Efficient scaling of neural-network interatomic potential prediction on CPU clusters
US11815945B2 · kind B2 · utility
Assignee
Inventors
Key dates
| Filing date | Sep 29, 2020 |
| Grant date | Nov 14, 2023 |
| Priority date | — |
| Expiry date | Jun 28, 2042 |
Classification
- Technology area (CPC G)Physics
- CPC primaryG06F2111/02
- WIPO fieldComputer technology
- WIPO sectorElectrical engineering
Abstract
Element simulation is described using a machine learning system parallelized across a plurality of processors. A multi-element system is partitioned into a plurality of partitions, each partition including a subset of real elements included within the partition and ghost elements outside the partition influencing the real elements. For each processor of the plurality of processors, force vectors are predicted for the real elements within the multi-element system by making a backward pass through a graph neural network (GNN) having multiple layers and parallelized across multiple processors, the predicting including adjusting neighbor distance separately for each of the multiple layers of the GNN. A physical phenomenon is described based on the force vectors.
Source: USPTO / EPO open patent data. Objective bibliographic and citation counts.