Inverse modeling of process parameters from data to predict the cooling behavior in injection molding

Abstract

AI methods, especially Deep Learning Methods (DLMs), present an excellent opportunity for the surrogate modeling of complex processes, like the injection molding process, based on simulation or measurement data. To overcome the need for large data sets DLMs usually require, the integration of domain knowledge into the learning process e.g., in the form of Partial Differential Equations (PDEs), is rising in popularity. In this study, an inverse approach based on Physics Informed Neural Networks (PINNs) is explored for parameterizing the influence of material and process parameters on the cooling behavior of an injection molded part. With the proposed method, the PDE, Initial Condition (IC), and Boundary Condition (BC) of the underlying physical process can be automatically parameterized based on data. To reduce modeling effort, a simplified generic representation of the physical process description is used. The effectiveness of the method is demonstrated by utilizing the inversely learned physical process model to regularize the surrogate model. When predicting the spatiotemporal temperature evolution dependent on different materials and process settings, a 25% lower Root-Mean-Squared-Error (RMSE) was achieved by the hybrid approach in comparison to a purely data-driven model. The use of the simplified process physics highlights the generalizability of the approach to other data types and processes.