Selecting Laminate Configurations for Training a DeepONet to Predict Incremental Hole-Drilling Calibration Constants in Fiber-Reinforced Plastic Laminates

Abstract Measuring residual stress is critical for assessing the structural integrity and performance of fiber-reinforced plastic (FRP) laminates. The incremental hole-drilling (IHD) method, a common technique for this purpose, relies on calibration constants that are typically determined through computationally intensive finite element (FE) analyses for each unique laminate configuration. While deep operator networks (DeepONet) can serve as efficient surrogate models, the optimal strategy for composing the necessary training data has not been fully established. In this work, a simple yet effective strategy for training data selection is proposed, which is shown to reduce error in predicted stress by 12.4–15.2% over uniform random selection. Further analysis of training data length with the proposed selection strategy shows that acceptable test error, within the inherent uncertainty of the IHD method, can be achieved using just a small fraction of the possible laminate configurations. In this study, strategically selecting only 15 laminate configurations for training, out of the 70 possible laminate configurations, provided acceptable accuracy for complex residual stress profiles, including steep gradients. These findings provide a practical framework for developing reliable surrogate models, making computationally demanding residual stress analyses more accessible for the design and validation of composite structures.