![[2602.18146] Stable Long-Horizon Spatiotemporal Prediction on Meshes Using Latent Multiscale Recurrent Graph Neural Networks](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2602.18146] Stable Long-Horizon Spatiotemporal Prediction on Meshes Using Latent Multiscale Recurrent Graph Neural Networks
Summary
This paper presents a novel deep learning framework for stable long-horizon spatiotemporal predictions on complex geometries, particularly in additive manufacturing applications.
Why It Matters
Accurate long-horizon predictions are critical in scientific machine learning, especially in fields like additive manufacturing where temperature control is vital for product quality. This research addresses the limitations of existing methods, offering a more efficient and stable approach that can generalize across various geometries.
Key Takeaways
- Introduces a deep learning framework for predicting temperature histories on meshes.
- Utilizes a latent recurrent graph neural network for capturing spatiotemporal dynamics.
- Demonstrates improved stability and accuracy over existing baseline methods.
- Framework is extensible to three-dimensional geometries and physics-driven systems.
- Addresses challenges in long-horizon predictions in scientific applications.
Computer Science > Machine Learning arXiv:2602.18146 (cs) [Submitted on 20 Feb 2026] Title:Stable Long-Horizon Spatiotemporal Prediction on Meshes Using Latent Multiscale Recurrent Graph Neural Networks Authors:Lionel Salesses, Larbi Arbaoui, Tariq Benamara, Arnaud Francois, Caroline Sainvitu View a PDF of the paper titled Stable Long-Horizon Spatiotemporal Prediction on Meshes Using Latent Multiscale Recurrent Graph Neural Networks, by Lionel Salesses and 4 other authors View PDF HTML (experimental) Abstract:Accurate long-horizon prediction of spatiotemporal fields on complex geometries is a fundamental challenge in scientific machine learning, with applications such as additive manufacturing where temperature histories govern defect formation and mechanical properties. High-fidelity simulations are accurate but computationally costly, and despite recent advances, machine learning methods remain challenged by long-horizon temperature and gradient prediction. We propose a deep learning framework for predicting full temperature histories directly on meshes, conditioned on geometry and process parameters, while maintaining stability over thousands of time steps and generalizing across heterogeneous geometries. The framework adopts a temporal multiscale architecture composed of two coupled models operating at complementary time scales. Both models rely on a latent recurrent graph neural network to capture spatiotemporal dynamics on meshes, while a variational graph autoencode...
