Abstract:
Modern quantum materials and quantum devices generate data that are simultaneously high-dimensional, heterogeneous, and constrained by rich physical structure. From diffraction movies and readout of quantum hardware to materials databases accumulated over decades, extracting interpretable insight from such data poses fundamental challenges for both physics-based modeling and machine learning. In this talk, I will present a unifying modeling perspective on learning quantum matter, highlighting how domain structure and machine learning tools an be combined to offer new insights and predictions. I will introduce several case studies developed at the interface of condensed-matter physics and machine learning: (i) Quantum Attention Networks (QuAN) that use self-attention to characterize of complex quantum states; (ii) X-TEC, a clustering evolving high-dimensional diffraction data, which has led to the discovery of Bragg glass order in disordered charge-density-wave systems; and (iii) GPTc that predict superconducting transition temperatures with calibrated uncertainty from heterogeneous experimental databases. Together, these examples illustrate how modern modeling—grounded in physics but enabled by machine learning—can turn complex quantum data into predictive understanding, while revealing new opportunities for collaboration between ML and the physical sciences.
Bio:
Eun-Ah Kim is the Hans Bethe Professor of Physics at Cornell University. A pioneer at the intersection of quantum many-body physics, quantum simulation and artificial intelligence. She is the director of NSF AI institute: AI-Materials Institute (AI-MI). Her contributions have been recognized with prestigious honors, including a Radcliffe Fellowship, two Simons Fellowships for Theoretical Physics, and election as a Fellow of the American Physical Society. She received her Ph.D. from the University of Illinois at Urbana-Champaign and completed postdoctoral research at Stanford University before joining the Cornell faculty in 2008.
Summary:
Focus: AI techniques for modeling quantum matter/materials
Materials the behavior of which is government by a many-body quantum state
Behavior is collective, not single particle
Very sensitive to values of external knobs
Ex: superconductors, topological materials, quantum magnets
Physics:
Wavefunction is a complex valued vector function in an exponentially large space
Evolves over time
Need ensemble of wavefunctions to finite temperature equilibrium state
Though, not all phenomena can be described as equilibrium states
Task:
Need to design a material from given set of chemicals with set of desired properties
Many forward modeling challenges
Exponential wavefunction statespace
Relevant latent variables are unknown and emergent
Never observe wavefunction directly
Traditional forward models require strong inductive bias (results in errors where bias is incorrect)
Data is invaluable since the inverse modeling problem (given data, infer best model) is very challenging and data intensive
A representation learning problem where we need to ingest diverse datasets that collectively constrain the behavior space of the physics into a common representation that can be used to constrain models
Representation Learning
Quantum attention networks for state characterization
Scenario: state space of many binary qubits
Given a fine number of samples, summarize them in the most informative representation
Approach: classification problem
Volume law vs area law state
Topological vs trivial state
Shallow circuit vs deep circuit
Model looks at correlation in the moments; has access to higher order moments without explicitly representing them
Exploit permutation invariance in the samples by training model on random batches of observations
X-TEC for X-Ray diffraction
Goal: infer temperature/disorder from X ray images of materials
Infer time series and its phase transition across temperature thresholds
Enables experimentalists to collect data at the key phase transition temperature range
GPTc: Gaussian process Tc Predictor
Accumulated Heterogeneous Data
Gather materials structure information across the literature
Molecular structure is the key differentiator for the differences in physical behavior beyond chemical composition
Elemental features: electron affinity, electronegativity, ionization potential, covalent radius, atomic weight, column, # valence s/p/d e
Considering many candidate 2nd order features (e.g. interatomic distance)
Using ML to evaluate various features by using them as inputs to an ML model to predict observable properties
Identify which features are useful
Using Gaussian Processes to use histograms as features
Was able to predict superconductivity and transition temperature of a new material where it was previously unknown