Physics-Informed Multi-Fidelity Graph Learning for Sequence-Aware Residual Bolt Preload Prediction in Hyperelastic-Sealed Flanges

Abstract

Highlights What are the main findings? A physics-informed graph-learning workflow was developed for sequence-conditioned residual preload fields using EICM pre-training, nonlinear-discrepancy transfer, and ultrasonic measurement-based uncertainty quantification. On the augmented development split, TransformerConv-L3 achieved an RMSE of 170.9 N and 93.9% coverage for the nominal 95% prediction interval, outperforming the recurrent baselines evaluated under the same protocol. What are the implications of the main findings? The full causal graph provides a compact computational representation of sequence-dependent bolt interactions and can be evaluated in milliseconds. The workflow integrates mechanics-informed simulation and ultrasonic measurements for sequence-aware preload prediction and uncertainty quantification. Abstract Sequential tightening redistributes load among bolts and produces a residual preload field that depends on the complete operation history. This study presents a physics-informed graph-learning framework for a 60-bolt circular flange. The model receives initial preloads and tightening order and predicts the residual preload field on a complete causal directed graph. All 1770 admissible later-to-earlier edges are retained. Among them, 240 local edges carry non-zero Elastic Interaction Coefficient Method (EICM) attributes, while 1530 distant edges remain available to learned message passing. The multi-fidelity dataset combines online EICM responses, prescribed nonlinear-discrepancy labels, and repeated ultrasonic preload measurements. On the augmented development split, TransformerConv-L3 achieved an MAE of 138.3 N, an RMSE of 170.9 N, an NLL of 8.839, and 93.9% coverage for the nominal 95% prediction interval. Its RMSE was 56.9% lower than that of the strongest recurrent baseline evaluated under the same protocol. These results demonstrate the value of combining causal topology with mechanics-informed edge attributes for residual preload prediction.