Welcome to the Feugmo Research Group (Physics-Informed Energy Materials Lab (PIEML))

Our laboratory develops interpretable, multiscale computational frameworks to understand and predict electrochemical and thermochemical phenomena across atomic, interfacial, and device length scales. By integrating Physics-Informed Neural Networks (PINNs), machine learning interatomic potentials (MLIPs), advanced statistical mechanics, and differentiable structural refinement, we construct physically grounded “digital twins” of complex energy materials and electrochemical systems. Our goal is to replace empirical fitting with mechanistic, equation-level discovery—enabling predictive modeling of corrosion, ion transport, degradation, and interfacial dynamics in batteries, fuel cells, molten salts, and solid oxide devices.

We specialize in embedding governing physical laws directly into neural architectures, allowing partial differential equations (PDEs), thermodynamic functionals, and transport constraints to shape the learning process. This eliminates the need for massive labeled datasets while ensuring thermodynamic consistency and interpretability. Beyond simulation, we leverage sparse regression and symbolic learning to transform experimental signals—such as cyclic voltammetry (CV) and electrochemical impedance spectroscopy (EIS)—into closed-form governing equations. By uniting atomistic simulation, continuum theory, and data-driven discovery, our laboratory provides a quantitative foundation for the rational design of corrosion-resistant materials, high-temperature electrochemical systems, and ultra-long-life energy storage technologies.

Key Research Themes

1. Physics-Informed Neural Networks for Electrochemical Systems

We develop PINN-based solvers to model coupled nonlinear PDEs governing electrochemical interfaces, transport, and reaction kinetics. By embedding conservation laws and constitutive relations into the loss function, our models resolve complex multiphysics interactions in corrosion, electrocatalysis, fuel cells, and batteries without relying on large experimental datasets. These frameworks uncover hidden nonlinear couplings between atomic-scale structural evolution and macroscopic device behavior.

In parallel, we implement sparse regression and symbolic learning to discover degradation laws directly from noisy  data. Rather than assuming empirical power-law fading or resistance growth, we extract interpretable mathematical expressions describing lithium plating, catalyst agglomeration, interfacial resistance evolution, and long-term degradation mechanisms.

2. Neural Multiphysics of Solid Oxide Interfaces

We design specialized neural architectures to model ionic and electronic transport across electrode/electrolyte interfaces in solid oxide electrolysis and fuel cells. These models incorporate nonlinear dependencies of current density, defect chemistry, and interfacial reactions while explicitly capturing long-term cation interdiffusion—one of the primary degradation drivers in high-temperature ceramic systems. By embedding transport physics into the training objective, our approach predicts performance decay trajectories that traditional empirical fits cannot resolve.

3. Machine Learning Interatomic Potentials & Atomic-Scale Dynamics

To simulate materials under extreme chemical and thermal conditions, we develop high-fidelity neural network potentials that achieve near–ab initio accuracy at a fraction of the computational cost. Using equilibrium and non-equilibrium molecular dynamics (NEMD), we extract macroscopic transport coefficients—including viscosity, thermal conductivity, and diffusivity—directly from atomistic simulations.

Applications include corrosion in molten salt environments relevant to advanced nuclear reactors, ion transport across polycrystalline and two-dimensional materials, and defect-mediated kinetics in solid electrolytes. These GPU-accelerated simulations provide atomistically grounded parameters that inform our continuum PINN models—forming a unified digital twin workflow spanning atoms to devices.

4. Advanced Statistical Mechanics & Neural Functional Theory

We advance the theoretical foundations of Classical Density Functional Theory (cDFT) to model inhomogeneous ionic fluids and electrochemical interfaces. By introducing neural-parameterized functionals within Fundamental Measure Theory (FMT), we enable adaptive free-energy landscapes capable of resolving crystallization, polymorphism, and confined phase behavior.

We extend these equilibrium models into the time domain via Dynamic Density Functional Theory (DDFT), coupling chemical potential gradients with Nernst–Planck transport to capture non-ideal ion correlations and steric exclusion effects. This framework allows accurate modeling of electric double layer dynamics, nanoporous adsorption, and ion transport in concentrated electrolytes where classical Poisson–Boltzmann theory breaks down.

5. Gradient-Based Structural Refinement of Disordered Materials

We address the long-standing challenge of resolving atomic structure in amorphous and disordered solids through differentiable optimization. Using our open-source framework TorchDisorder, atomic coordinates are treated as trainable parameters constrained by experimental scattering data and chemical priors. By replacing stochastic Reverse Monte Carlo approaches with automatic differentiation and constrained Lagrangian optimization, we achieve rapid, physically consistent structural refinement of glasses and solid electrolytes.

These refined atomic models provide high-quality structural inputs for downstream machine learning potentials and density functional theory workflows, completing a fully differentiable materials discovery pipeline.

Our laboratory’s mission is to transition electrochemical modeling from empirical curve-fitting to predictive, physics-grounded equation discovery—enabling rational materials design across nuclear, energy storage, and high-temperature electrochemical technologies.