Graph Autoencoders with Structure-Preserving Port-Hamiltonian Latent Dynamics for Thermal Reduced-Order Modelling in Directed Energy Deposition
Reduced-order models (ROMs) for transient finite element simulations are commonly built on Proper Orthogonal Decomposition (POD), which is restricted to linear subspaces and struggles with localised features such as the steep thermal gradient around a moving laser source. We combine hierarchical graph autoencoders, which operate directly on the unstructured finite element mesh via graph convolutions and fixed geometric pooling, with structure-preserving port-Hamiltonian neural networks (sPHNN) that guarantee energy dissipation and Lyapunov stability of the latent dynamics by construction. Four compression back-ends (POD, MLP autoencoder, graph autoencoder (GAE) and variational graph autoencoder (VGAE)) are compared under identical conditions on a Directed Energy Deposition thermal problem with 1,836 mesh nodes, trained on four trajectories and tested on 21 unseen laser paths. At default settings the GAE reaches the lowest test RMSE (2.81 K vs. 3.26 K for the MLP and 3.53 K for POD) with 57% fewer trainable parameters than the MLP; across five seeds GAE and MLP are statistically on par. KL regularisation in the VGAE degrades dynamics accuracy, and POD remains the most data-efficient option with only two training trajectories. The results indicate that compression methods for ROMs should be evaluated jointly with the latent dynamics model rather than by reconstruction quality alone.
Authors
- Obada Allaham
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-14
- DOI
- https://doi.org/10.5281/zenodo.24218585
- Primary Topic
- Model Reduction and Neural Networks
- Type
- preprint