Local pore prediction in laser powder bed fusion of metals using dual off-axis photodiode monitoring and spatiotemporal deep learning

Accurate pore identification and localization remain major challenges in laser powder bed fusion of metals (PBFLB/ M). This study presents an X-ray computed tomography (XCT)-labeled framework for local pore detection using dual off-axis photodiode signals. Each build layer was divided into patches, and each target patch was stacked with spatially corresponding patches from adjacent layers to form a sampling cuboid. These cuboids were registered to XCT-derived labels. DualFusionNet combines 3D-CNN and ConvLSTM encoders: the 3D-CNN extracts localized volumetric correlations across tracks and layers, whereas the ConvLSTM preserves layer order to model thermal evolution associated with successive melting and remelting events. Trained and tested on M789 maraging steel, the model achieved an area under the receiver operating characteristic curve (AUROC) of 0.93. Probability-of-detection (POD) analysis yielded a corrected a90 of 96 μm, where a90 denotes the spherical equivalent diameter at which the fitted POD curve reaches 90%. Ablation studies confirmed that encoder diversity and fusion of the two photodiode channels improved pore prediction performance. Gradient-weighted class activation mapping revealed distinct yet partially overlapping salient regions within the sampling cuboid, consistent with the encoders’ complementary roles in capturing spatial-interlayer structure and sequential dependencies. DualFusionNet required 0.25 ms per patch for inference and completed end-to-end processing of a representative 784-patch layer in 5.8 s, including data loading and preprocessing. This fits within the nominal 6 s recoating interval, supporting near-real-time layerwise prediction. DualFusionNet achieved an AUROC of 0.91 and 85.0% recall on an additional 9 × 9 × 12 mm³ specimen produced using different process parameters (laser power: 440–410 W; scan speed: 820–750 mm/s; hatch spacing: 97–95 μm; and layer rotation: 67◦ to 90◦). Recall reached 91.9% for pores larger than the established 96 μm detection size, demonstrating within-platform transfer under moderate process variations. Extension to substantially different process windows, machines, or materials requires further validation and may necessitate sensor recalibration, adjustment of the sampling-cuboid dimensions, and model fine-tuning or retraining. These results establish a practical route toward in-situ local pore detection and layerwise decision support in PBF-LB/M tooling applications, providing a foundation for future corrective control.