A universal force field leaves the supercomputer behind: ALIGNN-FF evaluates two billion atoms on desk-side GB10s
Synopsis
The universal graph neural network force field ALIGNN-FF was previously capped by single-GPU memory at 216,000 atoms with forces as one graph on a desk-side GB10. Cutting the periodic cell into cubes, each evaluated with a halo covering the model's 20 Å receptive field, makes per-piece memory independent of system size, and nine GB10s evaluate energies and forces of 2,000,376,000 silicon atoms, with the largest piece at 104,194 atoms. A retrained AFF-r2SCAN-smooth with smooth cutoffs also removes the force jump at the three-body cutoff, so relaxation converges and energy is conserved.
Interpretation
A universal machine-learning force field can now evaluate energies and forces of billion-atom systems on desk-side hardware one group can own, without a leadership-class supercomputer. Earlier machine-learning force field records above a billion atoms all ran on leadership-class machines and used potentials fitted to one system or one class of systems; universal force fields, with their higher per-atom cost, usually ran at the million-atom scale. On nine GB10s, the energy evaluation of 2,000,376,000 silicon atoms took 17.7 GB10-hours and 72.3 with forces, with at most 5.7 GB of host memory per task; the billion-atom silicon runs took 9.3 and 43.2 GB10-hours, the slowest force task 3.7 hours, and the largest piece of 71,578 atoms used no more than 31.9 GB of GPU memory.
The method cuts the periodic cell into cubes and evaluates each cube together with a halo of surrounding atoms that covers everything able to influence an atom in the model. Parallel molecular dynamics already used spatial decomposition with ghost atoms, but the ghost region was usually one cutoff wide; message-passing force fields require the halo to be at least the receptive field, which is 20 Å for ALIGNN-FF. On 8,000-atom cells that fit whole, chunked and full-cell results agree in double precision, and differ by no more than float32 roundoff in single precision; in water the 52-neighbor cap binds, so the 20 Å halo is no longer exact, leaving an error about five orders of magnitude below single-precision roundoff that a 25 Å halo removes.
The published model's forces jump at its three-body cutoff, and retraining with smooth switches yields AFF-r2SCAN-smooth, which removes the jump. The original AFF-r2SCAN keeps a line-graph bond only if it is at most 3.5 Å long and does not smooth the angle terms, so moving a bond 0.0004 Å across that length changes forces by 0.47 to 0.51 eV/Å, against 0.005 eV/Å for the same move far from any cutoff. On the 992 MatPES frames in both test sets, AFF-r2SCAN-smooth has mean absolute errors of 45.7 meV/atom and 0.144 eV/Å, against 52.0 and 0.207 for AFF-r2SCAN; the same relaxation converges in 658 steps and 1.35 hours with every force below 0.05 eV/Å, and energy drift in the two dynamics runs falls to 0.08 and 0.9 meV/atom.
Perspective
The demonstrated scope is single-point energy and force evaluation: silicon systems of 8 million, 64 million, one billion and 2,000,376,000 atoms, a billion-atom silicon polycrystal with 4,800 grains, a 102-million-atom silicon/silicon-dioxide/aluminum gate stack, and a 1.07-million-atom satellite tobacco mosaic virus in water, with the largest run at 2,000,376,000 silicon atoms on nine GB10s. The billion-atom timings for aluminum, copper and water are cost estimates extrapolated from samples of about a million atoms, not actual runs. The authors state that the decomposition is not specific to ALIGNN-FF, that other message-passing force fields can set the halo to their own receptive field, with long-range electrostatics models as the exception, and that scripts are released with the work.
The authors note that the largest systems are evaluated only as single points, and that a thousand 1 fs steps of a billion atoms would take about 4.9 years on one GB10; for structures read from disk each task holds the whole cell in host memory, about 60 to 74 GB per billion atoms, capping one GB10 at roughly 1.1 to 1.4 billion atoms; in water the 52-neighbor cap makes the 20 Å halo no longer exact; the virus run measures size and cost only, because the model was trained on inorganic materials and has not been validated on biomolecules; and the other published ALIGNN-FF models show the same force jump at 3.5 Å.
