In-Flight Auto-Batcher

Relaxing an atomic structure towards its ground state takes an unpredictable number of ionic steps, so a fixed batch leaves the GPU idle while it waits for the slowest member. In-flight auto-batching swaps each converged structure out for a new one as soon as it finishes, keeping the device full.

• Batch size varies with the memory each structure needs, set mainly by atom count and density (denser cells mean more edges in the graph representation, often scaling cubically)
• Memory scaling can be configured on n_atoms or n_atoms × density

See the TorchSim InFlightAutoBatcher docs.