Aplicação de técnicas elementares de machine learning à Física

Abstract

Recent advances in deep learning have ignited, once again, the hype related to this areaand, consequently, a large number of investiments have been made in the past few years.This is justified by the versatility of this method, which has comercial and academic appli-cations. This study is aimed to evaluate the performance of deep leaning techniques whenapplied to two physical systems: the first one being a temperature-forecasting problemand the second one being the computation of the ground states of a particle subjectedto a potential function. To this purpose, the Keras API was chosen, as this plataformsimplifies the implementation of neural networks. For the first system, the architectureselected consists of a combination of recurrent layers. The model started overfitting earlyand presented high validation loss during all the procces. In the second problem, thearchitecture selected consisted in a stack of Dense layers. For each potential, the trueground states were obtained and compared to the network prediction. Although neuralnetworks have the ability to identify complex patterns and create powerful models, re-sults showed that this network was not efficient in fitting the data for examples where theground state function presented a high level of heterogeneity.