Integrating Multi-Physics Simulations and Machine Learning to Define the Spatter Mechanism and Process Window in Laser Powder Bed Fusion

TL;DR

The paper tackles spatter formation in laser powder bed fusion by developing a high-fidelity multiphysics OpenFOAM model coupled with a spatter-tracking pipeline and machine learning classifiers, augmented by SHAP explainability, to relate spatter ejection to melt pool dynamics. FLOW-3D data provide faster, economy-focused validation and enable a reduced-order approach that leverages ML trained on OpenFOAM data to generate a comprehensive spatter process map across parameter space. The results demonstrate high classification accuracy ($97\%-99\%$) and identify key discriminative features (notably $z$-axis position, recoil pressure, and $z$-velocity) with interpretable insights from SHAP and PDP analyses. The resulting spatter process map highlights processing windows that minimize spatter and intersects with porosity regimes, offering practical guidance for achieving low-defect LPBF parts at scale.

Abstract

Laser powder bed fusion (LPBF) has shown promise for wide range of applications due to its ability to fabricate freeform geometries and generate a controlled microstructure. However, components generated by LPBF still possess sub-optimal mechanical properties due to the defects that are created during laser-material interactions. In this work, we investigate mechanism of spatter formation, using a high-fidelity modelling tool that was built to simulate the multi-physics phenomena in LPBF. The modelling tool have the capability to capture the 3D resolution of the meltpool and the spatter behavior. To understand spatter behavior and formation, we reveal its properties at ejection and evaluate its variation from the meltpool, the source where it is formed. The dataset of the spatter and the meltpool collected consist of 50 % spatter and 50 % melt pool samples, with features that include position components, velocity components, velocity magnitude, temperature, density and pressure. The relationship between the spatter and the meltpool were evaluated via correlation analysis and machine learning (ML) algorithms for classification tasks. Upon screening different ML algorithms on the dataset, a high accuracy was observed for all the ML models, with ExtraTrees having the highest at 96 % and KNN having the lowest at 94 %.

Figures (17)

• Figure 1: Spatter ejection and melt pool parameters from OpenFOAM simulations were used as input for machine learning classification tasks. The trained models were tested on FLOW-3D datasets, a commercial software that lacks realistic spatter simulation capability to generate a process map. A 5 $\mu m$ LPBF simulation in FLOW-3D and OpenFOAM requires 4 hours and 72 hours respectively on the same machine.
• Figure 2: Schematic illustration of the computational domain with random packing of stainless steel powder bed in for OpenFOAM simulation.
• Figure 3: A tracking algorithm is implemented to consistently identify spatter particles as they move through the domain. The tracking process is demonstrated for two sample particles $i$ and $j$, between the timesteps ${n}$ and ${n+1}$. During the tracking process, the velocity field of the particle is used to predict an estimated displacement $\delta x_{i}$ within $t_{n+1} - t_{n} = \Delta t$. Next, the actual particle positions at the next timestep, $i_{t_{n+1}}$, are correlated to the estimated displacement values $i'_{t_{n+1}}$ based on a nearest neighbors assignment. Following this, each particle $i_{t_n}$ is linked to $i_{t_{n+1}}$, completing the tracking process.
• Figure 4: Mesh dependence studies on the temperature profile (a) FLOW-3D (b) OpenFOAM
• Figure 5: Implementation of the two-color thermal imaging method as described in myers2023high for the comparison of the observed surface temperatures for SS316L with the corresponding surface temperatures obtained from modelling packages: a) Measurements taken at P = 150 W, V = 1000 mm/s b)Measurements taken at P = 300 W, V = 1000 mm/s. The error bars from the experimental measurements represent combined uncertainties in emissivity and signal, as detailed in myers2023high.