Prediction of Frequency-Dependent Optical Spectrum for Solid Materials: A Multioutput and Multifidelity Machine...

Area of Science:

• Computational materials science
• Machine learning for materials science

Background:

• Frequency-dependent optical spectra are crucial for optoelectronics and energy harvesting.
• Density functional theory (DFT) provides accurate but computationally expensive data.
• Existing machine learning (ML) models often predict scalar properties, not complex optical spectra.

Purpose of the Study:

• To develop a deep graph neural network (GNN) model for predicting frequency-dependent complex dielectric functions directly from crystal structures.
• To explore multi-output and multi-fidelity learning strategies to handle limited high-accuracy DFT data.
• To model solar cell absorption efficiency metrics and enhance learning of the absorption coefficient.

Main Methods:

• Utilized deep graph neural networks (GNNs) for predicting complex dielectric functions.
• Investigated various GNN architectures for spectral multi-output representation.
• Employed transfer learning and fidelity embedding for multifidelity learning.
• Integrated solar cell absorption efficiency metrics into the learning process.

Main Results:

• Accurate prediction of frequency-dependent complex dielectric functions across IR, visible, and UV spectra.
• Demonstrated improved learning of solar cell absorption metrics through integrated learning bias.
• Showcased the effectiveness of multi-output and multi-fidelity ML for optical spectra prediction.

Conclusions:

• Deep GNNs accurately predict optical spectra from crystal structures.
• Multi-output and multi-fidelity ML techniques overcome data scarcity challenges.
• This approach provides a versatile tool for rapid material screening in optoelectronics and solar energy.