| Abstract Scope |
Additive manufactured (AM) parts are typically inspected post-process using such techniques as X-ray computed tomography, optical and scanning electron microscopy, among others. This empirical build-and-test, cook-and-look approach for qualification of AM part quality is prohibitively expensive and cumbersome. Whilst purely data-driven AI models based on real-time sensor data have been successfully used for defect detection, however, their generalizability beyond simple shapes to practical AM parts remains an open challenge. As a departure from this status quo, we propose data and physics-integrated graybox AI models. The key idea is to train machine learning models that combine (digitally twin) real-time data from multiple in-situ sensors with predictions from physics-based simulations. We show that such an approach predicts the onset of porosity, meltpool depth, grain size, and microhardness with an accuracy exceeding 90% (R2). |