Predição da energia de ligação do pósitron com moléculas: uma análise exploratória utilizando algoritmos de machine learning

Abstract

The goal of this dissertation is to employ machine learning techniques to investigate the importance of specific molecular properties in determining the binding energy of the positron with molecules. Experimental data on positron binding energies (PBEs) and molecular properties such as polarizability, dipole moment, and the number of π bonds were used. Five machine learning algorithms were explored: support vector machines for regression (SVR), polynomial regression with ridge regularization, XGBoost, multilayer perceptron, and an ensemble model. The data were divided into three groups: nonpolar molecules, polar molecules, and a mix of both types. Different sets of molecular properties were tested for each group. The evaluation and tuning of the algorithms were performed using nested cross-validation and Bayesian optimization, respectively, with the benchmark metric being the mean absolute percentage error. Additionally, the SHAP approach was used for interpretability of the models used. The results suggest that anisotropic polarizability is an important descriptor for polar molecules, as well as π bonds for nonpolar molecules, reflecting the influence of structure and spatial distribution of the electron cloud in calculating positron binding energy. For polar molecules, it was also observed that the performance of algorithms improves when ionization potential is used as an attribute. This study demonstrates that machine learning techniques can provide valuable insights about the physical mechanisms involved in positron binding with molecules, complementing theoretical and experimental approaches. However, limitations such as the scarcity of experimental data need to be overcome in future research. Future work should focus on acquiring more data, exploring other models specific to few samples, and advanced model interpretability techniques.