Towards rapid quality control in additive manufacturing: Explainable multimodal deep learning framework for powder flowability

Powder flowability affects the properties of the products in powder-based manufacturing processes including additive manufacturing and powder metallurgy. Traditional flowability tests—including the Angle of Repose (AOR), Hausner Ratio (HR), Carr Index (CI), and Hall flow rate—are widely used but have limitations in speed and operator dependence. The study uses a dataset of 183,120 SEM image–record pairs from 38 types of metallic powders acquired at multiple magnifications. Each SEM image was available in several processed variants to ensure compatibility with standard deep-learning pipelines and to improve robustness under different contrast conditions. Five convolutional neural network backbones (ResNet-18, EfficientNet-B0, MobileNetV3-Large, ShuffleNetV2, RegNetY-400MF) are evaluated. Among the tested architectures, ShuffleNetV2 achieves the trade-off between accuracy and efficiency, with the error (MAE 0.405, RMSE 0.770, MedAE 0.274) and R2 scores (0.988), validating the approach’s effectiveness. The proposed method provides rapid prediction of flowability improves testing efficiency, as conventional tests typically require about three hours per powder type, whereas our image-based model system achieves predictions within 20 minutes. Baseline and ablation studies reveal that the multimodal fusion of SEM images and magnification metadata achieves better accuracy and robustness compared to single-modality models. Attribute mapping using Grad-CAM and Integrated Gradients enhances interpretability, showing model attention closely aligns with microstructural features relevant to flowability prediction, thereby linking AI reasoning with physical powder behaviour.