Reproduce “Surrogate Modelling for Injection Molding Processes using Machine Learning” on OpenResearch

About this paper

Arsenii Uglov, Sergei Nikolaev, Sergei Belov, Daniil Padalitsa, Tatiana Greenkina, Marco San Biagio, Fabio Cacciatori

The authors train a surrogate model on Moldflow simulations to predict spatial fields of plastic fill time and finished-part deflection. It combines gate features and a gradient-boosting predictor for fill time with a projected U-Net-based convolutional model for deflection. In reported comparisons, deflection prediction was 17.34 times faster on average; this baseline uses 158 simulations of one polymer without cooling.

Key insights

• Machine learning, specifically Gradient Boosting and CNNs, can serve as efficient surrogate models for complex injection molding simulations, significantly reducing computation time.
• Strategic feature engineering, such as incorporating geodesic distances to gates and gate opening times, is crucial for accurate fill time prediction in 3D geometries.
• Projecting 3D object properties onto 2D feature maps allows for leveraging powerful 2D convolutional architectures to predict spatial distributions like deflection, which can then be re-projected to 3D.

Summary by alphaXiv.