Exploring self-driving labs for optoelectronic materials

Abstract: Self-driving laboratories (SDLs), by combining automation with machine learning-guided experiment selection, have the potential to transform experimental materials science. To date, most SDLs have been optimisation-driven, designed to rapidly converge on performance metrics, by embedding multiple mechanistic layers within platform-specific surrogate models. Such approaches excel at process tuning yet offer limited insight into the underlying physics governing synthesis-property relationships. Here we articulate a complementary paradigm: the exploration-driven, or scientific, SDL, whose primary purpose is the generation of data for data-driven science. We exemplify this concept for the case of inorganic optoelectronic materials, arguing that defect physics, which forms the central mechanistic link between synthesis conditions and functional properties, provides the foundation for designing a suitable SDL. Because defect populations and their spatial organisation cannot generally be resolved directly - nor fully predicted from first principles - the task of the SDL is to generate datasets in which thermodynamic and kinetic synthesis variables are systematically perturbed and defect-sensitive observables measured in parallel. From this basis, we propose a set of design principles for scientific SDLs that will enable them to operate close to the physics of optoelectronic materials, thereby generating transferrable and reusable datasets offering radical insight. We use Cu2ZnSn(S,Se)4 as a case study, both to show the scale of the task of defect-aware materials exploration as well to highlight as the deficiencies in the current paradigm. We propose that properly designed SDLs can generate the structured datasets necessary to enable mechanistic inference and advance synthesis-aware materials design.