Work

A Simulation-Based Deep Learning Approach for Predicting Part Quality in Thermoplastic Injection Moulding

Year: 2026

Type: conference-paper

Abstract: In this work a deep learning approach for real-time warpage prediction in injection moulding is presented. A simplified geometry derived from an industrial HUD top cover is investigated as an example geometry. Simulation data from Autodesk Moldflow serve as the basis for training surrogate models. Two model variants are compared: a direct, fully data-driven prediction of the full displacement field by neural networks (NN) and a reduced-order approach combining Proper Orthogonal Decomposition with NN (POD+NN). Both approaches reach a $$R^2 \ge 0.94$$ R 2 ≥ 0.94 and have prediction times below 50 ms. The results show that the POD+NN model achieves high accuracy with reduced complexity, supporting efficient integration into digitalised manufacturing workflows. A prototype dashboard enables interactive analysis.

Source: Lecture notes in production engineering

DOI: 10.1007/978-3-032-19524-1_42

Authors: André Hürkamp, Tom Hartmann, Manuel Megnet, Klaus Dröder

Institution: Technische Universität Braunschweig

Language: English

Topic: Model Reduction and Neural Networks

Subfield: Statistical and Nonlinear Physics

Field: Physics and Astronomy

Domain: Physical Sciences

Open Access status: hybrid

Funder: Bundesministerium für Wirtschaft und Energie

Award: Verbundprojekt: Digitaler Zwilling für den KI-unterstützten Werkzeugentstehungsprozess (TooliNG) - Teilvorhaben: Skalierter Werkzeugentstehungsprozess