Predição de propriedades mecânicas do biovidro através de dinâmica molecular e de inteligência artificial

Abstract

Recently, research into glass phases with bioactive properties has increased, and in most applications, the material must have minimal mechanical performance. Advances in molecular simulation methods and hardware have enabled the creation of predictive models of the structure-property relationship that faithfully reproduce experimental tests. These molecular simulation models will allow experimental data to be obtained in critical areas of the compositional domain that are not available in the literature and are necessary for training the neural network model. In this study, we applied Molecular Dynamics (MD) simulations to obtain the PMMCS structures and force fields that have compatibility in the simulation of bioglass. The concentrations of the oxides vary, P2O5 from 4% to 7%, SiO2 from 30% to 70%, and CaO and Na2O from 0% to 66%, with a 4% increase in mass. The structure validation parameters were evaluated based on the RDF of the structures for the Si-O length, which ranged from 1.6 to 1.611 Angstrom, values similar to those of the experimental structures. The DM and literature of the 45S5 bioglass show a small deviation, being 57.2 Gpa and 54.8 Gpa respectively, which shows that the force field used is safe. The variation in sodium oxide has a negative impact on Young's modulus, while silicon and calcium oxides show an inversely proportional variation, and this variation is directly related. However, phosphorus oxide does not have a clear evaluation due to its small variation. Thus, in the use of predictive algorithms using models, in this case polynomial regression, XGBoost, random forest and artificial neural networks, the nested cross- validation technique was implemented, in which the training hyperparameters needed to be adjusted, making use of a series of test/validation and training splits. The use of machine learning proved to be more effective than the Makishima-Makenzie predictive method, which, compared to experimental values, has a coefficient of determination of around 0.751, while machine learning has a coefficient of over 0.92.