research
Autonomous materials discovery across scales
Autonomous materials discovery across scales.
My research develops autonomous, physics-grounded AI systems that discover and design materials for extreme environments, neuromorphic and next-generation microelectronics, thermal management of high-performance computing, and advanced manufacturing. I work on composition–structure–property–processing relationships that make materials predictable by design. Because the space of possible materials — compositions, structures, defects, interfaces, processing histories — is astronomically large, discovery has historically relied on trial-and-error and intuition-guided synthesis. I combine quantum mechanics, atomistic modeling, mesoscale simulation, and autonomous experiments with reinforcement learning, graph neural networks, and multi-agent AI to explore this space with scientific intent rather than luck.
The objective is not to replace physical modeling with black-box prediction. It is to create computational and experimental systems in which physics provides the structure, AI chooses efficient actions, and verification constrains autonomy.
Scientific discovery is a multiscale decision problem.
Materials behavior emerges across many coupled length and time scales, and the information needed to design a material is incomplete, expensive, and scattered across calculations, experiments, and prior knowledge. How the pieces are connected matters as much as how good each piece is.
Traditional materials workflows are largely sequential: calculate an electronic structure, fit a model, run atomistic simulations, build a continuum description, compare with experiment, then iterate manually. Each handoff hardcodes assumptions, drops uncertainty information, and puts a human expert in the loop for tasks that could — in principle — be reasoned about automatically.
I ask how these stages can be made connected, adaptive, and auditable: how information (including uncertainty and mechanism, not just fitted parameters) should flow between scales, when an agent should request a new calculation instead of extrapolating, and how autonomy at each scale can be verified rather than trusted.
From electrons to engineering performance.
The core of the program is a closed computational stack in which information moves both forward toward performance and backward toward the next most informative calculation.
Bonding, defects, interfaces, phase stability, reaction energetics.
Physics-informed, uncertainty-aware potentials for realistic atomistic scale.
Kinetics, transport, interfaces, nucleation, defect motion.
Microstructure evolution, domain dynamics, multiphase processes.
Structure–property relationships, process windows, device and component behavior.
Current bottlenecks
Methods I develop
Four interconnected thrusts.
These are not independent topics. They share methods, software, data, and a common goal: make multiscale materials research more autonomous, predictive, and experimentally grounded.
Autonomous Scientific AI
Modern simulation codes and experimental instruments can capture highly complex physics, but they still depend on substantial human expertise to translate scientific intent into valid inputs, choose settings, detect failure modes, interpret outputs, and decide what to run next. I develop AI-agent systems that operate across these full workflows while retaining deterministic execution and explicit verification.
AutoMOOSE is the flagship platform in this direction — it decomposes autonomous phase-field simulation into specialized roles for architecture, input generation, execution, review, visualization, and adversarial checking. TEM-Scientist applies the same compile–rehearse–execute philosophy to autonomous transmission electron microscopy. The broader program extends these ideas to DFT, molecular dynamics, HPC workflows, and heterogeneous multiscale simulations.
Multiscale Materials Modeling
Many important materials phenomena cannot be described at one scale. Electronic structure controls bonding and defect energetics; atomistic dynamics determine kinetics and interfacial processes; microstructure evolves collectively at mesoscale; engineering performance emerges only after these mechanisms interact.
My work connects DFT → machine-learned interatomic potentials → molecular dynamics → phase-field and multiphysics models. A particular focus is on how uncertainty and mechanistic information should be transferred between scales rather than simply passing fitted parameters downstream.
Materials Design & Discovery
I work on a set of concrete materials-science problems where predictive control of structure, composition, and processing would meaningfully change what is possible — functional oxides for information technology, structural alloys for extreme environments, neuromorphic materials for AI hardware, thermal-management materials for high-performance computing, and metals and alloys for autonomous additive manufacturing. In each system the immediate goal is scientific: understand the mechanisms that connect processing, structure, and function. The methods in Thrusts 01, 02, and 04 exist to make that understanding faster and more reliable.
AI Methods & Scientific Infrastructure
Autonomous science depends on the underlying algorithms and shared infrastructure it runs on. I develop reinforcement learning, inverse design, and search algorithms alongside the datasets, provenance systems, and software that make them reusable across problems. My earlier work established this decision-centric view through RL, tree search, and inverse design; I now extend the same principles toward autonomous scientific agents.
CASTING uses continuous-action-space tree search for crystal-structure prediction and inverse design. BLAST applies reinforcement-learning strategies to interatomic-potential development. The Quantum Cluster Database catalogs more than 70,000 DFT-computed structures across 55 elements as an open resource for data-driven discovery.
From search algorithms to autonomous scientists.
A continuous arc: sequential-decision methods for materials, applied to progressively more of the scientific workflow.
Closed-loop, trustworthy autonomous materials discovery.
The destination is a research ecosystem in which simulations, experiments, data, and scientific agents continuously exchange information: models identify uncertainty, agents select the next action, deterministic tools execute it, and new evidence updates the scientific hypothesis.
Selected publications →