Artificial intelligence strategies to accelerate research on next-generation photovoltaic technologies

The development cycle of new materials-based technologies typically spans 15–25 years due to the increasing complexity of experimental research processes. Accelerating this cycle is essential for the large-scale deployment of sustainable technologies supporting the green energy transition, in which photovoltaics (PV), particularly novel thin-film photovoltaic (TFPV) technologies, play a central role. Recent advances in artificial intelligence (AI), explainable artificial intelligence (XAI), robotics, and automation have enabled a new paradigm: the data-driven materials science, capable of accelerating materials discovery and optimization. However, the application of XAI to real experimental characterization datasets remains limited, and translating machine learning (ML) outputs into physically meaningful insights is still a major challenge.

This thesis addresses that challenge by developing AI-driven strategies for the analysis and interpretation of experimental characterization data, following a proposed interpretability framework aimed at both accelerating research and enhancing explainability of ML models. It begins by implementing and optimizing different training strategies for generative neural networks using benchmark datasets, providing general grounding in ML implementation for experimental datasets. Then, an initial XAI-driven approach for kesterite-based TFPV devices is introduced, demonstrating the potential of interpretable ML models for identifying correlations between the physicochemical properties of materials and the PV performance of devices.

Finally, building on this progression, an integrated XAI-driven methodology is implemented to analyze multimodal datasets from holistic characterization of kesterite-based solar cells, covering the entire research cycle from synthesis and characterization to interpretation and feedback extraction. The methodology identifies the physicochemical properties that govern device performance and reproducibility, linking the relevant spectral regions with the corresponding physical properties. The resulting feedback is experimentally validated through modified fabrication processes, resulting in improved device performance, and reduces optimization cycles from over a year to less than a week.

Taken together, the work shows how transparent, physically meaningful AI can accelerate the development of new materials, contribute to the development of self-driving laboratory approaches and, ultimately, support the green energy transition.