Request: Benchmark Comparison with PyTorch Geometric's radius_graph for Chemistry Applications

[ericyuan00000]

Hi Vesin developers,

First off, thank you for the excellent work on Vesin! I've been exploring it as a potential neighbor list generator for atomistic systems, especially in the context of machine learning models for chemistry.

As someone evaluating Vesin alongside PyTorch Geometric's radius_graph function, I'd love to see a direct performance and accuracy benchmark between the two—particularly in chemistry-relevant settings. Both tools seem to target similar use cases, but Vesin's atomistic focus and support for periodic boundary conditions could be major advantages.

Would it be possible to add a benchmark suite or extended comparisons between Vesin and radius_graph under conditions like:

• Varying number of atoms (1k – 1M)
• Use of periodic boundary conditions (enabled/disabled)
• Timing and memory usage (CPU and GPU if applicable)

This would really help users in computational chemistry and ML better choose the right tool for their workloads. I’d be happy to contribute example molecule datasets or collaborate on benchmarking scripts if helpful.

Thanks again for building this—it’s a great contribution to the community!

[Luthaf]

Hey!

I'd be happy to add another tool to the existing benchmark suite https://github.com/Luthaf/vesin/blob/main/benchmarks/benchmark.py. Could you give me an example / reference documentation for this function? Or even better, send a pull-request adding setup_pyg_radius_graph_{cpu,cuda} and pyg_radius_graph_run to this file! This benchmark suite is limited to periodic systems, and would likely require significant refactoring to properly record memory usage.

I'm not sure having a comprehensive benchmark suite makes sense for this repository (benchmarks should be unbiased and separate from the implementation IMO), so it might make sense to create this in a separate repository. If you want to give this a go, feel free to take inspiration from the code here. I'll be happy to link to external benchmark results from vesin's documentation!

[Luthaf]

Ok, reading a bit more about it, it looks like radius_graph does not support PBC at all. I could do a custom algorithm for minimal image convention, duplicating the cell 27 times, but I'm afraid this would kill performance. Might be worth to wait until pyg-team/pytorch_geometric#10145 is implemented before adding a benchmark.

[ericyuan00000]

Yeah that's me too. I can do non-periodic systems with radius_graph but I have to duplicate the cell 27 times for PBC, which sounds like a terrible implementation. I do have this naive code available and let me see if I can just add them here.

On the other hand, do you have plans to do simple benchmarks for non-periodic systems? Is that included in benchmark.py currently?

[Luthaf]

No, the existing benchmarks all assume periodic systems. Is there a use case for non-periodic system where performance matters a lot? I can't find large, non-periodic systems that one would want to run simulations for.

[ericyuan00000]

That's a valid point!