations of the
underlying Hamiltonian, there are states of matter where such a parameter can only be
defined in a complicated nonlocal way [8], such as the models for unconventional high Tc-
superconductivity [1] in cuprates, which is the intended system we wish to investigate in
this proposal. Theoretical condensed matter physicists wish to predict the details of phase
transition, such as the critical temperature for the onset of superconducting [1] or pseudo-
gap [45] phases, and the physics driving the transitions among the various phases
embedded in a many-body quantum model. However, the theories for describing phase
transition [10, 11, 12] generally involve non-local, long-range order effects which are often
computationally formidable.
Serious models for unconventional high-Tc superconductivity generically involve
quantum many-body effects. Quantum many-body (QMB) problems are known to be
computationally daunting due to the nontrivial correlations encoded in the exponential
complexity of the many-body wave function Ψ. An exponential amount of information is
needed to fully encode a generic many-body quantum state, rendering reliable numerical
solutions for the ground state technically difficult to come by. Conventionally, many-body
calculations are performed through highly sophisticated computational methods with some
extent of approximations, such as quantum Monte Carlo (QMC) methods [2,3,9], density
matrix renormalization group [4], tensor networks [5], matrix product states (MPS) [13, 14,
15] or general tensor networks [16, 17, 18]. However, there are many instances where these
conventional approaches fail, due to, e.g., the sign problem [19] or the inefficiency in
handling the exponentially huge degree of freedom inherent in these systems.
Fundamentally, the quandary in QMB lies in the failure of finding a general strategy to
reduce the exponential complexity of the full many-body wave function down to its most
essential features (this is known as the ‘curse of dimensionality’ [32]).
As an independent development in condensed matter and quantum statistical
physics, since 2017 or so we are witnessing a rapid rise of AI, a. k. a. machine learning (ML)
technique in the present context, as a powerful tool for solving many-body problems,
classical as well as quantum ones, that are very expensive to calculate via non-machine
learning approaches. Machine learning provides a complementary paradigm to the
conventional computational approach. The ability of modern machine learning techniques
to classify, identify, or interpret massive data sets such as images provides physicists with a
new and suitable tool to deal with the exponentially large data sets embodied in the state
space of condensed-matter systems. The fundamental benefit of applying machine learning
in physical problems is that this approach can extend beyond the limits of conventional
approaches by obtaining solutions based on partial or even no prior physical knowledge,
and thereby extrapolate them to unexplored data. This benefit has been demonstrated in
many papers as recent as 2016 [23], including identifying phase transition by unsupervised
learning approaches in classical systems [20, 21, 22, 23]. In [23], Wang applied principle
component analysis (PCA) to classify the two phases in the classic Ising model. Later, van
Nieuwenburg et al. [22] proposed the so-called confusion scheme to obtain critical points
successfully for several Ising-like models. The application of ML, as demonstrated in these
publications, represents a landmark in the study of phase transitions.
In the year 2017, another pioneering paper appearing in Science by Carleo and
Troyer [24] has created a spark of frenzy among computational condensed matter physicists
as it successfully shown a new direction by using artificial neural network (ANN), a subset of
artificial intelligence algorithm, to achieve correct phase classification and transitions in
certain prototypical quantum many-body systems, without knowing a-priori the boundary of
the phases. ANN is in general inspired by the human brain neural network. It uses the model
of neurons which are connected with each other by synapses. The neurons are divided into
the input, hidden and output layers connected with weighted synapses as seen in Fig. 1.
The key ingredient of the success of Carleo and Troyer’s [24] approach is that the ANN can
effectively and efficiently compress the essential information of a many-body wave function
in high-dimensional systems in terms of ‘weights’, a set of internal parameters representing
the neutral network. Neural network representation of the many-body ground state Ψ
tremendously reduces the dimensionality to represent it to a level much lower than the
maximum Hilbert space. The parameters of the neural network are then optimized either by
static variational Monte Carlo (VMC) sampling [25] or time-dependent VMC [26, 27]. The
artificial neural network architecture used by Carleo and Troyer [24] is known as restricted
Boltzmann machine (RBM). Such an AI-inspired strategy to represent the ground state of a
quantum many-body system could potentially reduce the computationally cost by many
Figure 1. Neurons layers and synapses
orders of magnitude, apart from the ability to access parametric regions that are otherwise
not possible using conventional numerical approaches such as the popular QMC.
Since the pioneering work by Carleo and Troyer, many works using neural network to
address QMB problems have been published. To name a few, the Bose-Hubbard model has
been studied by using a similar method in [28, 29], while the feedforward neural network
has been applied to the Fermi-Hubbard model [30]. Another method used to study QMB
problem, deep neural network DNN, is reported in [31] where a large class of many-body
lattice Hamiltonians is studied using deep Boltzmann machine. Similar approach employing
convolutional neural network (CNN) has also been recently reported by [23, 32], where CNN
is shown to efficiently encode phases of matter and discriminate phase transitions in
classical correlated many-body systems. In a follow-up of an earlier work [32], Carrasquilla’s
group has reported in [33, 42] an auxiliary field QMC (AFQMC) technique to sample
statistical instances of the wavefunction of a fermionic system. A CNN was then deployed to
discriminate between two fermionic phases of a fermionic Hubbard-type quantum lattice
models. The authors showed that their machine learning approach, coupled with auxiliary
field QMC, can locate the phase transition point in selected systems which are known to be
plagued by severe sign problem. Based on the literature findings since 2017 pertaining to
applying machine learning in quantum many-body problems, it seems that the formidable
fermion sign problem that haunts generations of condensed matter theorists seems to have
a new weapon, i.e., machine-learning, to tame the beast. At this point of writing, the most
up-to-date and technical summary of how machine learning rises to overcome the curse of
dimensionality in quantum many-body can be found in a March 2019 article in Physics
Today [48]. The essential ideas of solving quantum many-body problem with various ML
algorithms, which would be used in this proposal, are clearly explained in [48].
However, our literature search reviews that only prototypical many-body models,
e.g., Ising-like and Heisenberg spin-like models are explored via ML approaches. There has
not been much comprehensive work of using AI-inspired approaches to study the full-
fledged 2D Hubbard model [34, 35], which is arguably the most studied QMB model since its
first proposal in 1963. Specifically, there has not been any comprehensive works of using AI-
inspired approaches to explore the Hubbard model to explain the interplay of
superconductivity, pseudo-gap and other observed phases in HTSC, nor how the transition
temperatures associated with these phases are predicted as a function of the free
parameters of the model. It is conventionally believed that 2D Hubbard model provides a
working model that captures some if not all essential features in copper-oxide
superconductors. It remains a daunting model to be completely solved despite relentless
computational efforts spent for so many years to abstract the physics out of it due to the
inherent sign problem. To illustrate the significance of the Hubbard model, as recent as
2017, we still see heroic effort such as that by two back-to-back papers in Science (Zheng et
al. and Huang et al., [36]) reporting their consorted, yet independent effort for obtaining the
ground state of the prototype 2D Hubbard model using state-of-the-art computational
methods , including Auxiliary field quantum Monte Carlo (AFQMC), density matrix
renormalization group (DMRG), hybrid DMRG, density matrix embedding (DMET) and
infinite projected entangled pair states (iPEPS). These are all ‘conventional’, non-machine
learning methods. On the other hand, the physics of superconductivity, pseudo-gap, and
their dynamical interplay in the 2D Hubbard model have been explored numerically using
another non-machine-learning approach known as cluster dynamical cluster approximation
(DCA) in a paper by Gull in 2013 [37]. The paper uses the DCA technique to map out the
phases (including the pseudo-gap, superconducting, Fermi liquid and Mott phases) along
with other physical insights, e.g., the transition temperatures as a function of the dopant
concentration, in the 2D Hubbard model. It is well known that these non-AI inspired
approaches are highly expensive and expert knowledge-demanding.
Independently, there are also alternative views on the physical origin of the pseudo-
gap and superconductivity in high-Tc cuprates, such as that proposed by B. S Lee [6] and the
following-up paper by B. S. Lee, T. L. Yoon and R. Abd-Shukor [7]. B. S. Lee was a well-
recognized theoretical physics professor retired from USM, while Yoon is the leader of this
research proposal. [6] proposes a physically-motivated model, known as the Extended
Hubbard Model (EHM) for explaining the physical origin of the pseudo-gap phase observed
in HTSC cuprates, whereas [7] further explores the physical signatures and consequences in
the superconducting sector in the EHM via a hybrid approach of theoretical mean field
theory and conventional numerical procedures. EHM offers a physically motivated
mechanism for modelling the novel phases in HTSC, which are traced to a common origin,
namely a Jahn-Teller type interaction induced by an electron interacting with a nonlinear 𝑄2
mode of the oxygen clusters in the CuO2 plane of the HTSC cuprates. The EHM is an
extended version of the prototype 2D Hubbard model and admits additional terms for the
lattice-electron interactions induced by the 𝑄2 mode. Due to its theoretically attractive
features, the EHM is a fully qualified contending many-body model for solving the
unconventional high-Tc superconductivity conundrum, the ‘Holy Grail’ of condensed matter
physics. This proposal makes this highly original, Malaysian-grown model the main subject
of research. It serves a rare chance where our very own home-grown theoretical model
could strike a truly profound achievement in the international frontier of condensed matter
physics.
The use of AI has a long history and its applications are already causing continuous
yet profound impact to the world we are living in. However, based on the literature review
as illustrated above, the application of machine learning in the area of QMB is a very recent
one. This means that there is still a lot of unexplored avenues to pursue. We, a group of
aspiring physicists and computer scientists, are jumping at the opportunity of such
promising research frontier to apply AI-inspired approaches to solve a ‘Holy Grail’ problem
in theoretical condensed matter physics.
Relevance to Government Policy, if any
Our proposed project is advancing one of the prioritised research areas as stated in the
Malaysian Ministry of Education research grant policy, namely, Artificial Intelligence and
Internet of Everything, which is an integral part in any research policy aimed for long term
sustainability. The scope and scale of our proposal, which is computational-cum-theoretical
in nature, offers a natural advantage against the pragmatic constraints in the context of
Malaysian research environment.
Objective(s) of the Research
1. To numerically solve for the ground states, classify and map out the quantum
phases in the prototype 2D Hubbard model via the AI-inspired machine learning
algorithms as originally proposed by the pioneering papers [24], [32], [33].
2. To work out the dependence of the superconducting transition temperature in
the EHM
3. To elucidate the physics and interplay of pseudo-gap, Mott, superconducting and
other observed phases in HTSC cuprates within the framework of EHM.
4. To establish a successful showcase of applying AI techniques in solving highly
non-trivial computational problems in the pure academic frontier of theoretical
condensed matter physics.
Description of Methodology
a. This is a computational condensed matter physics project. All calculations are to
be implemented in high-performance computing (HPC) facilities where open-
source AI platforms and Quantum Monte Carlo packages are to be deployed. To
this end, availability of HPC hardware, specifically, large amount of RAM, hard
disks, GPUs and powerful parallel machines are highly desirable.
b. The technical details and procedures of calculations in this project are to be
based heavily on, but not exclusively, that published in [24], [32], [33], [36], [37].
c. Python programming will be learned and used as it is the ‘default’ programming
language in most AI platforms. We shall also develop our own programming
codes to batch the implementation of ML calculations and perform some
numerical tasks. Mathematica, Fortran and shell script are the candidate
programming language of choice, depending on circumstances and necessity.
d. The generic machine learning platform developed by Google, which is popularly
used by the condensed physics research community, TensorFlow [46], will be
used for performing our machine learning calculations, in conjunction with a list
of open-source codes, as listed below:
(i) NetKet [38]. It is listed as one of open-source computing software
developed in the Center for Computational Quantum Physics (CCQ)
(https://www.simonsfoundation.org/flatiron/center-for-computational-
quantum-physics/) under the Simon Foundations
(https://www.simonsfoundation.org/about). NetKet comes with a
complete set of well-documented manual plus tutorials. It is the machine
learning package used for the publication of the pioneering paper [24].
(ii) QUEST (QUantum Electron Simulation Toolbox) [39]. The theoretical
background and other technical details of QUEST are described in the
documentation within the QUEST portal [39]. QUEST is a determinant
QMC (DQMC) implementation that is particularly suitable for generating
data from a generic 2D Hubbard model.
(iii) ALF package [40] which implements the auxiliary field QMC (AFQMC).
(iv) An AFQMC library, which is still under development at the time of this
proposal, by CCQ (see
https://www.simonsfoundation.org/flatiron/center-for-computational-
quantum-physics/software/auxiliary field-quantum-monte-carlo-2/).
(v) CPMC-Lab [43]. This is a Matlab package with a graphical interface made
readily available by the project leader of the AFQMC in CCQ, Shiwei
Zhang. It is a constrained-path and phaseless auxiliary field Monte Carlo
code for the Hubbard model.
We shall make use of the above QMC codes for generating 2D Hubbard model
configurations as sample data for training, testing and classifying for phases in
selected Hamiltonians. Our TensorFlow platform will be installed in a GPU-
enabled HPC. All software packages required for this research project are open-
source and require no cost for the licenses. We shall also maintain these
packages ourselves as our research group has the necessary expertise to do so.
e. Stage 1: To begin with, the numerical results of the classical Ising model as
reported in [24] shall be reproduced using NetKet and TensorFlow as a warming-
up training to familiarize with the computational handling of the package. The
numerical results to recognize the phases in classical spin systems using fully
connected and convolutional neural networks in the prototype Ising model as
reported in the 2017 Nature Physics paper [32] shall also be reproduced as part
of the Stage 1 exercise.
f. Stage 2: Once the classical results of phase classification are derived and verified,
we shall proceed to reproduce the quantum results as reported in [33] (which is
a follow-up of [32]), where the fermionic sign problem was reported to have
been circumvented. To this end, we shall make use of the NetKet and TensorFlow
packages which were used by the authors of [32] and [33] to obtain these
published results.
g. The proposed tasks in Stage 1 and Stage 2 are meant as warming-up exercises for
this research project. It is expected to last for a few months to a year, during
which the know-how to deploy machine learning algorithms in NetKet, use
Python programming and run computations on a GPU-enabled TensorFlow
platform to solve the QMB problems would have been picked up. Specifically, the
know-how to deploy Restricted Boltzmann machine (RBM) and Deep Boltzmann
machine (DBM) on prototype spin models, including transverse-field Ising model,
antiferromagnetic Heisenberg model, Bose-Hubbard model (these models are
well documented as part of the tutorials exercise in the NetKet code) would
become familiarized.
h. Stage 3: The next stage in the planned progress line is to apply the relevant
machine learning algorithms as reported in [24], [32], [33] on the 2D Hubbard
model based on the experience gained from the warming-up stage. The results
for the phases classified via various ML algorithm should be compared and
verified against that obtained via the non-AI approaches as reported in [36] and
[37].
i. Stage 4: The next stage would be to confront the EHM directly. EHM is a
generalisation of the prototype 2D Hubbard model which have never been
subjected to any ML investigation before. Armed with the knowledge and
experience picked up until this stage, we shall proceed to classify the phases
inherently embedded in the model Hamiltonian using ML algorithms.
j. Stage 5: We will then attempt to numerically investigate the two sectors
embedded in the EHM, i.e., the superconducting and the pseudo-gap sectors,
and the interplay between them. The numerical investigation at this stage does
not necessarily involve AI algorithms but relies on conventional computational
methodology. The dependence of the critical temperature with doping
concentration and other free parameters in the EHM model can be calculated by
following the theoretical method as detailed in [41]. The computation approach
in [41] requires the generation of QMC data from the model Hamiltonian to
determine two independent quantities, i.e., the helicity modulus and the pairing
correlation function. Critical temperature can be deduced once these two
quantities are known. To carry out the calculations as prescribed in [41], we
would have to develop our own numerical code to implement the computation
procedure.
(e) Expected Results/Benefit
Novel theories/New findings/Knowledge
1. The knowledge of applying various machine learning tools for the purpose of
solving quantum-many body problem, which can be easily translated into other
general, real-life applications, such as productivity optimization, pattern
recognition, identification of signals buried under noisy background, etc.
2. A practical know-how for using AI-spired techniques for predicting phases of
generic quantum or classical many body systems.
3. A resolution for the unconventional high-Tc superconductivity conundrum.
Specific or Potential Applications
The knowledge of applying various machine learning tools for the purpose of solving
quantum-many body problem, which can be easily translated into other general, real-life
applications, such as productivity optimization, pattern recognition, identification of signals
buried under noisy background, etc.
Impact on Society, Economy and Nation
This research proposal basically has two aspects which are in principle distinct from each
other, namely the AI aspect and the physics of HTSC. The former is a very, very powerful and
versatile tool while the latter concerns academic researchers committed to understand the
fundamentals of how nature work.
As for the latter aspect, in the context of a developing country like Malaysia who is
stuck at the bottle neck of a middle-income range, the disinterest to advocate pure
fundamental research is hampering the nation in the long run. We need to establish a
society worthy of protecting, one which embraces intellectual appreciation for endeavour
that do not appear to provide immediate, tangible gain. Supporting pure, fundamental
research (which this research proposal is one), has a long lasting, though gradual, impact on
the overall intellectual level of the nation.
On the other hand, the benefit and potential significance of the AI component in this
research proposal is immediately apparent. The highly non-trivial experience gained from
practically meddling with the AI tools can immediately be translated into practical
applications which the nation and industries are desperately needing. Being a very versatile
tool, the experience gained from applying AI in solving theoretical many-body problems,
which are highly mathematical and non-trivial in nature, can be easily translated into
applications for optimizing productivity in almost all aspect of economic and research
activities. Our nation is yearning for capable human capital in this AI age. This research
proposal, which is AI intensively by nature, rightly fits the bill for a nation which is already
far lagging behind in the face of the AI era.
Intellectual Property (IP)
This is a theoretical cum computational research project on fundamental physics that
produces universal knowledge that cannot be filed for IP.
References
1. Bednorz, J.G., Muller, K.A.: Rev. Mod. Phys. 60, 585–600 (1988)
2. D. Ceperley, B. Alder, Science 231, 555–560 (1986).
3. W. M. C. Foulkes, L. Mitas, R. J. Needs, G. Rajagopal, Rev. Mod. Phys. 73, 33–83
(2001).
4. U. Schollwöck, Rev. Mod. Phys. 77, 259 (2005).
5. S.-J. Ran, E. Tirrito, C. Peng, X. Chen, G. Su, and M. Lewenstein, arXiv:1708.09213.
6. B.S. Lee, J Supercond Nov Magn (2010) 23: 333–338.
7. B. S. Lee, T. L. Yoon and R. Abd-Shukor, J Supercond Nov Magn (2017) DOI
10.1007/s10948-017-4087-4.
8. X.-G. Wen, Quantum Field Theory of Many-Body Systems (Oxford University Press,
Oxford, 2004).
9. J. Carlson et al., Rev. Mod. Phys. 87, 1067–1118 (2015).
10. L. Landau, Phys. Z. Sowjetunion 11, 26 (1937).
11. L. P. Kadanoff, Physics 2, 263 (1966).
12. K. G. Wilson, Rev. Mod. Phys. 47, 773 (1975).
13. S. R. White, Phys. Rev. Lett. 69, 2863–2866 (1992).
14. S. Rommer, S. Ostlund, Phys. Rev. B 55, 2164–2181 (1997).
15. U. Schollwöck, Ann. Phys. 326, 96–192 (2011).
16. R. Orús, Ann. Phys. 349, 117–158 (2014).
17. F. Verstraete, V. Murg, J. I. Cirac, Adv. Phys. 57, 143–224 (2008).
18. K. H. Marti, B. Bauer, M. Reiher, M. Troyer, F. Verstraete, New J. Phys. 12, 103008
(2010).
19. M. Troyer, U.-J. Wiese, Phys. Rev. Lett. 94, 170201 (2005).
20. N. Walker, K.-M. Tam, B. Novak, and M. Jarrell, Phys. Rev. E 98, 053305 (2018).
21. G. Torlai and R. G. Melko, Phys. Rev. B 94, 165134 (2016).
22. E. P. Van Nieuwenburg, Y.-H. Liu, and S. D. Huber, Nature Physics 13, 435 (2017).
23. Discovering phase transitions with unsupervised learning, L. Wang, Phys. Rev. B 94,
195105 (2016).
24. Solving the quantum many-body problem with artificial neural networks, G. Carleo
and M. Troyer, Science 355, 602 (2017).
25. W. L. McMillan, Phys. Rev. 138, A442–A451 (1965).
26. G. Carleo, F. Becca, M. Schiró, M. Fabrizio, Sci. Rep. 2, 243 (2012).
27. G. Carleo, F. Becca, L. Sanchez-Palencia, S. Sorella, M. Fabrizio, Phys. Rev. A 89,
031602 (2014).
28. H. Saito, J. Phys. Soc. Jpn. 86, 093001 (2017).
29. H. Saito, J. Phys. Soc. Jpn. 87, 074002 (2018).
30. Y. Nomura, A. S. Darmawan, Y. Yamaji, and M. Imada, Phys. Rev. B 96, 205152
(2017).
31. G. Carleo, Y. Nomura, and M. Imada, arXiv:1802.09558.
32. Machine learning phases of matter, Juan Carrasquilla and Roger G. Melko, Nature
Physics 13, pages 431–434 (2017)
33. Machine learning quantum phases of matter beyond the fermion sign problem, Peter Broecker, Juan Carrasquilla, Roger G. Melko and Simon Trebst, Scientific Reports 7: 8823 (2017)
34. J. Hubbard, Proc. Roy. Soc. London, A, 276 (1963), pp. 238-257.
35. J. Hubbard, Proc. Roy. Soc. London, A, 281 (1964), pp. 401-419.
36. Numerical evidence of fluctuating stripes in the normal state of high-Tc cuprate
superconductors, Huang et al., Science 358, 1161–1164 (2017); Stripe order in the
underdoped region of the two-dimensional Hubbard model, Zheng et al., Science
358, 1161–1164 (2017).
37. Superconductivity and the Pseudogap in the Two-Dimensional Hubbard Model, Gull,
Parcollet and Millis, PRL 110, 216405 (2013).
38. NetKet webpage, https://www.netket.org/
39. QUEST webpage, http://quest.ucdavis.edu/
40. The ALF Documentation for the auxiliary field quantum Monte Carlo code, Martin
Bercx et al., SciPost Phys. 3, 013 (2017). URL:
https://scipost.org/SciPostPhys.3.2.013/pdf
41. Critical temperature for the two-dimensional attractive Hubbard model, Thereza
Paiva, Raimundo R. dos Santos, R. T. Scalettar, and P. J. H. Denteneer, RPB 69,
184501 (2004).
42. Machine Learning Phases of Strongly Correlated Fermions, Ch’ng et al., Physical
Review X 7, 031038 (2017).
43. Homepage of the CPMC-Lab Package, https://cpmc-lab.wm.edu/; Journal-ref:
Computer Physics Communications 185, 3344 (2014).
44. Discovering phases, phase transitions, and crossovers through unsupervised machine
learning: A critical examination, Wenjian Hu, Rajiv R. P. Singh and Richard T.
Scalettar, PHYSICAL REVIEW E 95, 062122 (2017).
45. Lawler, M.J., Fujita, K., Lee, J., Schmidt, A.R., Kohsaka, Y., Kim, C.K., Eisaki, H., Uchida,
S., Davis, J.C., Sethna, J.P., Kim, E.A.: Nature 466(7304), 347 (2010).
doi:10.1038/nature09169
46. TensorFlow webpage, https://www.tensorflow.org/
47. The great quantum conundrum, P. M. Grant, Nature 476, 37 (2011).
48. Machine learning meets quantum physics, S. D. Sarma, D.-L. Deng and L.-H. Duan,
Physics Today 72, 3, 48 (2019). https://doi.org/10.1063/PT.3.4164.
Flow Chart of Research Activities
Verify that these classical results are successfully reproduced and comparable to that published in the relevant papers listed in the references.
Plan, prepare, design and test-run ML calcultions of selected prototype classical systems, e.g., ferromagnetic Ising model in [32,34] Ising. The purpose is to learn how to to use these reported ML algorithm s to reproduce the phases in these classical systems.
Purchase, upgrade and install high-performing computing (HPC) facilities purchased with grant money. Recruit a graduate student.
Results justified?
Troubleshoot
Generate the phase classification and transition of selected classical systems that have never published before using the novel ML technique (e.g., unsupervised machine learning techniques of [44]), such as that of water or other selected Lennard-Jones systems.
Publication of the ML derived results on the phase classification of these selected classical systems.
Stage 1 completes
No
Yes
Troubleshoot
Plan, prepare, design and test-run ML calcultions of selected prototype quantum systems, e.g., transverse-field Ising (TFI) model and antiferromagnetic Heisenberg (AFH) model of [24], the Hubbard-like fermion model of [33] and Fermi-Hubbard Hamiltonian of [42]. ML algoriths to use: Restricted Boltzmann machine (RBM) and Deep Boltzmann machine (DBM).
Results justified?
Verify that these quantum results are successfully reproduced and comparable to that published in the relevant papers listed in the references.
Stage 2 completes
Stage 2 begins
Stage 3 begins
Apply the relevant machine learning algorithms as reported in [24], [32], [33] to perform a full-fledged classification and transition of phases in the prototype 2D Hubbard model.
No
Yes
Stage 1 begins
Troubleshoot Results justified?
Explore the prototype 2D Hubbard model to explain the interplay of superconductivity, pseudo-gap and other observed phases in HTSC, and how the transition temperatures associated with these phases are predicted as a function of the free parameters of the model.
Compare and verify the results for the prototype 2D Hubbard model phases classified via selected ML algorithms against that obtained via the non-AI approaches as reported in [36] and [37].
Stage 3 completes
No
Yes
Numerically evaluate the dependence of the critical temperature with doping
concentration and other free parameters in the EHM model by following the
theoretical method as detailed in [41]. This will involve application of AFQMC
on the EMH to generate sample data from the model Hamiltonian to
determine the helicity modulus and the pairing correlation function.
Stage 4 completes
Stage 4 begins
Begin stage 5
Numerically investigate the interplay of the two sectors embedded in the EHM, i.e., the superconducting and the pseudo-gap sectors. Propose resolution of the HTSC conundrum within the EHM.
Publish of obtained results
Stage 5 completes
Classify the phases inherently embedded in the model Hamiltonian of EHM
using ML algorithms and experience picked up during the previous stage of
investigating the prototype 2D Hubbard model.