Data-driven strategies for next-generation scientific discovery in clean energy research

Abstract

The transition to a low-carbon economy requires fundamental advances in clean energy technologies. Functional materials are critical to designing more efficient devices for energy generation, storage, and transmission. However, discovering energy materials is an inherently time-consuming and resource-intensive process. Current strategies are primarily based on the synergistic interplay of physical models describing macroscopic materials characteristics and experimental techniques to probe their properties. This dissertation targets the amplification of cutting edge technologies for materials discovery with data-driven tools founded on machine learning (ML). We focus on formulating restructured discovery workflows and discussing their potential to approach clean energy challenges, including natural photosynthesis and artificial light-harvesting. In the first part, we outline the benefits of data-driven methods to the acquisition and interpretation of empirical evidence in the absence of tractable computational or experimental approaches. We demonstrate how ML tools can inexpensively estimate excitation energy transfer properties in natural light-harvesting complexes at higher accuracies than approximative physical models. We also highlight opportunities to simplify experimentation for organic electronics by identifying and exploiting statistical relations between experimentally inaccessible electronic properties from easily accessible optical properties. Finally, we showcase how ML can evidence mechanistic insights into the dynamics of light-matter interactions. We further focus on the development of data-driven tools enabling autonomous workflows to discover advanced solar cell materials. This next-generation approach to scientific discovery integrates automated platforms with data-driven methods to iteratively design, synthesize, and characterize materials candidates in closed-loops. In the second part, we formulate data-driven strategies to suggest promising materials candidates with real-time experimental feedback. In the third part, we detail algorithmic frameworks for the orchestration of autonomous workflows and demonstrate their benefits to clean energy research on two applications: (i) the discovery of conductive thin-film materials for perovskite solar cells, and (ii) the discovery of photostable polymer blends for organic photovoltaics. In both applications, the autonomous approach identified promising materials with more complex compositions in larger design spaces at reduced operations costs. Our findings suggest that transitioning to autonomous experimentation with predictive, intuitive, and interpretable data-driven tools can extend the boundaries of forefront technologies for the discovery of clean energy materials.