AI-Driven Spectroscopic Characterization of Defects and Disorder

May 6, 2026, 10:40 AM
25m
Building 402 Lecture Hall

Building 402 Lecture Hall

Argonne National Laboratory, 9700 S Cass Ave, Lemont, IL 60439

Speaker

Mingda Li (Massachusetts Institute of Technology)

Description

Defects and structural disorder govern materials functionality, yet their quantitative, non-destructive
characterization remains a major challenge. For defect configuration, we introduce DefectNet, a foundation
model that predicts the chemical identity and concentration of multiple coexisting substitutional point defects
directly from phonon density-of-states spectra. Trained on over 16,000 simulated spectra across 2,000
semiconductors, the model identifies up to six defect species over a wide concentration range and generalizes to
unseen materials. Validation with experimental inelastic scattering data demonstrates accurate, transferable
defect quantification from vibrational spectroscopy.
For defect dynamics, we develop an integrated experiment–theory–AI framework combining coherent x-ray
photon correlation spectroscopy (XPCS) with theory-informed stochastic simulations and semi-supervised
domain adaptation. The approach quantitatively extracts grain-boundary diffusivity, stiffness, and effective
boundary concentration directly from measured two-time correlation functions, overcoming the domain gap
between simulation and experiment. This framework enables robust characterization of slow, non-equilibrium
grain-boundary relaxation in nanocrystalline silicon and provides a general route for bridging theory and
experiment in complex spectroscopic measurements.

Presentation materials

There are no materials yet.