Development and Implementation of Data Augmentation Methods for Predictive and Generative Models in Chemistry and Materials Science

Abstract

The use of machine learning (ML) in the study and discovery of new materials has emerged as a fundamental approach, offering valuable alternatives and complements to traditional experimental and computational methodologies. ML's ability to analyze vast amounts of data and identify complex patterns enables the efficient prediction of properties and the generation of new materials, significantly accelerating the discovery process. While experimental approaches are often limited by time and cost, and traditional computational techniques, such as Density Functional Theory (DFT) calculations, are computationally expensive, machine learning methods offer faster and more scalable solutions. However, the effectiveness of these models heavily depends on the availability of high-quality data that adequately covers the problem space. In this context, data augmentation techniques become essential, allowing for the expansion and enrichment of training datasets, thereby improving the accuracy and reliability of predictive and generative models in materials chemistry.Despite their importance, compared to other areas of machine learning application, such as computer vision, there is a significant lack of studies integrating advanced data augmentation methods in the context of chemistry and materials science. Therefore, this research project aims to fill this gap by evaluating existing methods and proposing new data augmentation methods for this specific scenario. In particular, these methods will aim to enhance the training of predictive and generative models in materials chemistry, increasing the quantity and diversity of training data, thereby improving the accuracy and applicability of the trained models. In addition to the computational study of the methods, the project intends to develop a customized toolbox for chemical data augmentation. It is noteworthy that this project is part of the Center for Innovation in New Energies (CINE), funded by FAPESP and SHELL.