Alf Rodin Christensen will present his master's thesis

Title: Industrial Implementation of an Inverse Physics-Informed Neural Network for Viscosity Modelling in Injection Molding Processes

Author: Alf Rodin Christensen

Supervisor: Anders Heyden

Examiner: Alexandros Sopasakis

Abstract:

Injection molding is a widely used manufacturing process in which viscosity constitutes one of the most important material properties governing flow behavior. Accurate viscosity modelling is essential for reliable simulation and process optimization. Direct measurement of viscosity, however, can be difficult to perform in industrial environments due to the high cost and complexity of sensor installation. As a result, indirect estimation methods based on available machine data are of increasing interest. This thesis investigates the application of an inverse physics-informed neural network (iPINN) for estimating apparent viscosity in an injection molding process. The work builds upon a previously developed iPINN framework trained on simulation-generated data from a computational fluid dynamics (CFD) program, and focuses on adapting the method to real machine measurements. A central challenge addressed in this work is the discrepancy between idealized CFD data and the sparse, noisy sensor data collected from industrial injection molding equipment. Several domain-alignment strategies are introduced to enable the use of machine data as inputs to
the iPINN, including data selection procedures, inference of velocity from indirect measurements, and pressure-domain alignment based on measurement availability, CFD simulation data, and process physics. The iPINN combines a power-law description of apparent viscosity with simplified Navier–Stokes equations, which are enforced through physics-informed loss terms during training.
Additional optimization strategies, including bounded parameter optimization and scheduling of physics-loss contributions, are applied to improve training stability and robustness. The results indicate that the adapted model can produce physically feasible viscosity estimates from industrial machine data. Although quantitative validation is limited by the absence of direct reference meassurements, the predicted behavior is consistent with known rheological characteristics of polymer melts under injection molding conditions.