Optimizing Hyperparameters of a Multi-Scale Convolutional Neural Model Tailored to Describe Amorphous Materials Behavior

Abstract

This paper presents an optimization procedure of parameters in a multi-scale neural model, which is crucial for accurately characterizing the behavior of amorphous materials; more specifically glass transition kinetics, which is considered one of the most important yet not fully understood phenomena of solid-state physics and chemistry, with wide-ranging applications. Through systematic exploration of a hyperparameter grid space and rigorous evaluation of resultant models, a highly effective configuration was identified – Conv1D kernel size=8-24-48, Conv1D #filters=16 & Dense #neurons=16 – which offers optimal performance with remarkably short mean epoch times. Our findings suggest a promising strategy of increasing kernel size while decreasing the number of filters and neurons in the dense layer, supported by the superior performance of key competitors sharing this trend. However, further examination reveals comparable performance levels among subsequent competitors, indicating the need for additional samples to draw definitive conclusions. Our experiment highlights the intricate relationship between model accuracy and computational resources, emphasizing the necessity for further results to gain a comprehensive understanding of hyperparameter impact. Three promising combinations for future experimentation emerge, with Conv1D kernel size=8-24-48, Conv1D #filters=16 & Dense #neurons=16 standing out as the clear winner for practical application.