Research

Introduced a committee-of-experts architecture for machine learning potentials: multiple specialized models trained on curated subsets, with predictions weighted to favour the most informed expert. This divide-and-conquer philosophy yields potentials that are more accurate than monolithic models with essentially no additional computational cost—a new paradigm for scalable atomistic simulation.

Challenged the assumption that ML potentials merely interpolate training data, demonstrating that 80–100% of test configurations actually lie outside the convex hull of training points. Proposed a probability-density framework to rationalize why these models remain accurate in extrapolation regimes—a finding that redefines how we think about trustworthiness in atomistic ML.

Developed automatic protocols to select optimal atomic fingerprints from large candidate pools using data-intrinsic correlations, dramatically simplifying the construction of neural-network potentials. Applied to water and Al-Mg-Si alloys, the method finds the best trade-off between accuracy and efficiency—making high-fidelity ML potentials accessible without expert hand-tuning.

Systematically benchmarked four classes of molecular descriptors—from classical to geometric—to predict drug solubility in lipid excipients, uncovering that hydrogen-bond basicity and melting-point-related solid-state properties are the dominant drivers. By making atomistic interactions interpretable, this work provides rational design rules for lipid-based drug formulations.

Built a machine-learning framework that decomposes the electronic density of states into local, atom-centered contributions—enabling prediction of electronic structure for systems far too large for direct quantum-mechanical calculation. Applied to amorphous silicon across extreme thermodynamic conditions, revealing structure–property relationships invisible to conventional analysis.

Fused Bayesian statistics with the ShiftML chemical-shift predictor to quantify confidence in NMR crystal structure identification. Demonstrated on six organic molecular crystals that machine-learned NMR predictions can replace expensive first-principles calculations, opening a practical route for polymorph characterisation in pharmaceutical development.

Replaced the conventional thermodynamic convex hull with a data-driven, probabilistic construction that maps the full structural diversity of candidate compounds without bias. By working in learned coordinates, the framework identifies synthesizable structures and the experimental constraints most likely to stabilize them—turning crystal structure prediction from a brute-force search into a guided discovery process.

Community-wide benchmark of 22 groups ranking crystal structures by stability. Showed that periodic DFT-D methods match experiment within expected margins, while system-specific ML potentials emerged as a viable fast alternative—establishing the state of the art for polymorph energy ranking at industrial scale.

The generation counterpart to the ranking blind test: 21 groups sampled crystal structure landscapes for increasingly complex pharmaceutical targets. Revealed that random search and ML-guided generation now reliably find the experimental polymorph for rigid molecules, but flexible and multi-component systems remain an open frontier.

Presented at NeurIPS AI4MAT, this work automates the classification of pharmaceutical solid forms—a bottleneck in early drug development—using ML models trained on diffraction and spectroscopic data. Enables rapid polymorph triage without expert intervention, directly accelerating the path from candidate molecule to developable drug.

Developed a machine-learning reranking scheme that accelerates crystal structure prediction by orders of magnitude, validated on a diverse set of real pharmaceutical molecules. The speedup unlocks both higher-throughput screening of drug candidates and the use of more accurate levels of theory, pushing CSP from an academic exercise toward an industrial-scale tool.

Introduced a first-of-its-kind agentic AI system that couples large language models with experimental databases and machine learning to automate the preformulation pathway for drug candidates. APPA reasons over scientific publications, experimental data, and ML predictions to propose optimal strategies, cutting weeks of manual evidence gathering to minutes.