A THEORY-GUIDED MACHINE LEARNING METHOD FOR CURE CYCLE OPTIMIZATION TO MINIMIZE PROCESS-INDUCED DEFORMATIONS IN COMPOSITES

Abstract

One of the most prevalent challenges faced in aerospace manufacturing is the accurate prediction, control, and mitigation of residual stresses and resulting process-induced deformations (PIDs) in composites. Unwanted dimensional changes in composite parts can cause significant geometry mismatches during aerostructure assembly, leading to decreased throughput and a loss of structural performance. In recent decades, process simulation tools have been developed to facilitate predictions and mitigations of PIDs through process optimization as well as methods such as tool compensation. However, such numerical approaches often rely on time-consuming, expensive, and deterministic characterization and modelfitting techniques. As a result, simulation tools may be error-prone in industrial settings, making PID predictions unreliable. Therefore, manufacturers rely on laborintensive methods, such as shimming, to compensate for geometry mismatches during the assembly process. This paper presents an alternative methodology to predict and minimize PIDs in composites without using process simulation. The proposed method is solely based on a limited amount of element-level experimental tests and theory-guided machine learning (TGML). Experimental methods include autoclave-curing of L-shaped laminates with different cure cycles and quantifying the resulting PIDs using laser profilometry. Probabilistic ML models are then built to correlate temperature cycles directly to PIDs using Gaussian Process Regression (GPR) guided by the closed-form theory available for PID predictions. Finally, the theory-guided ML models are iteratively calibrated using additional targeted tests specified by the GPR algorithm to efficiently converge on an optimal cycle that minimizes PIDs and satisfies cost specifications. The proposed ML-based approach is a generalizable probabilistic method to optimize process parameters and mitigate PIDs without using expensive process simulations or conducting exhaustive characterization experiments.