Machine Learning Quantum Matter
Machine learning can classify configurations, infer observables, represent states or propose models of quantum matter. The research question is what physical information is learned, under which data assumptions, and whether the method generalizes beyond the examples used to train it. High classification accuracy can reflect a simulation artifact rather than a phase distinction.
Required background. Many-body simulation supplies the origin and limitations of numerical data; quantum machine learning supplies access and generalization language.
A phase label needs a physical definition
Section titled “A phase label needs a physical definition”A finite dataset contains finite systems, finite temperatures and correlated samples. A phase transition is an appropriate limiting statement. A classifier may learn energy, magnetization, boundary conditions or the Monte Carlo update history; these features can be useful but need to be identified.
For a predicted observable , report error on a held-out parameter region, not merely a random split of neighboring samples. Effective sample size can be far smaller than the number of stored configurations. Data from one Markov chain should not be spread indiscriminately across training and testing.
Evidence and bounded guarantees
Section titled “Evidence and bounded guarantees”Carrasquilla and Melko demonstrated machine-learning approaches to phases of matter. Huang and colleagues proved efficient learning results for properties of gapped Hamiltonians under specified locality, data and phase assumptions. Established: both empirical applications and conditional guarantees exist. The theorem does not imply efficient prediction across arbitrary gap closings or from arbitrary inadequate measurements. Phase-learning study, learning theorem.
Active: interpretable features, transfer between sizes and models, uncertainty under distribution shift, and data-efficient inference from experiments. Conjectural: discovery of a new phase from a classifier boundary without independent physical diagnostics.
Open questions
Section titled “Open questions”Which measured features distinguish phases when local order parameters fail? Can a model know when its training phase assumptions are violated? How should symmetry be built into the architecture? Can learned predictions improve a scientific calculation after accounting for the cost of generating training data?
Graduate project: reveal what a classifier uses
Section titled “Graduate project: reveal what a classifier uses”Train a simple model on spin configurations from two regions of a small phase diagram. Compare with magnetization, energy and a low-dimensional linear classifier. Hold out complete temperatures, system sizes and simulation runs. A laptop or workstation and a validated sampler suffice.
Deliver learning curves, feature-ablation tests, uncertainty calibration and finite-size behavior of the decision boundary. Failure tests randomize labels, change boundary conditions and remove obvious order-parameter information. If the model only rediscovers magnetization, that is a validated baseline rather than a new phase diagnostic. Research begins when a physically interpretable statistic or uncertainty method succeeds where known diagnostics fail and is checked against independent theory or data.
Entry points and scope
Section titled “Entry points and scope”Read the empirical and theorem papers together to see the gap between useful experiments and provable domains. Source checks: 11 October 2026; selected evidence: 2017–2022. A new phase claim should update only after independent physical observables and scaling support it.