Neural network surrogates for finite element models in loaded tooth contact analysis of polymeric gears

This paper introduces a neural network (NN) model for loaded tooth contact analysis in polymeric gears. The NN model is trained on 10,000 finite element (FE) simulations, which utilize a large deformation framework and span a 17-dimensional input space, including geometry, material and load related parameters. Combining a parametric meshing scheme and a robust implicit solver implementation enables the automated extraction of static transmission error (STE) curves. Leveraging the symmetry of periodicity and a conversion from a dimensional to a dimensionless parametric space, an approximately 9,000-parameter fully connected neural network achieves a mean absolute percentage error of 0.49% on a test set of 1,000 previously unseen STE curves. This error represents an order of magnitude more accurate replication of FE results than typical, analytical, physics-based solvers, with the effects of corner contact being predicted more faithfully. The computational cost of the NN model remains comparable to simple, linearly approximated formulas, indicating that data-driven approaches can be both more accurate and less computationally intensive than physics-based surrogates to FE simulations for large deformations.