The search for better materials for energy conversion and storage is limited less by our understanding of the underlying physics than by the sheer size of the chemical and structural space to be explored. In this talk I will show how high-throughput first-principles calculations, large curated databases and machine-learned surrogate models together change the cost of this search by orders of magnitude. I will use the Alexandria database, with millions of DFT calculations, and graph-neural-network models for stability and property prediction as examples of how such tools turn exhaustive exploration into a practical strategy.

Screening identifies candidates, but device performance is decided elsewhere: by defects, surfaces and interfaces. I will therefore discuss how targeted higher-level calculations complete the picture. Examples include point-defect thermodynamics and doping limits in the earth-abundant photovoltaic absorber Zn₃P₂, and the surfaces and interfaces that control function. I will conclude by asking what still separates a computational prediction from a material that can be synthesised and used, and where data-driven methods can help close that gap.