QuICS Special Seminar: Shunshun Liu

Title:  AI-Guided Design of Compositionally Complex Alloys with Targeted Properties
Speaker:  Shunshun Liu (University of Virginia)
Date & Time:  April 27, 2026, 11:00am
Where to Attend:  ATL 3100A and Virtual Via Zoom: https://umd.zoom.us/j/9893676372?pwd=VVNOd2xNZ3FCblk4aFdTMjkzTllvQT09&omn=99440594483 Meeting ID: 989 367 6372 Passcode: abc123

The design of structural materials for extreme environments is fundamentally constrained by limited data availability and the high cost of generating reliable experimental and first‑principles datasets. This challenge is particularly pronounced in compositionally complex systems like refractory high-entropy alloys (RHEAs), where the combinatorial design space far exceeds what can be explored experimentally. Although data-driven models have accelerated property screening across vast compositional spaces, their dependence on empirical correlations restricts transferability to novel alloy regimes with limited training data.

In this talk, I will discuss the development of a hybrid workflow that integrates machine learning (ML) with mechanistic solid-solution strengthening models, centering on three fundamental physical descriptors: (i) lattice parameter, (ii) elastic constants, and (iii) atomic misfit volume. The HEA literature lacks consolidated databases for these descriptors. An automated prompt engineering workflow leveraging large language models (LLMs) was developed to extract experimental lattice constant data from scientific publications. Temperature-dependent elastic constants were determined with quantified uncertainty by combining ML and Bayesian inference with the phenomenological Varshni model. An active learning workflow coupled with density functional theory (DFT) calculations was employed to capture the non-linear electronic interactions critical for predicting atomic misfit volumes. The integration of these descriptors into a parameter-free edge dislocation model establishes a robust predictive methodology to address experimental data gaps and facilitate the accelerated design of high-strength alloys. In addition, my current focus is on utilizing the DFT data generated for BCC RHEAs to develop machine learning interatomic potentials (MLIPs). These models enable accurate property predictions at substantially lower computational costs compared to direct DFT calculations.

*We strongly encourage attendees to use their full name (and if possible, their UMD credentials) to join the zoom session.*