Researchers have developed a new physics-informed AI framework that predicts in-process distortion in metal additive manufacturing (AM) in real time, paving the way for adaptive control and digital twins in complex builds, tells this 3D Printing Industry article.
Traditional numerical simulation methods such as finite-element analysis offer accurate predictions of thermal stress and deformation, but they are too slow to run live during a print. Data-driven machine-learning models, meanwhile, often lack physical interpretability or struggle to capture the coupling between thermal history and mechanical stress.
The new method, named PIDeepONet-RNN, combines a neural-operator architecture with embedded physics constraints. It takes temperature history as input and predicts z– and y-direction distortions up to 15 seconds into the future. By encoding the heat-conduction equation as a soft constraint, the model stays physically consistent, avoids unphysical artifacts, and gives interpretable results tied to thermal history and stress evolution, not just black-box outputs.
In testing, the model achieved high accuracy: maximum absolute errors of under 1 millimeter in the z-direction and ~0.2 millimeter in the y-direction, even over extended horizons.
Because inference is fast enough for real-time use, the framework can integrate with digital-twin systems, enabling predictive monitoring, adaptive control of print parameters, and early detection of potential defects or distortions before they irreversibly form.
For engineers and manufacturers, this promises a shift from post-process inspection or slow simulation-based prechecks to live distortion control, improving yield, reducing scrap, and paving the way for more reliable, complex metal AM components.