Explorando descritores moleculares para previsão de bandgaps de materiais cristalinos 2D

Abstract

Among the materials studied in materials science, two-dimensional materials have particularly relevant applications. With the advancement of computational techniques, many of these materials have been theoretically predicted even before their experimental synthesis. Due to the numerous potential applications of two-dimensional materials, it is important to understand electronic properties such as the bandgap. This work proposes investigating ways to predict bandgap values using geometric parameters of materials; for this purpose, molecular descriptors were employed to represent these parameters. For the prediction task, a neural network based on a Multi-Layer Perceptron (MLP) algorithm was used. The descriptors were selected according to the characteristics of the problem. In this work, the chosen descriptors were the Coulomb Matrix, SOAP, and MBTR. Among the studied descriptors, the best performance was achieved using the MBTR descriptor. The Coulomb Matrix descriptor was used as a baseline for comparison with descriptors that are more suitable for the proposed task. The best obtained R2 value was 0.7492, indicating that the model has a good ability to capture the general behavior of the data. The present work shows that there are limitations in training neural networks for bandgap prediction, arising from the size of the database, the distribution of values between the minimum and maximum limits, limitations in representing crystal structures through molecular descriptors, and the techniques employed during network training, among other factors. However, it also demonstrates that this approach has a promising future for bandgap prediction, since databases are continuously expanding, there are ways to overcome descriptor limitations, and more precise neural network tuning can considerably improve prediction quality.