Abstract

Designing the new semiconductor technologies calls for advanced materials having superior electrical conductivity, heat stability, carrier mobility, dielectric properties, and robust structures. Traditional techniques for material synthesis and parameter optimization can be requiring numerous tests, computational efforts, and parameter tuning. State-of-the-art methods of computational materials design usually experience difficulties in handling the non-linear relationship between material properties and the performance of the final device. A Deep Learning-Assisted Semiconductor Material Intelligence Network (DLSMIN) that uses multimodal material representation, physics-informed attention-guided learning, physics-informed learning approach, and multi-objective optimization framework is presented. The method combines information on chemical composition, crystal structure, electronic properties, and fabrication process of material to predict its performance. Physics-informed regularization technique is used to encode knowledge about physical processes, and uncertainty-aware optimization is applied to select the most suitable material candidates for experiments. In the comparison study, the new material intelligence framework demonstrates an R2 of 0.92, mean absolute error of 0.08 eV, and root mean square error of 0.15 eV for predicting semiconductor’s bandgap with a 70% portion of training data. Additionally, the model attains the material suitability classification accuracy of 94.2% and screening efficiency of 50%.

Authors

Mohammad Abdur Rasheed
Shaqra University, Saudi Arabia

Keywords

Deep Learning, Semiconductor Materials, Materials Discovery, Physics-Informed Learning, Multi-Objective Optimization