Machine Learning Applicationfor Active Exploration of Weld Sequence Scenarios

Abstract

Distortion is a common problem in welded structures, and the process of finding an effective weld sequence to mitigate the distortion is a challenging task given a large number of possible combinations. Despite efficient simulation tools and powerful supercomputers, yet simulation tools have been limited by CPU time to optimize a welding sequence and therefore not mature for practical designs. To this end, we constructed and integrated machine learning (ML) algorithms with the simulation capability. These ML models were then trained to increase the fidelity by a wisely chosen training-set of simulation to construct a meta-model for active exploration of various weld sequence scenarios in real time. As opposed to existing ML algorithms that require an extensive data set to train, our algorithm picks relatively small training-set to construct a metamodel. We present an example of our algorithm implemented in a real welded structure project. Introduction Today’s structures are more complex and demanding much tighter fabrication tolerances than our routine practice. Welding plays a challenging role in meeting those tolerances in particular when it comes to distortion. Welding sequence and intermittent welding design, which determines the best welding pattern in multi-pass welds, are familiar techniques to control the distortion when dealing with multi-pass welded structures. Finding the best solution for such a design is usually intuitive and based on the similarity of previously welded structures because this is not feasible through shop trials. An alternative is welding simulation tools that model several sequences to select the one with minimal distortion. Excellent simulation software is now available to capture and couple thermal, microstructure and stress effects of welds based on 3D transient temperature and thermal stress-strain analysis [1]. Despite powerful supercomputers, yet welding simulation tools are limited by computational time, and therefore they are not mature for practical designs. For example, having “n” welds requires choosing from 2^n n! possible scenarios or combinations of the welds where n! counts for permutations and 2^n counts for change in the direction of welding, i.e., several million for typical weld consisting of 10 weld passes or more. More affordable approaches have been developed to generate a sufficient and reliable level of understanding of the behavior of structures in order to find an optimal sequence with a few numbers of simulation. One approach is to use a fast but low-fidelity model that captures the most dominant physics of the problem, for example, depositing each weld pass at once [2]. Although 1 Corresponding author and presenter; Mahyar.Asadi@applusrtd.com 1002 Contributed Papers from Materials Science and Technology 2019 (MS&T19) September 29–October 3, 2019, Oregon Convention Center, Portland, Oregon, USA DOI 10.7449/2019/MST_2019_1002_1009 Copyright © 2019 MS&T19®