An Integrated Machine Learning Workflow Based on Deep Generative Models for Discovery of Inorganic Nonlinear Optical...

Area of Science:

• Materials Science
• Condensed Matter Physics
• Computational Chemistry

Background:

• Discovering nonlinear optical (NLO) materials for deep-ultraviolet (DUV) and mid-infrared (MIR) applications is challenging due to strict structural and property demands.
• Existing machine learning (ML) methods for predicting NLO properties are limited by insufficient data, hindering exploration of novel chemical spaces.

Purpose of the Study:

• To develop an advanced computational workflow for efficient discovery of new NLO materials.
• To overcome data limitations in machine learning by utilizing deep generative models for exploring uncharted chemical territories.

Main Methods:

• Developed an integrated workflow combining deep generative models and machine learning predictors.
• Generative models were trained without requiring pre-existing NLO property data, enabling exploration of novel structures.
• Generated crystal structures were filtered using ML predictors and validated through first-principles calculations.

Main Results:

• Identified 27 potential deep-ultraviolet (DUV) and 13 mid-infrared (MIR) nonlinear optical (NLO) materials.
• The discovered materials meet the stringent requirements for DUV and MIR NLO applications.
• The identified crystals exhibit significant chemical compositional and structural diversity.

Conclusions:

• The developed deep generative model workflow effectively accelerates the discovery of novel NLO materials.
• This approach bypasses the need for extensive experimental data, enabling broader exploration of chemical space.
• Paves a new pathway for discovering innovative NLO-active structural units for advanced optical applications.