Interpretable Machine Learning for Rapid Deformation Prediction in AlSi10Mg Laser Powder Bed Fusion Using Layer-Wise Cross-Sectional Feature Sequences

Thermal deformation in laser powder bed fusion (L-PBF) critically affects part quality, yet accurate prediction remains challenging because finite element (FE) simulations are computationally prohibitive. This study proposes a lightweight and interpretable machine learning framework for rapid deformation prediction in AlSi10Mg L-PBF using layer-wise cross-sectional feature sequences. Because L-PBF proceeds layer by layer, ten geometric feature curves extracted from cross-sectional slices along the build direction are treated as time-series sequences, from which 7890 descriptors are generated using tsfresh, an open-source Python 3.9 library for automated time-series feature extraction. Multi-stage feature selection then selects 25 key descriptors, and a Gaussian process regression model with a Matérn kernel is trained using mean and maximum deformation values obtained from inherent strain simulation. On a single held-out internal test set generated by a sample-level 80%/20% split of the development data, the trained model achieved R2 values of 0.89 and 0.85 for mean and maximum deformation, respectively. On the geometry-disjoint independent test set containing three unseen geometries, the model achieved average PCC values of 0.759 and 0.831, RMSE values of 0.0082 mm and 0.0394 mm, and R2 values of 0.521 and 0.532 for mean and maximum deformation, respectively. The corresponding pairwise orientation-order accuracies were 0.753 and 0.812, both higher than the random-ordering baseline of 0.5. These results indicate that the model can capture orientation-dependent deformation trends for preliminary build-orientation screening. Once trained, feature extraction and deformation prediction for a supported part can be completed within seconds, enabling rapid deformation assessment for design and build-orientation screening. Moreover, SHapley Additive exPlanations (SHAP)analysis is used to examine model-specific associations between selected layer-wise geometric descriptors and predicted deformation. Overall, the framework provides a rapid and interpretable surrogate for deformation-aware L-PBF design, enabling early-stage build-orientation screening while reducing the need for repeated FE simulations in industrial process planning.