Physics-Informed Machine Learning for Rapid Characterisation of Semiconductors

The development of more efficient and stable semiconductor energy generation devices, such as photovoltaics (PVs), is key to achieve the EU’s 2050 decarbonisation targets. Tandem architectures, which stack multiple semiconductor junctions, are a promising technology that can enable more efficient and longer-lasting optoelectronic devices that far surpass those of single junction architectures. However, optimizing such complex multi-junction systems requires navigating a vast material and deposition parameter space that cannot be addressed through trial-and-error or conventional modeling alone. MatLearn introduces a novel physics-informed digital twin framework that combines drift-diffusion simulations with machine learning. By embedding semiconductor physics into Physics-Informed Neural Networks (PINNs), we aim to achieve physically consistent surrogate models that go far beyond black-box approaches, reducing data needs and providing interpretable insights. These models will enable processing and rapidly inferring key material and device parameters from large mutlimodal operando microscopy data, which would be impossible with existing approaches. This will allow us gain insight into nanoscale disorder in tandem devices under real-world operation to link to device performance and stability. Dr. Brenes brings unique expertise from academia and industry, spanning semiconductor physics, high-performance computing (HPC), device modeling, and automation. In this project, he will expand his skills in advanced characterization, supervision, and collaborative research, while transferring industry-driven approaches to data and modeling. By merging physics-informed modeling, machine learning, and operando measurements, MatLearn will deliver transferable, open-source tools for the rapid optimisation of tandem semiconductor devices, driving Europe’s progress toward clean energy and sustainable technology.