Warpage prediction for fiber reinforced injection molding via geometric feature learning and differentiable FEM
- Injection molding is a popular mass production process for short fiber reinforced components. One of the main production defects is warpage, unwanted deformation, resulting from the thermal history of the injected polymer and the geometry of the part. Ideally, this deformation should be predicted and compensated for before mass production to minimize defective parts. Conventional process simulation is able to predict warpage, but is too computationally expensive to be used in iterative optimization procedures. Hence, we propose a fast approximation method based on machine learning and a custom finite element solver to predict warpage in arbitrary 3D geometries with any injection location. It combines nodal geodesic values and spatial moments to capture local geometry on different scales. Neural networks then predict fiber orientation and initial strain fields for a subsequent warpage calculation in a differentiable finite element software. The latter enables training of the neuralInjection molding is a popular mass production process for short fiber reinforced components. One of the main production defects is warpage, unwanted deformation, resulting from the thermal history of the injected polymer and the geometry of the part. Ideally, this deformation should be predicted and compensated for before mass production to minimize defective parts. Conventional process simulation is able to predict warpage, but is too computationally expensive to be used in iterative optimization procedures. Hence, we propose a fast approximation method based on machine learning and a custom finite element solver to predict warpage in arbitrary 3D geometries with any injection location. It combines nodal geodesic values and spatial moments to capture local geometry on different scales. Neural networks then predict fiber orientation and initial strain fields for a subsequent warpage calculation in a differentiable finite element software. The latter enables training of the neural network through the solver with warpage based losses, enabling the training with observable deformation data. The approximation pipeline exhibits relative warpage errors of less than 1% on typical injection molded geometries while being six times faster than the conventional simulation. Retraining through the solver with a warpage enhanced dataset to emulate real world data leads to significant improvements in the predictive accuracy.…
