Modeling of tensile strength and wear resistance in friction stir processed MMCs by metaheuristic optimization and supervised learning

This study presents a systematic framework for modeling the tensile strength (TS) and wear resistance (WR) of friction stir processed (FSP) AA1100 metal matrix composites (MMCs) reinforced with aluminum oxide (Al₂O₃) powder reinforcement. Using a design of experiments approach, the influence of tool rotation speed, feed rate, and reinforcement content on TS and WR was investigated across 20 independent experiments, with the optimal process combination resulting in a 20.7% increase in TS and a minimum WR of 7.3 × 10⁻³ mm³/m. To capture the nonlinear relationships between process parameters and material performance, TS and WR were initially modeled using second-order polynomial regression, further optimized by two metaheuristic (ME) algorithms. Furthermore, two supervised machine learning (ML) models were developed, optimized, and benchmarked. ME and ML models were validated using both the experimental dataset and ten synthetic cases generated by Gaussian mixture models (GMM) within the experimental features’ latent space. For TS modeling, ME-optimized regressions showed the mean percentage deviations in 15.6% ~ 27.7% range while the ML formulations in the 12.4% ~ 17.9% range. For WR, a similar reduction in error range and variance was also confirmed, highlighting a higher modeling reliability than the TS counterpart. Results demonstrate that while both ME and ML approaches can be employed for FSP process modeling within the latent space, data-driven approaches, and especially neural networks architecture, have a clear advantage in prediction accuracy and robustness under data-limited conditions. Although modeling with only 20 cases is challenging for both ML and ME methods, such data scarcity is typical in industrial settings; thus, this study reflects a realistic scenario where reliable predictions must be achieved from limited data. Overall, this work provides a practical modeling benchmark between ME-optimized and ML modeling for the FSP and may help in identifying the best modeling approach also for other manufacturing engineering processes.