| Author(s) |
Yuxiang Luo, Akshar Kudva, Eun Jang, Salman Matan, Nikolas M. Vega Michalak, Andrew Perrault, Joel A. Paulson, Boian Alexandrov |
| Abstract Scope |
The accuracy of welding and additive manufacturing process simulations heavily depends on the heat source (HS) model, which dictates the spatial distribution of applied heat and influences both thermal and mechanical predictions. However, heat source calibration has been largely overlooked, with parameter selection traditionally relying on empirical heuristics. This under-calibration can lead to poor simulation accuracy. This paper presents a surrogate gradient-based machine learning framework for efficiently calibrating HS parameters. The proposed approach employs a two-stage hybrid optimization strategy: first, Bayesian optimization is used to globally explore the parameter space and potentially optimal regions; second, a gradient-based optimization method locally fine-tunes the parameter values from the best-found parameters. This hybrid approach enables reliable and sample-efficient calibration based on experimental fusion zone data. Furthermore, this work introduces a set of enhanced HS models that offer greater expressiveness and practical applicability than the conventional double-ellipsoidal Gaussian model. The proposed methodology is validated through case studies involving three processes, gas tungsten arc welding, spray transfer gas metal arc welding, and pulsed gas metal arc additive manufacturing—demonstrating improved agreement between simulation predicted and experimental HAZ results.
The validation results show that (1) machine-learning optimization-based HS calibration significantly enhances simulation accuracy compared to using empirical HS models, (2) expressive HS models, that can represent complex energy distribution, better align with experimental data compared to the double-ellipsoidal Gaussian model, and (3) the proposed approach is sample-efficient and adaptable to different welding and additive manufacturing processes. Experimental validation, including comparisons of thermal histories, microstructure, and mechanical properties in the heat-affected zone, confirms strong alignment between simulation and real-world observations. This work provides a systematic framework for improving welding simulation accuracy by addressing the limitations of empirical HS models and facilitating the application of more expressive and precise modeling approaches. |