Exploring the mechanical behavior of friction material composites using artificial intelligence.

The application of artificial intelligence techniques is becoming increasingly valuable for several areas of knowledge, allowing to extract information, predict patterns, work with complex problems, and generate solutions that would not be achievable by other techniques without an extremely high computational cost or else, numerous physical experiments, with high execution costs and which often do not achieve the desired result. In this work, artificial intelligence (AI) was used to create mathematical models for four properties of friction material composites from an extensive database containing the chemical composition and the respective mechanical properties. An algorithm capable of predicting mechanical results of friction materials based on chemical composition, optimizing an existing composition, and proposing new (previously non-existent) compositions was proposed based on the desired values of each mechanical property. The algorithm combines rule-based instructions, neural networks, and particle swarm optimization. Physical samples based on the algorithm's prediction were produced, making it possible to assess the predictive power of the models and further understand the need for improvements in the tool built for predicting new friction materials. The root mean square error (RMSE) of predictions from multilinear models was significantly higher than those from artificial neural networks when predicting results of non-existent previous compositions. The smallest observed increase in RMSE was 20.9 %, while the largest was 90.2 %. On average, the RMSE for multilinear models was 48.6 % larger, highlighting the superior accuracy of neural network predictions from previously non-existent compositions.