Geometric Deep Learning for Predicting Thermo-mechanical Performance in Cold Spray Deposition
Автор: Akshansh Mishra
Загружено: 2026-03-17
Просмотров: 9
Описание:
My latest paper, now on arXiv, applies geometric deep learning to the cold spray deposition process.
The core idea is relational. Rather than treating each simulation case as an isolated observation, I constructed a k-nearest-neighbour graph in the input feature space. Each process condition learns from its neighbours. The geometry of the parameter space becomes part of the model itself.
Four architectures were tested, i.e., GraphSAGE, Chebyshev spectral network, TDA-augmented MLP, and a geometric attention network across four output targets, including average plastic strain, maximum plastic strain, temperature, and deformation ratio.
GraphSAGE and GAT consistently achieved R² above 0.93. GAT reached R² = 0.97 for maximum equivalent plastic strain. Spectral and topological approaches failed on several targets entirely.
Particle velocity dominates the response. But temperature and friction interact in ways that require the full three-dimensional input space to resolve, and that is exactly where graph-based neighbourhood aggregation proves its value.
Paper Link: https://arxiv.org/abs/2603.14478
Повторяем попытку...
Доступные форматы для скачивания:
Скачать видео
-
Информация по загрузке:
![Эффект Джанибекова [Veritasium]](https://imager.clipsaver.ru/N9HlQ-XVnFk/max.jpg)
