A Theory-Guided Probabilistic Machine Learning Method to Minimize Process-Induced Deformations in Composite Structures

Abstract

Abstract Although manufacturing technologies have significantly evolved in recent decades to enable composite-intensive aircraft designs, many production challenges remain outstanding. One such challenge is accurately predicting, controlling, and mitigating residual stresses and process-induced deformations (PIDs) in composite structures. Several physics-based tools have been developed with the intent of enabling predictions and mitigations of PIDs. However, predictions made by numerical simulations are often unreliable or cost-deficient since they rely on time-consuming, expensive, and deterministic characterization and model-fitting efforts. This paper presents a simulation-free methodology to predict and find the optimal stacking sequence for minimizing PIDs. The proposed method is comprised of limited element-level experiments directed by a theory-guided machine learning (TGML) framework. Throughout the optimization procedure, a Gaussian process regression (GPR) model is built based on underlying theory and iteratively calibrated to converge on a solution efficiently. In this study, the TGML method is used to find an optimal cross-ply stacking sequence (i.e., layup) that minimizes PIDs for L-shaped composite parts made from T800S/3900-2B. The method proposed in this paper is meant to provide a generalizable optimization procedure to potentially mitigate PIDs in composite structures.