University of British Columbia
A deep learning-assisted inherent strain method for predicting welding distortions in large stiffened panels
Abstract
dc:descriptionWelding-induced distortion in stiffened marine structures affects fabrication accuracy, assembly fitup, and structural performance. High-fidelity thermal elastic-plastic analysis can resolve the local welding response, but direct application to large stiffened panels remains computationally expensive. This thesis develops a deep-learning-assisted inherent strain workflow for efficient weldingdistortion prediction in large welded panels. Local unit-cell T-joint simulations of double-sided gas metal arc welding are used to generate residual plastic-strain fields, and five neural network surrogate architectures are evaluated to predict these fields from T-joint geometry. The predicted strain fields are converted into equivalent inherent strain parameters and imposed in global elastic shell models through an equivalent thermal strain formulation. This local-to-global strategy preserves the mechanics-based interpretation of inherent strain modelling while avoiding full transient thermal-mechanical simulation of each large panel. Among the evaluated architectures, the transformer triple-head model achieves the lowest overall prediction error, while continuity tests across plate- and web-thickness variations produce smooth and bounded error trends. Against thermal elastic-plastic reference solutions, distortion predictions based on surrogate-generated strain fields achieve peak-normalized path errors of 3.19%, 3.90%, and 4.61% for the unit-cell, 2-by-2, and 3-by-3 stiffened-panel cases, respectively. The corresponding elastic shell analyses achieve wall-clock speedups by a factor of over 1,000, demonstrating that field-level deep learning surrogates can extend local welding simulations to efficient distortion analysis of larger stiffened-panel assemblies.
Degree
thesis:*- Name thesis:degree_name
- Master of Applied Science - MASc
- Level thesis:degree_level
- master's
- Discipline thesis:degree_discipline
- Mechanical Engineering
- Grantor dc:publisher
- University of British Columbia
- Year dc:date
- 2026
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Shi, Runze
- Contributors dc:contributor
-
- Jelovica, Jasmin
Rights
dc:rights- Statement dc:rights
-
- Attribution-NonCommercial-NoDerivatives 4.0 International
- Language dc:language
- eng
Identifiers
dc:identifier.*- Handle dc:identifier
- http://hdl.handle.net/2429/95537