Physics-based features and pretrained tabular models improve melt pool prediction in metal 3D printing
An arXiv study reports that dimensionless features combining laser conditions with material properties, paired with pretrained tabular models, improve melt pool prediction across alloys.
In plain words
In laser powder bed fusion, a metal 3D printing method, the size and type of the melted spot affect whether parts come out defect-free. The study tested whether describing the process with physics-based numbers, rather than raw machine settings, helps machine learning models predict the melted spot from less data. The authors report that this helped most when models trained on one alloy were applied to another.
What was announced
The authors, writing on arXiv, note that machine learning for advanced manufacturing is held back by small experimental datasets and by process-property relationships that shift across materials and operating conditions. Conventional models use dimensional process parameters, which capture laser-material interactions only implicitly and may limit transfer across materials. The study tested two dimensionless features, normalized enthalpy and the Peclet number, which combine laser processing conditions with material thermophysical properties, against raw process parameters for predicting melt pool width, depth, and melting mode in laser powder bed fusion.
The team benchmarked seven models: four classical methods (Gaussian process, support vector machine, random forest, and XGBoost) and three pretrained tabular foundation models (TabPFN, TabDPT, and TabICLv2). Experiments covered 316L stainless steel, IN625, IN718, and combined-material datasets, with regression for width and depth and classification for melting modes. Dimensionless features gave comparable or better performance in most settings, with the largest gain in cross-material depth prediction at R² = 0.96. The foundation models generally classified better than classical baselines, especially on multi-material data. The authors say the combination could support more transferable process-window identification.
What it does not show
- The evidence is the paper's abstract only; full methods, data sources, and validation details are not available in the supplied passages.
- Reported results, including the R² = 0.96 figure, come from the authors' own benchmarks and have not been independently confirmed.
- The passages do not state whether code, trained models, or the underlying datasets are publicly available.
For security and IT teams
- Scope
- The reported experiments cover 316L stainless steel, IN625, IN718, and combined-material datasets for LPBF melt pool width, depth, and melting mode prediction; findings are limited to these materials and tasks.
Sources
Source-support review completed 20 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 20 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.