New work: Truncated automatic sparse differentiation

Hello, I'm not sure if anyone is actually reading this "blog" but I thought I should probably start also posting news...

A new preprint has arrived: Truncated automatic sparse differentiation for machine learning interatomic potentials with Adrian Hill and Michele Ceriotti. We make the calculation Hessians of MLIPs (which are needed for various kinds of phonon-based observables) faster.

The idea is simple: MLIPs (if they are based on MPNNs) have a finite cutoff, and the underlying physics (and message passing itself) decays with distance. So Hessians are both sparse (they have known zeros for atoms that don't interact) and decay (entries between faraway atoms are small). We exploit this in two related ways: sparsity can be used for automatic sparse AD, which computes the Hessian faster, yet exactly, by not bothering to compute known zeros. Then we go one step further: we truncate the sparsity pattern for faraway atoms, and only compute the remaining, significant, Hessian entries. This leads to noise, and large speedups. We find that the noise is manageable, and controllable by the severity of truncation (because Hessian entries decay).

We test the method on MOFs, COFs, and zeolites, with PET-MAD-1.5 and MACE. Production code will soon be in pet-jax. All the other information at tasd4mlip.