A new arXiv preprint presents a mesh-based super-resolution method that uses graph transformers to reconstruct multiscale detonation flow data. The authors argue such data-driven reconstruction is useful for subgrid closure modeling, faster spatiotemporal forecasting, compression, and as an upsampling step in simulations. The work appears as a cross-listed revision in the cs.AI and cs.LG categories.
A new arXiv paper examines how different pre-training approaches affect graph transformer performance on biochemistry tasks. The authors report that pre-training with supervision, using computed molecular properties as labels, outperformed the other strategies tested. The finding comes from a set of comparative experiments run in that domain.
A new arXiv paper investigates how well Graph Transformers generalize when trained on graphs of one size and evaluated on larger or smaller ones. It focuses on positional encodings derived from graph neural networks, a common design element in these attention-based models for graph data. The work examines whether such encodings help preserve performance as graph size changes.