AI that connects data, models and experiments: • My research programme addresses a recurring difficulty in materials science: experiments are costly, data are sparse, the design space is enormous, and the variables that determine performance often involve not only composition or structure but also synthesis history, processing conditions and microstructure. • Small-data learning and molecular design: SMILES-X demonstrated that molecular properties can be learned directly from SMILES strings without task-specific human-engineered descriptors, while retaining interpretability through attention mechanisms. This work belongs to a broader effort on data-efficient molecular-property prediction and inverse molecular design. • Active learning and Bayesian optimisation: I develop approaches in which machine-learning models do not simply fit previous experiments but actively determine which experiment to perform next. These methods have been applied to adhesive formulations, permanent magnets, thermoelectric materials and a variety of composition- and process-optimisation problems. In thermoelectric research, active learning has enabled substantial performance improvements within only a few experimental cycles. In Nd–Fe–B magnets, machine-learning-assisted active learning was used to optimise fabrication conditions for high magnetic performance. • Autonomous and robotic experimentation: In rechargeable Li–O₂ batteries, data science was coupled directly to an automated electrochemical experimentation platform, allowing efficient exploration of a large multicomponent electrolyte space and leading to experimentally validated electrolyte formulations. • Generative AI across materials representations: My work has investigated generative modelling not only for molecular structures but also for materials microstructures. For example, I adapted a style-based generative adversarial network to generate realistic alloy microstructures from limited microscopy datasets. Current research extends generative approaches toward inorganic crystal discovery, distribution-aware model evaluation and independent physical validation. • Scientific language models and knowledge extraction: I investigate how large language models can understand materials science, retrieve scientific knowledge and transform literature into structured experimental information. This includes systematic benchmarking of LLM expertise in materials science as well as information extraction from scientific publications. • Toward trustworthy and synthesis-first AI: Current methodological work increasingly addresses uncertainty calibration, out-of-distribution generalisation, numerical uncertainty and the validation of AI-generated candidates. A central direction is developing AI methods that learn relationships among synthesis protocols, structures, and properties, with the goal of generating experimentally executable and reproducible routes to new materials.

Collaboration, translation and international research training: • A defining feature of my research is that AI methodology is developed together with domain scientists and then tested on real materials problems rather than isolated benchmark datasets. • Within NIMS, this involves collaborations across data-driven materials research, magnetic materials, energy materials, thermoelectrics, structural materials and experimental synthesis. International activities include collaborations within the CNRS–Saint-Gobain–NIMS LINK international laboratory and with researchers and universities in France and other countries. Industrially relevant work spans steels and alloy microstructures, magnets, polymers and adhesives, functional materials, formulations and manufacturing-process optimisation. Public examples include collaboration with JFE Steel on generative modelling of alloy microstructures. • My role frequently lies at the interface between materials expertise, physics, computation and experimentation: translating an experimental scientific question into a tractable AI problem; identifying what information can actually be learned from available data; constructing and validating computational workflows; determining the next informative experiment; and helping interpret the resulting scientific insight. • This philosophy is also central to my supervision. I regularly supervise and co-supervise international students and research interns coming to NIMS from universities in Europe, North America and Asia. Their projects span active learning and Bayesian optimisation, generative materials models, crystal generation and validation, synthesis–process–property relationships, scientific information extraction, uncertainty quantification and AI-assisted experimental design. The goal of these internships is not simply to teach a machine-learning technique. Students are trained to carry a project across the complete scientific chain: scientific question → data and hypotheses → computational method → uncertainty and validation → experimental or physical interpretation → reproducible research output. • Several of these projects are deliberately designed to mature toward journal publications, reusable software, research datasets, patents or longer-term international collaborations. • My broader objective is therefore to build an AI-enabled materials research environment in which computation enhances, rather than replaces, scientific expertise, experimental judgement and physical understanding.