Using Physics-Informed Machine Learning to Optimize 3D Printing Processes

In this work, we present the design and evaluation of a physics-informed machine learning (ML) approach for 3D printing of metal components based on real experimental measurement data. For this purpose, different thermal processes during the manufacturing of different metal components are modelled and solved in space and time using physics-informed neural networks (PINNs), with special attention to the geometric domain of the heat source. In this way, a digital twin is created as a simulation model of the thermal process with fixed input parameters, which represents the physical behavior and allows interpretable conclusions to be drawn with regard to the temperature distribution. The presented approach includes discrete and continuous models that are compared with a numerical solution method (finite difference method, FDM) for partial differential equations (PDEs) and real measurement data from Siemens. In addition, learning is accelerated by designing the loss function with information about initial and boundary conditions (Neumann and Dirichlet) and the heat source. This approach does not require discretization and is robust to limited noisy data. The results have confirmed the advantages of PINNs for data-driven discovery of the heat equation and have shown that the temperature distribution during additive manufacturing (AM) processes can be estimated and predicted very well.