Conrard Giresse Tetsassi Feugmo

Conrard Giresse Tetsassi Feugmo is an Assistant Professor in the Department of Chemistry at the University of Waterloo, where he has led a research group since 2022.

He studied chemistry at the Université de Yaoundé I in Cameroon, taking a BSc in 2007 and an MSc in 2010, then moved to Belgium for an M.Phil. in nanotechnologies at UCLouvain and a PhD in computational chemistry at the University of Namur, awarded in 2018. He came to Canada as a postdoctoral associate at Western University and was a research officer at the National Research Council of Canada before joining Waterloo.

His group works on how materials degrade and fail in demanding electrochemical environments: the molten salt coolants proposed for advanced nuclear reactors, the interfaces that set the lifetime of batteries and fuel cells, the oxide films that protect reactor piping. The common difficulty is that the governing equations are known but cannot be solved at the scales where measurements are made, so the field substitutes empirical fits that carry no mechanism and fail outside the conditions they were calibrated in.

His response is to write the physics as differentiable programs, so that one piece of code can predict forwards, recover mechanism backwards from sparse experimental data, and hand back a design. Machine learning enters as a way to solve and invert the governing equations rather than to replace them. Embedding those equations in the model keeps it thermodynamically consistent, makes its parameters mean something, and removes the need for the large labelled datasets that materials problems rarely have.

He is as interested in where the approach fails as in where it works. Much of his published work reports a boundary rather than a capability: which parameters a corrosion measurement can and cannot determine, how many decades of data an anomalous-transport claim actually requires, when a neural solver’s advantage belongs to the network and when it belongs to the comparison it was tested against. Stating those limits is what separates a mechanistic model from a fit that happens to look mechanistic.

His group releases its methods openly at github.com/Feugmo-Group, and works with industrial partners in the electric-vehicle battery supply chain alongside its academic collaborations.

• Grouped rather than listed. A visitor should be able to tell in five seconds whether you do what they came looking for.

• Methods. Differentiable physics · Physics-informed neural networks · Spectral and spectral-element solvers · Automatic differentiation · Machine-learned interatomic potentials · Classical and dynamical density functional theory · Neural functional theory · Differentiable optimization and inverse design · Sparse regression and symbolic discovery

• Physics. Ion transport · Electric double layer · Charge transfer kinetics · Corrosion and passive-film growth · Degradation mechanisms · Defect chemistry · Phase behaviour · Anomalous diffusion and fractional transport · Identifiability and inverse problems

• Systems. Solid-state and lithium-ion batteries · Solid oxide electrolysis cells · Fuel cells · Molten salts · High-entropy alloys · Nuclear materials · Nanoporous adsorbents and separations

• Scales. Electronic structure → atomistic dynamics → interfacial statistical mechanics → device response

Electrochemistry rests on equations we trust and cannot solve at the scales where measurements are made. Ångström-thick charged layers sit inside micrometre devices, and rates span twelve orders of magnitude. The field closes that gap by fitting: equivalent circuits for impedance, Langmuir isotherms for adsorption, power laws for oxide growth. A fit carries no mechanism and fails outside the window it was calibrated in.

Our group writes the physics as differentiable programs instead. One code then performs forward prediction, inverse recovery of mechanism from sparse experimental data, and gradient-based design, because gradients pass through converged solutions. We pair every inversion with a measurement of how much the data actually determine, so a recovered parameter arrives with a statement of whether it is identifiable at all.

Solvers that survive stiffness

Physics-informed neural networks break on exactly the equations electrochemistry needs: coupled Poisson–Nernst–Planck transport with kinetic boundary conditions, where residual scales differ by many orders of magnitude. We traced that failure to three structural causes and built the Neural Spectral-Element Method, which replaces random collocation with fixed spectral nodes and automatic differentiation with a precomputed differentiation matrix. The loss becomes deterministic, and residuals reach 10⁻⁹ to 10⁻¹⁰ on stiff benchmarks, roughly seven orders below the floor where standard formulations stall. A companion study benchmarks eleven stabilisation strategies against a finite-volume reference and identifies which of them earn their cost.

Statistical mechanics of the interface, without fitted parameters

Adsorption and double-layer behaviour are normally described by fitted isotherms and point-ion theories that say nothing about why a material works. Classical density functional theory predicts the structure of a fluid at an interface from a free-energy functional instead. We have rebuilt it as GPU-native differentiable code that takes any host and any fluid through a single interface, accepting analytic force fields or pretabulated first-principles grids. For CO₂ in aluminium formate it reproduces the 298 K isotherm to 0.44 mmol g⁻¹ with nothing fitted, and settles what the material is doing: an equilibrium selectivity near 4 against a measured 350 to 600 means the separation is kinetic, not thermodynamic. In a parallel line we make the functional itself learnable, treating fundamental measure theory coefficients as neural functions optimised end-to-end through the solver, which recovers Percus–Yevick behaviour unprompted and wall-contact densities within one to two percent of molecular dynamics.

Structure by gradient rather than by chance

Disordered and multi-component structures are built by stochastic search, which scales badly and gives little control over what the result satisfies. TorchDisorder replaces reverse Monte Carlo moves with gradient descent through a differentiable structure-factor engine, so a measured diffraction pattern is inverted rather than sampled towards. Silica, germania and lithium thiophosphate electrolytes fit above R² = 0.955 in 5,000 steps from a single dataset each. Our Neural Evolution Structure method does the same for high-entropy alloys, roughly 1000 times faster than special quasi-random structures, and two independent groups have built on it. These structures feed machine-learned interatomic potentials that reach near-quantum accuracy in GPU-accelerated molecular dynamics, from which we extract viscosity, thermal conductivity and diffusivity for molten salts, solid electrolytes and irradiated alloys.

What a measurement can and cannot determine

The theme that runs through the group’s work is knowing where a model stops being supported by the data. Inverting the point defect model for passive-film growth recovers film thickness within 2.2% at every applied potential under 5% measurement noise, and shows that parameter recoverability tracks stiffness: a boundary-stiff kinetic constant is identifiable from these data, and a weakly coupled interior coefficient is not, however well the fitted curve reproduces the measurements. We extend the same question to anomalous transport, where degradation and diffusion data are routinely fitted with fractional-order and general-kernel models. Using conservation-consistent nonlocal Stefan formulations and serial-correlation-robust intervals, we ask when an anomalous exponent is a property of the system rather than an artifact of the model or the fitting convention, and we document how physics-informed inversion fails quietly, since a soft inversion can satisfy the data while violating the equation and is caught only by re-solving forward.

Where it is applied

The methods run on solid-state and lithium-ion batteries, solid oxide electrolysis cells, fuel cells, molten salt systems for advanced nuclear reactors, high-entropy alloys, and nanoporous materials for carbon capture and hydrogen storage. All of it is released openly at github.com/Feugmo-Group, because a method nobody else can run is not a contribution to the field.

Over the next decade we aim to make transport modelling in electrochemical systems an inverse science: parameters recovered from experiment with a stated confidence, one set of equations carried from the double layer to the full cell, and the model run backwards to design.

• 2018, PhD, Computational Chemistry, University of Namur, Belgium

• 2011, Mphil, Nanotechnologies, Louvain School of Engineering, Belgium

• 2010, MSc, Chemistry, University of Yaounde I, Cameroon

• 2007, BSc, Chemistry, University of Yaounde I, Cameroon

• 2022, Talent Bursaries Alberta AI-week

• 2019, Center for Nonlinear Studies at Los Alamos National Laboratory travel grants

• 2018, Western’s Postdoctoral Fellowships Program

• 2014, C.G.B. (Comité de Gestion du Bulletin) - C.B.B travel grants

• 2014, Gordon Research Conferences travel grants

• 2012, Institutional PhD CERUNA grants

• 2011, Special Research Fund (FSR) Scholarship

• 2023-present, Science Faculty Council, Chemistry Representative

• 2024-present, Chemistry Awards Committee

• 2022-2024, Equity and Diversity Council, Science Representative

• Member, Waterloo Artificial Intelligence Institute

• Member, Waterloo Institute for Nanotechnology

• CHEM 120 - General Chemistry 1
  • Taught in 2025
• CHEM 123 - General Chemistry 2
  • Taught in 2023, 2024
• CHEM 400 - Special Topics in Chemistry
  • Taught in 2026
• CHEM 740 - Selected Topics in Theoretical Chemistry
  • Taught in 2026
• NE 451 - Simulation Methods
  • Taught in 2023, 2024, 2025, 2026
• NE 452 - Special Topics in Nanoscale Simulations
  • Taught in 2026

* Only courses taught in the past 5 years are displayed.

• Gore, A., Gouws, X. & Feugmo, C. G. T. TorchDisorder: A Differentiable Framework for Generating Physically Realistic Disorder Structures from Experimental Diffraction Data. J. Chem. Theory Comput. 22, 6859–6873 (2026).

• Feugmo, C. G. T. & Pankaczy, D. Neural spectral element methods for stiff multiphysics PDEs with electrochemical transport benchmarks. Mach. Learn.: Sci. Technol. 7, 045053 (2026).

• Farooqi, M., Bösing, I. & Feugmo, C. G. T. Physics-informed neural networks for the point defect model: Solving and inverting passive-film growth kinetics. APL Mach. Learn 4, 036110 (2026).

• Roy, A., Mathanamohan, L., Bao, S., and Tetsassi Feugmo, C. G.* A modular classical density-functional framework for gas adsorption in nanoporous materials: from first-principles binding energies to kinetic molecular sieving. J. Phys. Chem. C 2026; https://doi.org/10.1021/acs.jpcc.6c03450 .

• Katai, A. B., Ramasimha Varma, A. V., and Tetsassi Feugmo, C. G.* Computational insights into the corrosion behaviour of NbMoTaW and NbMoTaWV high-entropy alloys in molten fluoride salts. Faraday Discussions 264, 400–421 (2026).

• Tetsassi Feugmo, C. G.*, and McConville, T. Materials discovery is a composition problem: the case for agentic AI over bigger models. Digital Discovery, 2026, https://doi.org/10.1039/D6DD00338A

• Tetsassi Feugmo, C. G.* Neural-network parameterization of fundamental measure theory: from hard spheres to Lennard-Jones fluids. Physical Review E 114, 025302 (2026).