§ Research · Four pillars

The work.

Machine learning systems that reason about physical structure — from atomic coordinates to crystal lattices to drug candidates. The question across it all: how do we design architectures that are accurate, interpretable, and reliable enough to trust at production scale?

§ 01

Research pillars

4 themes · 0 publications

010 papers

Building the models that simulate atoms.

ML Potential Architectures

Neural networks that predict forces on atoms accurately, efficiently, and — crucially — know when they cannot. Committee-of-experts architectures, automatic descriptor selection, extrapolation diagnostics.

Committee-of-ExpertsDescriptor SelectionExtrapolation Theory

020 papers

From atomic coordinates to macroscopic behaviour.

Structure–Property Learning

Frameworks that decompose bulk properties into atom-centred contributions — making predictions interpretable and the physics visible. Electronic DOS, Bayesian NMR, solubility in lipid excipients.

Electronic DOSBayesian NMRMolecular Descriptors

030 papers

Which polymorph will a drug molecule form?

Crystal Structure Prediction

A drug's crystal form decides whether it dissolves, how it's manufactured, and whether it's patentable. ML reranking pipelines, blind-test benchmarking, solid-form classification — turning CSP into an industrial decision tool.

ML RerankingBlind-Test BenchmarkSolid-Form Classification

040 papers

Autonomous systems that reason like scientists.

Agentic AI for Drug Discovery

The frontier: LLM-based agents that automate multi-step scientific reasoning. APPA couples language models with experimental databases and ML predictions to replace weeks of manual drug preformulation.

LLM AgentsTool-Use ReasoningEvidence Synthesis

§ 02

Publications by theme

Data via Google Scholar · auto-refreshed