Machine-Learned Interatomic Potentials with Long-Range Physics
Automated active-learning workflows for MLIP training, reciprocal-space corrections, and robust simulations beyond short-range local environments.
Research program
Machine learning that connects atoms, experiments, and materials design.
Our research program develops physically grounded machine-learning models, validates them against real measurements from beamline experiments, and applies them with collaborators to support discovery across structural, functional, and biomedical materials.
Research Thrusts
AI-Guided Biomedical Materials Discovery
Computational-experimental co-design with UB School of Dental Medicine collaborators and external partners, using molecular machine learning, uncertainty-aware screening, and targeted validation.
MLIP-Driven Characterization of Beamline Experiments
Pairing machine-learned potentials with synchrotron, neutron, and high-temperature experiments on technologically relevant disordered oxides and layered chalcogenides to build models that reproduce measured structure.
Example workflow
How experiments become atomistic models
In prior Argonne work, high-temperature X-ray and neutron measurements were used to guide active learning for machine-learned interatomic potentials. The key idea remains central to our lab and to collaborations with beamline scientists: start from the data, train selectively, and validate models for disordered oxides and layered chalcogenides against experiment before using them to predict what is difficult to measure directly.
X-ray + neutron data
→
Active learning
→
ML potential + MD
What Students Work On
Train models that know when they are uncertain
Build active-learning loops that decide which atomic configurations need expensive reference calculations.
Connect scattering data to atomic structure
Use simulations to interpret diffraction, pair-distribution functions, and temperature-dependent structural changes.
Add long-range physics to local ML models
Develop reciprocal-space corrections and attention mechanisms for materials where local environments are not enough.