An intelligent design framework for multi-principal element alloys using optimized deep learning

Abstract

Multi-principal element alloys (MPEAs) are a new type of alloys which have shown promising applications but they lack a comprehensive theoretical framework due to insufficient experimental analysis. To circumvent these issues, machine learn-ing has proven to be effective in assisting design of MPEAs by generating a prediction mechanism of properties of alloys. The core object of this research is to develop a prediction framework for MPEAs via dense neural networks which uses atomic composition to predict alloy phase and its physical properties, making this a supervised learning approach. This framework will then be supplemented with hyperparmeter optimization to improve its prediction power and allow us to get a better theoretical understanding of interactions be-tween atoms in an MPEA. For the purposes of optimization, genetic algorithm (GA) is utilized with novel hyperparameters which serve to include data pre-processing within the framework of hyperparameter optimization. Ablation studies were con-ducted which indicate improvement in prediction by including novel hyperparameters, thus justifying our approach. This model is then tested on recently discovered MPEAs and is able to predict their alloy phases with high accuracy, making this a viable approach for experimental design. Building on our prediction model, we further explored hyperparameter optimization by using teachnig-learning-based optimization algorithm (TLBO), which is a population based heuristic algorithm. TLBO has less control parameters than GA making it more feasible, but it suffers from premature convergence. To this affect, we introduce modifications in conventional TLBO to create balance between local and global search. Each approach is tested and compared with conventional TLBO, and the results depict notable improvement. Combinations of the three modifications are compared with each other and the best combination is termed as enhanced TLBO (eTLBO), which when compared with GA illustrates improvement in prediction accuracy. The outcomes of this reseach provide a deeper understanding of complex mechanisms surounding MPEAs while signifying the importance of optimization in supervised learning models. This thesis leverages the findings to discuss implications regarding high-throughput alloy design, and determine feasible machine learning systems to support it.

The outcomes of this reseach provide a deeper understanding of complex mech-anisms surounding MPEAs while signifying the importance of optimization in su-pervised learning models. This thesis leverages the findings to discuss implications regarding high-throughput alloy design, and determine feasible machine learning systems to support it.