A aplicação de deep learning na previsão de propriedades mecânicas de materiais — revisão sistemática e aprofundada

Abstract

The application of deep learning in materials science represents a new frontier for accelerating the discovery and optimization of emerging technologies. This study aims to conduct a systematic literature review to map the state of the art, methodologies, challenges, and future trends in the use of deep neural networks for predicting mechanical properties. The methodology involved the analysis of 73 scientific articles published between 2017 and 2025. The results indicate an exponential growth of the field since 2020, with a predominant focus on predicting properties such as tensile strength and fatigue. Convolutional Neural Networks (CNNs) dominate the landscape, driven by the use of microstructure images as input data, alongside hybrid approaches that combine experimental and synthetic datasets for database construction. Despite the high predictive accuracy of the models, the main challenges lie in their "black-box" nature, which limits interpretability, and in ensuring generalization to real-world scenarios. Future trends point to the consolidation of Explainable Artificial Intelligence (XAI) techniques and the integration of such models into Inverse Materials Design platforms, marking the transition from a purely predictive data-driven science to a prescriptive approach.