About this Abstract |
| Meeting |
MS&T26: Materials Science & Technology
|
| Symposium
|
Additive Manufacturing Modeling, Simulation, and Machine Learning: Microstructure, Mechanics, and Process
|
| Presentation Title |
Toward Transferable Dimensionless Process Maps for Active Learning-Based Optimization in Wire-Arc Additive Manufacturing |
| Author(s) |
Pallock Halder, Tiana Tonge, Wyatt Ballweber, Satyajit Mojumder |
| On-Site Speaker (Planned) |
Satyajit Mojumder |
| Abstract Scope |
Wire Arc Additive Manufacturing (WAAM) holds significant promise for large-scale component production across aerospace, energy, and defense sectors. Nevertheless, process parameter optimization remains a critical bottleneck, as optimized conditions rarely transfer across materials or machines without costly recalibration. This work presents an active learning framework employing dimensionless process maps to reduce experimental burden in WAAM.
Acoustic signals captured during deposition are classified via a 2D Convolutional Neural Network to provide automated bead quality labeling, eliminating reliance on manual inspection. These quality scores train a Gaussian Process Regressor, with an Expected Improvement acquisition function guiding subsequent experiments. Validated on mild steel (ER70-S) within 40 experiments, the resulting dimensionless map served as a transferable prior for aluminum (Al4043), requiring only minimal additional trials to achieve comparable bead quality and geometry. These findings demonstrate that dimensionless process maps substantially reduce cross-material optimization effort in WAAM. |