Research

At LIAC, we develop artificial intelligence and machine-learning methods for chemical reasoning, synthesis planning, molecular and materials discovery, and experimental optimization. We aim to connect predictive models to the decisions that chemists make and, ultimately, to experiments that generate new knowledge.

Our main research directions include:

• language models, chemical reasoning, and tool-augmented agents;
• reaction prediction, synthesis planning, and reaction-space navigation;
• generative molecular design with synthesis constraints;
• adaptive experimentation for reactions, catalysis, and materials.

Selected Research Highlights

NoteSynthegy: strategy-aware chemical reasoning

Matter, 2026. Large language models guide chemical search algorithms using strategies written in natural language, enabling steerable retrosynthesis and reaction-mechanism elucidation.

Paper · Preprint

NoteSample-efficient generative molecular design using memory manipulation

Nature Machine Intelligence, 2026. Saturn combines the Mamba architecture with augmented memory to learn more from each scored molecule and directly optimize high-fidelity objectives.

Paper · Code

NoteTANGO: direct optimization of constrained synthesizability

Nature Computational Science, 2026. TANGO guides generative models toward molecules that use required starting materials or intermediates, making synthesizability part of the design objective.

Paper

NoteSwarm intelligence for chemical-reaction optimization

Chem, 2026. α-PSO combines interpretable particle-swarm dynamics with machine-learning guidance for parallel experimental optimization.

Paper · Code

NoteHighly parallel optimization of chemical reactions

Nature Communications, 2025. Minerva couples Bayesian optimization with automated 96-well experimentation to navigate large reaction-condition spaces.

Paper · Code

NoteA multi-modal transformer for predicting global minimum adsorption energy

Nature Communications, 2025. AdsMT predicts global minimum adsorption energies without enumerating binding sites and uses cross-attention to identify favorable adsorption locations.

Paper

NoteData-driven synthesis predictions of colloidal nanocrystal shapes

Journal of the American Chemical Society, 2025. Machine learning, electronic-lab-notebook data, Bayesian optimization, and experiments were combined to navigate copper-nanocrystal synthesis.

Paper

NoteChemCrow: augmenting large-language models with chemistry tools

Nature Machine Intelligence, 2024. ChemCrow demonstrated how a language model can plan and execute complex chemistry tasks by using purpose-built tools.

Paper · Code

Publications

For the complete and current publication record, see Google Scholar.