A transfer learning approach to the prediction of porosity in additively manufactured metallic components

Abstract

We implement a physics-informed neural network (PINN) pretrained on a synthetic dataset to quantify distributed porosity in additively manufactured AlSi10Mg components using experimental ultrasonic pulse-echo data. The proposed PINN framework directly processes raw ultrasonic data to estimate volumetric porosity and average pore size. Due to the significant data requirements of neural network (NN) models, training is initially conducted on a dataset generated through finite element simulations. Then the pretrained PINN is transferred to the experimental data after using a portion of them for retraining. The PINN integrates physics constraints based on Sayers scattering model, which relates wave speed to porosity and pore radius. Notably, the two material-dependent constants in the model are treated as learnable parameters, which converge to their true values during the training process. Results indicate that the PINN achieves superb prediction accuracy, reflected in high r2-scores and low RMSEs. Additionally, a performance evaluation study is conducted by varying training set sizes; the PINN consistently outperforms the corresponding data-driven model (without physics constraint) across all training set sizes, with its advantage becoming more pronounced as the training set size decreases. Our findings suggest that a feed-forward neural network informed by wave physics can accurately quantify the porosity and pore radius within our samples from their raw ultrasonic responses even when the amount of labelled data for training is limited.