Speaker
Description
Recent advances have demonstrated that neural networks can accurately predict optical spectra of a wide range of materials directly from atomic configurations. However, most existing approaches are limited to equilibrium structures at 0 K and therefore fail to capture temperature-dependent effects. Modeling such effects requires accounting for electron–phonon interactions, which is computationally demanding due to the need for extensive sampling of thermally perturbed configurations or explicit first-principles phonon calculations. Accurate temperature-dependent optical spectra are, however, crucial for applications such as photovoltaics and optoelectronic devices, where material performance is strongly influenced by thermal effects under operating conditions.
In this work, we present a computational framework that utilizes machine learning interatomic potentials to efficiently generate large ensembles of thermally perturbed atomic configurations via molecular Dynamics simulations. These configurations are combined with density functional theory (DFT) calculations to produce reference optical spectra, which are then used to train a neural network model. This approach enables the prediction of temperature-dependent optical spectra with accuracy comparable to the underlying DFT calculations at a fraction of the computational cost.
We demonstrate the performance of our method on III–V semiconductors, showing that the neural network captures the temperature response of optical spectra even when trained on limited data. Our framework provides a scalable pathway for studying finite-temperature optical properties and significantly accelerates the exploration of thermally driven phenomena in materials.