From Photodiode Signals to CT-Quality Porosity Maps: A Deep Learning-Based Image-to-Image Translation Approach for Porosity Prediction in Laser Powder Bed Fusion

Understanding and predicting thermal characteristics of the melt pool including thermal emission, temperature gradients and cooling rates, is essential for ensuring the formation quality and defect control of metal components produced by Laser Powder Bed Fusion (LPBF). In this study, we introduce a novel generative adversarial network (GAN) based image-to-image translation model that converts in-situ photodiode thermal emission signals into visualized Micro-CT porosity maps. Initially, a single-layer input model was developed to enable real-time prediction of pore distribution. Based on this, a multi-layer network structure was further developed to improve the accuracy of 3D pore reconstruction of the samples. To our knowledge, this is the first application of an image generation model for real-time defect detection in LPBF, enabling CT-level information from low-cost photodiode sensors. Our method rapidly generates high-resolution porosity maps with a minimum resolution of 10 microns at the millisecond level and accurately predicts the spatial distribution and size of pores in LPBF-fabricated AlSi10Mg. Experimental results demonstrate exceptional performance, with some samples achieving an F1-score of 1.0, indicating perfect defect detection, and high spatial alignment between predicted porosity maps and ground truth CT scans. These findings underscore the potential of generative models to transform defect detection and process optimization in additive manufacturing.