Relative-density prediction in multi-material laser powder bed fusion based on a physics-informed continual-learning framework
Year: 2026
Type: article
Abstract: Laser powder bed fusion (LPBF) enables the fabrication of high-performance metallic components across stainless steels, nickel-based superalloys, titanium, aluminum, and copper alloys. Accurate relative-density prediction is critical for LPBF process optimization, supporting defect control, suppression of lack-of-fusion, avoidance of keyhole porosity, and qualification of process windows. However, most machine-learning models for LPBF density prediction are static: they assume all alloy data are available during initial training and require retraining when new material domains are introduced, leaving them vulnerable to catastrophic forgetting. To address this limitation, we develop a physics-informed continual-learning framework for multi-material LPBF relative-density prediction. A dataset spanning six alloy domains (316 L, IN718, Ti-6Al-4 V, 18Ni300, AlSi10Mg, and CuCrZr) was constructed, with a 42-variable feature contract integrating process parameters, energy-density descriptors, geometric factors, alloy composition, and thermophysical constants. A residual-corrected initial model (M⁎) achieved CPA @5% = 98.56 ± 0.61% on 316 L. Building on this foundation, a physics-informed elastic-weight consolidation strategy (PI-EWC) weighted historical samples by thermophysical distance. PI-EWC improved final accuracy from 75.90 ± 0.56% to 94.84 ± 0.26% and reduced backward transfer (forgetting rate) from −21.68 ± 0.59% to −0.51 ± 0.60%, remained stable across nine task orders, and was comparable to a static model jointly trained on all alloys (95.00 ± 0.77%) without requiring simultaneous access to all alloy data.
Source: Advanced Light Materials
Language: English
Subfield: Mechanical Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7223400175
