Predicting Internal Porosity in Batch Metal Binder Jetting from Early-layer Stereovision Measurement via Graph-based Modeling

Abstract Internal defects, such as porosity within parts fabricated by metal binder jetting, often exhibit spatial variations that significantly influence part quality. Recent advances in stereovision sensing have enabled layer-wise surface measurements, providing new opportunities for efficient early-stage defect estimation during printing. Although prior studies have explored in-situ sensing for layer-wise inspection and defect detection, few studies have addressed the problem of predicting the full spatial distribution of internal porosity using only early-layer observations. In particular, the use of stereovision-based surface measurements to infer unobserved internal defects across the entire part before build completion remains largely unexplored. This paper proposes a two-scale modeling framework that leverages stereovision data for early prediction of internal defects, thereby enhancing process responsiveness and production efficiency. At the micro scale, a dual graph network framework is developed to predict unprinted deeper layers from limited early-layer data. At the macro scale, Gaussian processes are employed to capture spatial dependencies among parts within the same build volume. By integrating information from both printed and predicted layers, the framework captures the overall influence of spatial variations on internal defects. Outputs from both scales are fused through an ensemble meta-model to improve prediction accuracy and robustness. The method is validated through experimental case studies that evaluate the accuracy of part porosity prediction within a printing batch. It further demonstrates the early-stage detection of high porosity without requiring full-part printing, as shown in metal binder jetting with HP Inc.