Interpretable Surrogate Learning for Electronic Material Generation

ACS Nano. 2024 Nov 1. doi: 10.1021/acsnano.4c12166. Online ahead of print.

Abstract

Despite many accessible AI models that have been developed, it is an open challenge to fully exploit interpretable insights to enable effective materials design and develop materials with desired properties for target applications. Here, we introduce an interpretable surrogate learning framework that can actively design and generate electronic materials (EMGen), akin to producing updated materials with requirements by screening all possible elements and fractions. Taking the materials system with required band gaps as a case study, EMGen exhibits a benchmarking predictive error and a running time of 1.7 min for designing and producing a structure with a desired band gap. Using EMGen, we establish a large hybrid functional band gap database, and more uplifting is that the proposed EMGen effectively designs GaxOy with a wide band gap (>5.0 eV) for deep ultraviolet (DUV) optoelectronic devices, enabling a breakthrough extension of the applicability of GaxOy films in photodetectors to DUV light below 240 nm. The augmented band gap also helps improve the breakdown voltage and the heat resilience performance of the amorphous GaxOy film, thereby achieving considerable potential within the realm of power electronics applications. The proposed EMGen, as a specialized, interpretable AI model for the generation of electronic materials, is demonstrated to be an essential tool for on-demand semiconductor materials design.

Keywords: active learning; band gap; electronic materials; ensemble learning; first-principles calculations; machine learning.