Real-Time Prediction of Distortion & Residual Stress via Machine Learning

In the fabricated structures community, there is a relationship between how the structure is made and how it will perform in service. One of the inherent benefits of automation is the consistency of the fabrication process that robotics afford, such as consistent heat input and process control. Combined with physics-based simulation and thermomechanical analysis structures maybe effectively optimized. However, the consistency that is possible with the application of robotic welding is compromised as the components to be fabricated vary in their shape and how they fit together. This input variation often results in distortion and residual stress profiles that are not intended. Recent developments in machine learning and artificial intelligence enable the understanding of a presented assembly variation and update of the welding plan in near real-time relative to that part condition. These developments would enable the target processes to be optimized and executed as if the physics-based simulation and optimization were computationally feasible. This takes advantage of both new approaches for machine learning based frameworks, as well as the ability to execute at the rate of production, leading to improved operational efficiencies and optimized fabricated products. This talk seeks to share progress in current work that seeks to apply this hybrid physics-based simulation with a novel learning framework that seeks to understand presented articles, how they deviate from plan and update welding plans on the fly to optimize measured distortion and residual stress. The goal is to create a system that may be added to welding systems that enable dynamic planning as presentations and conditions change in the real world, thereby making the upfront simulation work more valuable in the production phase, benefitting structural designers, fabricators as well as additional stakeholders concerned with as-fabricated quality.