ESAFORM Benchmark 2025 : predicting stainless steel PBF-LB part density using statistical, data-driven, and physics-informed machine learning models derived from process parameters and in-situ monitoring data

Tools

Monu, Medad Chiedozie C. and McCarthy, Eanna and Madathil, Abhilash Puthanveettil and Chekotu, Josiah C. and Ilic, Irina and Doğu, Merve Nur and Hughes, Cian and Juan, Rongfei and Amini, Ehsan and Lian, Junhe and Vassiades, Constantinos and Bylya, Olga and Darosa, Kim and Kromer, Robin and Seidou, Abdul Herrim and Mohanty, Sankhya and Habraken, Anne Marie and Mertens, Anne and Laitinen, Otto and Tucker, Michael R. and Brabazon, Dermot (2026) ESAFORM Benchmark 2025 : predicting stainless steel PBF-LB part density using statistical, data-driven, and physics-informed machine learning models derived from process parameters and in-situ monitoring data. International Journal of Material Forming, 19 (2). 31. ISSN 1960-6214 (https://doi.org/10.1007/s12289-026-01995-y)

[thumbnail of Monu-etal-2026-ESAFORM-Benchmark-2025]
Preview
 Text. Filename: Monu-etal-2026-ESAFORM-Benchmark-2025.pdf
Final Published Version
License: Creative Commons Attribution 4.0 logo
Download (13MB)| Preview

Abstract

This study benchmarks multiple data-driven methodologies for predicting relative density (RD) of 316 L stainless steel fabricated via Powder Bed Fusion–Laser Beam (PBF-LB), as part of the ESAFORM Benchmark 2025 AMDmodel initiative. Two datasets (DS-01 and DS-02), each with 256 specimens from a 4-factor, 4-level design of experiments, were produced on different PBF-LB systems equipped with equivalent in-situ infrared (IR) melt-pool pyrometry. Failed builds (RD = 60%) were retained to allow models to learn from both nominal and catastrophic processing conditions, a scenario rarely addressed in PBF-LB machine learning (ML). Statistical analysis of variance (ANOVA) confirmed that conventional process parameters alone are weak predictors (R² ≈ 0.49). In contrast, sensor-driven supervised ML models using melt-pool thermal descriptors performed substantially better. Recursive feature elimination highlighted the interquartile range and mode of thermal signatures as dominant predictors; an XGBoost model using only these achieved R² = 0.93 on DS-01. Hybrid models combining parameters and IR descriptors performed slightly worse (R² = 0.92), indicating mild redundancy. Cross-system transferability was limited: ML models trained on DS-01 underperformed on DS-02 due to IR input-domain divergence despite RD distributions between both domain sources showing high inter-laboratory consistency. To address this, a physics-informed ML framework (PIML) using symbolic regression (QLattice) embedded dimensionless physical priors. Resulting compact expressions dominated by normalized laser power and volumetric energy density achieved R² = 0.83–0.93 under cross-system validation. Overall, sensor-driven ML models are effective for machine-specific monitoring and layer-wise closed-loop control, whereas PIML provide system-agnostic process parameter-window estimation for design-stage optimization.

ORCID iDs

Monu, Medad Chiedozie C., McCarthy, Eanna, Madathil, Abhilash Puthanveettil ORCID logoORCID: https://orcid.org/0000-0001-5655-6196, Chekotu, Josiah C., Ilic, Irina, Doğu, Merve Nur, Hughes, Cian, Juan, Rongfei, Amini, Ehsan, Lian, Junhe, Vassiades, Constantinos, Bylya, Olga ORCID logoORCID: https://orcid.org/0000-0003-1229-1741, Darosa, Kim, Kromer, Robin, Seidou, Abdul Herrim, Mohanty, Sankhya, Habraken, Anne Marie, Mertens, Anne, Laitinen, Otto, Tucker, Michael R. and Brabazon, Dermot;
CORE (COnnecting REpositories)