Dr. Nabil Khossossi
Autonomous Energy Materials Discovery · Computational Electrochemistry · Equivariant Machine Learning
The transition to a carbon-neutral energy system demands catalysts and materials capable of driving electrochemical CO₂ reduction, N₂ fixation, hydrogen and oxygen electrocatalysis, and molecular energy storage with high activity, selectivity, and durability. The fundamental obstacle is not synthesis—it is discovery. The space of viable candidates spans millions of bimetallic, intermetallic, high-entropy alloy, and molecular configurations, and the Brønsted–Evans–Polanyi scaling relations that couple successive intermediate binding energies severely constrain the catalytic performance achievable within any single material class. Identifying compositions and molecular architectures that circumvent these constraints requires a systematic, theory-grounded exploration strategy that no purely experimental approach can provide.
My research builds autonomous computational frameworks that couple quantum-mechanical first-principles calculations with equivariant machine learning potentials and multi-objective optimization to navigate this space efficiently and rigorously—spanning extended solid surfaces, two-dimensional materials, and redox-active molecules. The aim is not incremental improvement of known systems, but the systematic identification of new material and molecular chemistries whose properties are predicted from electronic structure, validated computationally at scale, and delivered to experiment as actionable synthesis targets.
The Sabatier principle establishes that optimal catalysts bind reaction intermediates at a precise free-energy window—the volcano peak. For HER, this maps to a single descriptor ΔGH*, placing platinum near the apex and motivating the search for earth-abundant alternatives. For OER/ORR, the four-electron pathway introduces at least two independent adsorbate energies (ΔGOH*, ΔGOOH*), and the near-universal linear scaling between them imposes a thermodynamic overpotential floor of ~0.37 V that no binary metal surface can fully escape. CO₂ reduction and N₂ fixation present even richer constraint landscapes, with multi-step intermediate sequences whose binding energies are coupled through Brønsted–Evans–Polanyi relations across the entire reaction network.
I address these challenges by constructing end-to-end autonomous pipelines in which SE(3)-equivariant graph neural networks—trained on the OC20 dataset of over 260 million DFT-relaxed configurations—predict adsorption energies across >500,000 adsorbate–surface combinations at a cost three orders of magnitude below explicit DFT. Candidate materials are ranked through multi-objective Pareto optimization that simultaneously targets activity, thermodynamic and electrochemical stability, and elemental abundance—yielding synthesizable intermetallic and bimetallic leads as prioritised experimental targets.
Architectures that encode SE(3) or E(3) equivariance directly—such as EquiformerV2—achieve physical correctness by construction, yielding transferable potentials that reach near-DFT accuracy while enabling nanosecond-scale molecular dynamics at a fraction of the ab initio cost. I use these ML potentials to simulate electrode–electrolyte interface dynamics, solvation shells around adsorbed intermediates, and ion migration pathways in solid-state electrolytes at timescales inaccessible to AIMD.
In parallel, I develop generative models for crystal structure prediction—diffusion-based and flow-matching architectures that map from target property specifications to stable crystal geometries via inverse design. This shifts the discovery paradigm from screening existing databases to generating hypothesis-driven crystal candidates, including metastable phases and unconventional surface terminations that may break the scaling-relation ceiling for OER and CO₂R.
Beyond extended surfaces, I develop computational frameworks for designing functional molecules with precisely targeted electrochemical properties. In molecular electrocatalysis, transition-metal complexes—metalloporphyrins, MN₄ centres, metal-organic frameworks—offer programmable coordination environments that can decouple intermediate binding energies and circumvent the scaling relations constraining heterogeneous surfaces. Using GNN representations of molecular electronic structure with generative and active-learning approaches, I screen large libraries of complexes for CO₂R, N₂R, and ORR activity.
For electrochemical energy storage, I apply generative molecular design to redox-active organic molecules for flow batteries—targeting high and tunable redox potential, electrochemical stability over thousands of cycles, and electrolyte compatibility. The unifying principle: build quantitative structure–property relationships from first principles, encode them in transferable ML models, and navigate molecular design space autonomously toward property targets that experiment alone cannot efficiently reach.
Using the computational hydrogen electrode formalism, I construct atomically-resolved free-energy diagrams for the full multi-proton–electron transfer sequences of HER, OER, ORR, CO₂R, and N₂R—identifying the potential-determining step, quantifying the theoretical overpotential, and deriving electronic structure descriptors (d-band center, projected DOS, Bader charge transfer) that rationalise activity trends across material families.
These descriptors are benchmarked against experimental observables: Tafel slopes, exchange current densities, onset overpotentials, Faradaic efficiencies, and operando signatures (XAS, in situ Raman, DEMS). The feedback is bidirectional—experiments refine descriptor accuracy, and refined descriptors redirect the autonomous screening. Closing this loop between computation and characterisation is the central methodological challenge: every experimental measurement improves the model, and every model prediction generates a testable hypothesis.
The overarching goal is a tightly integrated computational–experimental laboratory in which the boundary between prediction and measurement dissolves. Autonomous screening pipelines generate ranked catalyst and molecular candidates grounded in first-principles theory; experimental groups synthesise and characterise the top predictions under operando conditions; and the resulting structural, spectroscopic, and electrochemical data flow back into the ML models—continuously refining the descriptor space and redirecting the next round of computation.
The four research directions above build the methodological infrastructure—equivariant potentials, generative crystal and molecular models, multi-objective optimisation, and operando-informed feedback loops—that makes this integrated programme possible. The long-term scientific objective is a quantitative, predictive theory of electrochemical catalysis that is fast enough to guide real experimental campaigns and accurate enough to explain why every prediction succeeds or fails.