Comparison between single- and multitask models induced by machine learning for the prediction of glass properties

• Grant number: 24/06604-0
• Support Opportunities: Scholarships in Brazil - Scientific Initiation
• Start date: October 01, 2024
• End date: September 30, 2025
• Field of knowledge: Engineering - Materials and Metallurgical Engineering - Nonmetallic Materials
• Principal Investigator: Daniel Roberto Cassar
• Grantee: Gustavo Uchôa Barros
• Host Institution: Centro Nacional de Pesquisa em Energia e Materiais (CNPEM). Ministério da Ciência, Tecnologia e Inovação (Brasil). Campinas , SP, Brazil

Abstract

Predictive models of material properties can be used for inverse design, that is, the search for materials that satisfy certain desired design constraints. Therefore, predictive models with better performance are desired, since the better the performance, the more efficient the inverse design search will be. Recently, a new predictive model capable of predicting 85 properties of glass materials has been reported. This model was named GlassNet and is a multitasking model. One question raised in the GlassNet article is whether multitasking models perform better than single-tasking models; however, it has not been possible to answer this question with confidence. The goal of this project is to investigate this question. To do so, we will train several single-task models of glass properties by going through the steps of feature engineering and hyperparameter optimization (the latter was not done in the article reporting on GlassNet). The results of this work are of particular interest to the glass materials community, but could also be used by the materials community in general as the strategies presented are agnostic with respect to the nature of the compositions studied.