26th Australian Institute of Physics Congress and the Australian and New Zealand Conference on Optics and Photonics

End-to-End Machine Learning Framework for Rapid Spectrum-Adaptive Photovoltaic Cell Design

Not scheduled

1h 30m

Belinda Hutchinson Building (The University of Sydney )

Contributed Oral COMMAD - Optoelectronic and Microelectronic Materials and Devices Parallel sessions

Description

Photovoltaic devices are increasingly required to operate under diverse illumination environments, including standard solar irradiation, indoor lighting, space spectra, and thermophotovoltaic thermal emission. These spectrum-dependent operating conditions create a high-dimensional optimization problem in which material properties, device architecture, and incident spectral profiles are strongly coupled. Conventional optimization workflows rely on repeated physical simulations or surrogate-model-assisted iterative searches, which remain computationally expensive and inefficient for rapid scenario-specific device design.

Here, we present a universal end-to-end machine learning framework for direct spectrum-to-design optimization of photovoltaic devices. The framework first constructs accurate fully connected neural network surrogate models for monocrystalline silicon and perovskite solar cells, enabling rapid prediction of key device responses including external quantum efficiency, current–voltage behavior, and maximum power density. To generate reliable reference optima, we develop a hybrid optimization strategy combining topological persistence analysis with Nelder–Mead simplex refinement. This approach identifies robust high-performance regions in multidimensional design landscapes and avoids noise-sensitive or initialization-dependent local optimization. The resulting optimized datasets are then used to train an end-to-end neural network that directly maps incident spectral profiles to optimal device parameters through single-pass inference, eliminating the need for iterative online optimization.

The framework is validated on both silicon and perovskite photovoltaic systems, demonstrating accurate surrogate prediction and near-optimal end-to-end design performance across held-out spectral conditions. A thermophotovoltaic case study further confirms its applicability to non-standard thermal radiation environments, where optimal device parameters strongly depend on the emission spectrum. This work establishes a general and computationally efficient paradigm for rapid, spectrum-adaptive photovoltaic design, providing a practical route toward accelerated optimization of solar, indoor, space, and thermophotovoltaic energy conversion devices.