Faculty Statement
Research Vision
AI4Science and Science4AI: Theory and Computation for Driven Open Systems
Physical systems in modern technologies operate far from equilibrium. Examples include electrochemical interfaces that drive catalysis and ultrastable glass coatings—all driven by forces such as high applied voltages and extreme pressures that push them to performance limits and failure. The ideal tools for predicting their behaviors are ab initio (first-principles) physics-based simulations, which remain intractable at realistic scales. While machine learning (ML) has begun to close this gap, most models are trained on equilibrium data and fail under far-from-equilibrium conditions. Intriguingly, AI systems face a structurally identical challenge: AI “agents” that autonomously propose experiments and retrieve external knowledge see their reliability degrade when driven far from their training distribution. These examples show how physical and AI systems share a common problem in which being pushed far from their initial conditions makes predictive control intractable.
To address this challenge, my proposed research proceeds in a bidirectional framework: I will advance AI to enable predictive control of systems far from equilibrium, while using the tools of physics to establish foundations for robust and explainable AI systems. My proposal therefore motivates not only AI4Science applications but also a Science4AI direction. It builds on my experience in quantum dynamics of photonic systems, statistical physics of disordered systems, and the integration of AI with molecular simulations (including Open Molecules 2025 and work on equivariant neural networks for electronic structure and active learning for rare-event sampling). I propose three thrusts:
- Thrust 1: Artificial Intelligence for Ab Initio Electrochemistry. Developing datasets and models for simulating molecular dynamics at electrochemical interfaces, with applications in electrocatalyst discovery.
- Thrust 2: Autonomous Photonic Control of Glass Formation. Combining multiscale simulation with AI to discover photonics-based protocols to manufacture new forms of glasses.
- Thrust 3: Statistical Physics for Robust Agentic AI. Developing statistical-physics theory for failure modes in multi-agent and tool-augmented AI, with applications in robust materials discovery.
Over the next five years, I will build computational infrastructure for driven physical systems and develop a fundamental understanding of the reliability of AI systems deployed in this setting.
Thrust 1: Artificial Intelligence for Ab Initio Electrochemistry. Electrochemical interfaces govern clean-energy conversion and storage, yet their molecular reactivity remains difficult to predict and control. The key difficulty is that interfacial disorder (hydration, defects, roughness, adsorbate fluctuations) couples directly to nonequilibrium electron transfer under bias and current. As experiments move toward spectroscopy under operating conditions, theory and simulation must produce observables that connect quantitatively to measurement. Machine learning interatomic potentials offer a route toward scalable, ab initio-quality simulation of reactive interfaces, but existing datasets underrepresent driven, disordered, nonequilibrium states that dominate electrochemical environments. In this thrust I will make predictive ab initio simulation of electrochemical interfaces feasible with AI by advancing: (1.1) theory for coupled electron–nuclei dynamics under bias (MASH–NEGF and QM/MM); (1.2) the Nonequilibrium Electronic Structure Database (NeESDB) and equivariant models that accelerate nonequilibrium interfacial MD; and (1.3) active learning targeting rare events and corrosion-resistant materials design, with an application to autonomous discovery of multi-component electrocatalysts for CO₂ reduction.
Thrust 2: Autonomous Photonic Control of Glass Formation. Every amorphous material passes through the supercooled liquid state before kinetic arrest below the glass transition. In the arrested state, glasses relax toward equilibrium through aging—a process that can take billions of years—which prevents annealing glasses to their ideal limit and limits access to ultrastable glasses. This motivates the “billion-year question:” Can we bypass aging to access the most optimal ultrastable glass? I propose to bridge glass engineering and molecular polaritonics (strong light–matter coupling in optical cavities) to establish control of glass formation through polaritonic coupling, using AI to discover optimal protocols. I will pursue: (2.1) theory for cavity-assisted glass formation and the cavity configurational feedback (C²F) protocol; (2.2) a Maxwell–MD framework coupling molecular dynamics with real-time electromagnetic fields and ML surrogates for the electromagnetic solver; and (2.3) ML-driven discovery of glass-forming protocols via Floquet engineering and Bayesian optimization, charting nonequilibrium phase diagrams and targeting states such as time-crystalline glass.
Thrust 3: Statistical Physics for Robust Agentic AI. Agentic AI systems (LLMs interfacing with tools and retrieval) are increasingly used for scientific discovery with minimal human supervision, but their performance remains brittle—systematic failures include inappropriate benchmark selection, resource exhaustion, and agents getting stuck in repetitive tool loops. I propose treating agentic AI from the lens of statistical physics as open systems driven out of equilibrium, where retrieval and tool use push the model beyond its training distribution at each step. I will: (3.1) use transition path theory (TPT) and committor analysis to characterize agent trajectories, identify mechanistic bottlenecks, and turn reliability into an interpretable diagnostic (validated on catalyst discovery); and (3.2) develop mean-field theory and scaling laws for multi-agent discovery, testing the hypothesis of a coupling-driven phase transition from diverse exploration to synchronized “groupthink.”
Teaching Philosophy and Course Contributions
I teach with empathetic rigor: meeting students where they are, using AI transparently, and training them to verify, explain, and own results they trust.
Philosophy and Approach. My teaching philosophy centers on empathetic rigor: be kind, understand student backgrounds, and pull them into rigorous scientific reasoning. This philosophy is especially crucial as students navigate how to use AI critically and responsibly in their work. To achieve this, I require that every homework assignment include a brief AI note documenting the tools used, prompts, and verification steps. Periodic in-class AI audits then identify hidden assumptions in model-generated solutions, reinforcing the importance of verifying against limiting cases and using mathematical/physical intuition. This approach cultivates the ability to critically evaluate and take ownership of one’s work; a skill increasingly vital as materials science and quantum chemistry rely more heavily on machine learning and computational modeling.
Teaching Experience. At Berkeley, I co-taught undergraduate chemical engineering thermodynamics with Prof. Carlo Carraro, working with students from diverse backgrounds. I designed discussion worksheets to complement his lectures, working through sample problems, limit cases, and dimensional/‘sanity’ checks to build their intuition. Through this approach, we covered thermodynamic postulates through azeotropic phase diagrams and reaction thermodynamics without sacrificing rigor. Since then, embedded in a chemistry department, I’ve further developed my communication skills by engaging with students studying electronic structure theory and spectroscopy. This interdisciplinary immersion has positioned me to communicate fluently across all mathematics–physical sciences interfaces.
Course Contributions. My interdisciplinary background in computational materials science, nonequilibrium statistical mechanics, and energy materials positions me to contribute meaningfully to both foundational and specialized courses. At the undergraduate level, I can teach courses in thermodynamics and phase equilibria, transport phenomena and diffusion, continuum mechanics (solids and fluids), polymer science and soft matter, ceramics and glasses, and electrochemical processes. At the graduate level, I can teach core courses in materials science principles, thermodynamics of materials, and statistical mechanics, as well as specialized courses in glass physics, soft matter physics, and computer simulations of materials. My background also positions me to offer specialized topics courses in computational electrochemistry, nonequilibrium statistical mechanics for energy materials, and deep learning in materials science.
New Course Development. I am passionate about developing courses that bridge theory, computation, and applications across disciplines. I propose three graduate courses that provide fundamental training in nonequilibrium phenomena and applications to energy science and engineering:
Course 1: Nonequilibrium Statistical Mechanics: Classical. This course establishes the foundations for understanding transport and relaxation in classical many-body systems, ranging from soft matter to electrolytes. Beginning with the origin of irreversibility and the second law, the course develops: (i) Irving–Kirkwood theory for transport phenomena, obtaining microscopic expressions for stress, heat flux, and pressure; (ii) Onsager regression hypothesis connecting spontaneous fluctuations to linear response; (iii) Green–Kubo theory for computing transport coefficients from equilibrium correlation functions; (iv) stochastic processes and Fokker–Planck equations for modeling thermal fluctuations; (v) Mori–Zwanzig projection formalism for memory effects and generalized Langevin equations; and (vi) mode-coupling theory for slow structural relaxation. Students will implement these methods in molecular dynamics simulations and derive transport coefficients in materials from first principles.
Course 2: Nonequilibrium Statistical Mechanics: Quantum. This course develops semiclassical and quantum dynamical methods for treating electron-nuclei dynamics beyond the Born–Oppenheimer approximation. Topics include: (i) time-dependent perturbation theory and Fermi’s golden rule for transition rates; (ii) semiclassical approximations (WKB, stationary phase, uniform approximations) and their validity regimes; (iii) path integral formulation of quantum mechanics and imaginary-time methods; (iv) Linearized Semiclassical Initial Value Representation (LSC-IVR) for real-time quantum correlation functions; (v) mixed quantum-classical approaches including Ehrenfest dynamics, trajectory surface hopping (fewest switches, decoherence corrections), and mapping Hamiltonians; and (vi) ring polymer molecular dynamics (RPMD) for computing quantum thermal rate constants and approximating real-time dynamics. Students will implement these methods and critically assess when semiclassical approximations break down.
Course 3: Asymptotics, Numerics, and Informatics for Materials Scientists. This methods course equips materials scientists and engineers with essential mathematical and computational tools, bridging traditional statistical reasoning with modern AI systems. Topics include: (i) asymptotic analysis (boundary layers, multiple scales, WKB theory) for quantum transport, electromagnetic problems, and phase-field models; (ii) numerical ODE/PDE methods including finite element and spectral methods with convergence analysis; and (iii) informatics spanning classical probability and statistics through machine learning, uncertainty quantification, and AI-agent engineering (retrieval-augmented generation, model context protocols, agentic workflows) for experimental data analysis. Readings from The Signal and the Noise by Nate Silver and Patterns, Predictions, and Actions by Hardt and Recht support critical reflection on modeling choices. This course trains students to build, validate, and deploy models while wielding AI critically; all are essential skills for contemporary materials research.
Throughout all three courses, I will implement AI-aware pedagogy: students document AI tool usage, verify model-generated code against analytical limits, and defend their computational choices through dimensional analysis, convergence tests, and comparison with experimental data.
Closing. My goal is to cultivate students who take ownership of their calculations and models, who understand not just what the answer is but why it must be so, and who can defend their work against scrutiny while remaining open to critique. My training brings quantum mechanical rigor, statistical mechanics intuition, and multiscale thinking. Paired with empathetic pedagogy, AI-aware instruction, and a commitment to inclusive advising, I am prepared to contribute to advancing materials research through computational innovation. I aim to train the next generation to be not only technically proficient but also collaborative, curious, and committed to innovation and sustainability through computational materials design.