A physics informed machine learning framework towards density prediction of additively manufactured components

Tools

Puthanveettil Madathil, Abhilash and Bylya, Olga and Sherlock, Andrew (2026) A physics informed machine learning framework towards density prediction of additively manufactured components. Key Engineering Materials, 1047. pp. 73-82. ISSN 1013-9826 (https://doi.org/10.4028/p-vvwf5p)

[thumbnail of Madathil-etal-2026-A-physics-informed-machine-learning-framework-towards-density-prediction]
Preview
 Text. Filename: Madathil-etal-2026-A-physics-informed-machine-learning-framework-towards-density-prediction.pdf
Final Published Version
License: Creative Commons Attribution 4.0 logo
Download (892kB)| Preview

Abstract

Laser powder bed fusion (LPBF) has become a key manufacturing route for high-value components, yet accurate prediction of part quality, particularly relative density, remains challenging due to complex interactions among process parameters, material properties, melt-pool physics and shielding environment. Traditional physics-based simulations offer mechanistic insight but suffer from model-form uncertainty and computational cost while purely data-driven machine learning (ML) models often lack interpretability, physical consistency and transferability across build conditions. To address these gaps, we propose a physics-informed machine learning (PIML) framework that integrates structured domain knowledge with symbolic regression to derive compact, interpretable analytical expressions for LPBF density. The framework constructs a knowledge base (KB) comprising dimensionless and normalized physics-informed process descriptors (PIPDs) that encode energy input, thermal diffusion and melt-track geometry; these descriptors form a physically consistent feature space for learning. The framework also serves as a foundation for the proposed future community-driven, knowledge-graph-based modelling of LPBF processes. The capability of the framework is demonstrated by modelling the relative density of additively manufactured maraging steel and evaluating cross-atmosphere transferability from Argon-shielded to Nitrogen-shielded builds. The resulting symbolic models provide a transparent, extensible and physically meaningful alternative to black-box ML, achieving high predictive performance under the training regime (R = 0.964) and strong generalization across printing atmospheres (R = 0.896).

ORCID iDs

Puthanveettil Madathil, Abhilash ORCID logoORCID: https://orcid.org/0000-0001-5655-6196, Bylya, Olga ORCID logoORCID: https://orcid.org/0000-0003-1229-1741 and Sherlock, Andrew;
CORE (COnnecting REpositories)