High-throughput computation and experiment now generate enormous datasets that machine learning can mine for structure–property patterns and use to guide the search for new materials with targeted functionalities. This workshop examines the full discovery pipeline, from dataset construction and surrogate model training through uncertainty-guided search to experimental validation, addressing the practical realities of deploying ML in a materials discovery programme.

The session covers the major open materials databases — the Materials Project, AFLOW, OQMD — and examines how surrogate models for formation energy, band gap, ionic conductivity, and transport properties are constructed, validated, and deployed. The comparative merits of high-throughput DFT screening and ML- guided search are assessed, together with multi-objective optimisation strategies for materials with competing property requirements. Case studies span new stable inorganic compounds (the GNoME campaign), solid electrolytes for battery applications, and photovoltaic absorber materials.

A critical theme throughout the workshop is the gap between computational prediction and experimental realisation — the “validation bottleneck” — and the strategies being developed to bridge it. Participants will leave with a comprehensive understanding of the ML-accelerated discovery pipeline, the infrastructure and data resources available, and the practical steps required to move from predicted candidates to synthesised and characterised materials.

Extended Key Takeaways & Outputs

• Comprehensive knowledge of open materials databases and their use in ML-guided discovery

• Understanding of surrogate model construction, validation, and deployment

• Comparative assessment of high-throughput DFT screening and ML-guided search strategies

• Practical knowledge of multi-objective optimisation for materials with competing requirements

• Realistic assessment of the experimental validation bottleneck and strategies to address it