Research

Developing practical machine learning tools for accelerating catalyst screening at scale. ARK (Angular Relational Knowledge Distillation) compresses large teacher MLIPs into 7× smaller student models with 50× speedup, preserving angular interaction fidelity critical for surface reactions. CatBench systematically evaluates 14 universal MLIPs on 47,000+ adsorption reactions with multi-class anomaly detection, revealing failure modes invisible to standard metrics.

We also develop domain-specific finetuning pipelines that adapt pre-trained MLIPs to target catalytic systems achieving DFT-level accuracy from limited training data, and generative diffusion models for designing thermodynamically stable bimetallic nanocatalysts. CatAgent, our autonomous multi-agent LLM framework, orchestrates hypothesis generation, surrogate-model evaluation, and critic feedback for electrocatalyst discovery, achieving 2.27× discovery rate over random search.

Closing the loop between AI predictions and experimental reality. Using MLIP-driven molecular dynamics, we simulate complete catalyst nanostructures up to 10,000 atoms at nanosecond timescales — bridging atomic-scale DFT accuracy with mesoscale dynamics that traditional surface slab models cannot capture.

Our multiscale approach spans from single-atom catalysts on oxide supports to realistic nanoparticle morphologies, revealing size-dependent oxygen transfer mechanisms, lattice oxygen kinetics in hierarchical ceria architectures, and facet-dependent reconstruction under reaction conditions. We are extending this framework to electrochemical interfaces, modeling explicit solvation, proton-coupled electron transfer, and electrolyte ion effects at catalytic surfaces. All predictions are validated by ambient-pressure XPS, in-situ Raman, XANES, and synchrotron characterization through direct collaboration with experimentalists.