ForgeNet: A Point Cloud Autoencoder for Predicting Plastic Deformation in Open-die Forging Processes

Abstract Open-die metal forging is challenging to automate due to the complex dynamics involved during deformation. Modern approaches use finite-element methods (FEM) to manually pre-plan deformation sequences on a per-part basis, requiring significant development and computation time. Inspired by autoencoder-based point cloud regression and segmentation methods, we introduce ForgeNet, a point cloud autoencoder that predicts the material displacement field given a state-action pair consisting of a metal workpiece and a deformation boundary condition. Standard point cloud autoencoders are trained with a permutation-invariant chamfer loss, which does not preserve point-to-point correspondence and therefore reproduces physically inconsistent displacement fields. ForgeNet uses barycentric point sampling to preserve point correspondence between deformations, enabling training with a standard mean-squared error loss that yields physically consistent displacement fields. We train our model on a large synthetic dataset simulated using JAX-Forge [1], a GPU-accelerated FEM package for open-die forging. Empirically, ForgeNet learns the dynamics of plastic deformation, accurately reconstructing displacement fields over long sequences, whereas standard point-based autoencoders reproduce only statistically similar point clouds after a single action. This allows ForgeNet to serve as a state-transition model that accelerates planning for optimal forging sequences, which we demonstrate on geometries outside our training dataset.