| Abstract Scope |
This work is focused on robotic GMAW in industry with multi process parameters to multi welding metrics. 25 robotic GMAW process parameters were co-optimized against 17 competing objectives and constraints, reaching a deployable procedure in ~750 forward evaluations. Production welding is never single-objective; competing quality, productivity, and integrity targets must be balanced. Single-objective optimization improves one metric while degrading the rest. This produces procedures that can't be efficiently deployed in production. Instead of a single-objective approach, we applied a customized multi-objective multi-agent system (MOMAS), an inverse solver that resolves all 17 objectives simultaneously.
The 25 decision variables include welding current, arc voltage, travel speed, wire feed rate, shielding-gas selection, and number of passes; and the 17 metrics span penetration depth, heat-affected-zone width, tensile strength, throughput, energy per part, angular distortion, and peak residual stress. The forward model is a physics-based digital twin grounded in six published welding-science correlations (Gunaraj-Murugan, Easterling, Lancaster, AWS D1.1, Okerblom, Masubuchi). The solver learns this model once, stores a reusable inverse representation, then inverse-solves for new objectives and constraints.
Joint optimization yields a Pareto front, including three regimes along a count-versus-magnitude trade-off frontier. The deepest-margin regime improves 10 of 17 with the largest gains: heat-affected-zone width falls 70%, throughput rises 22%, and predicted angular distortion and peak residual stress fall 96% and 58%. Two higher-coverage regimes each improve 11 of 17 with smaller margins; only the balanced one also lowers defect rate, suiting structural welds under radiographic inspection. Six single-objective baselines improve just 6.8 of 17 on average, with the worst worsening 13 of 17.
The model reproduced 20 of 23 published reference points within accepted ranges (87%). Simultaneous improvement saturates at 11 of 17; the improved set is identical at 750 and 20,000 evaluations (~25× more), indicating a physics-constrained frontier. |