Extending the limit of molecular dynamics with ab initio accuracy to 10 billion atoms

Abstract

High-performance computing, together with a neural network model trained from data generated with first-principles methods, has greatly boosted applications of ab initio molecular dynamics in terms of spatial and temporal scales on modern supercomputers. Previous state-of-the-art can achieve 1 − 2 nanoseconds molecular dynamics simulation per day for 100-million atoms on the entire Summit supercomputer. In this paper, we have significantly reduced the memory footprint and computational time by a comprehensive approach with both algorithmic and system innovations. The neural network model is compressed by model tabulation, kernel fusion, and redundancy removal. Then optimizations such as acceleration of customized kernel, tabulation of activation function, MPI+OpenMP parallelization are implemented on GPU and ARM architectures. Testing results of the copper system show that the optimized code can scale up to the entire machine of both Fugaku and Summit, and the corresponding system size can be extended by a factor of 134 to an Permission to make digital or hard copies of part or all of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. Copyrights for thirdparty components of this work must be honored. For all other uses, contact the owner/author(s).

Combined QM/MM, Machine Learning Path Integral Approach to Compute Free Energy Profiles and Kinetic Isotope Effects in RNA Cleavage Reactions

“…18,19 Our implementation of the DPRc correction also supports GPU acceleration via the tensorflow library 102 and several custom operators. 60,61,103,104 The DFTB2 QM/MM+DPRc performance is only 26% slower than DFTB2 QM/MM when the correction is evaluated on a V100 NVIDIA GPU. In comparison to ab initio sampling, the DFTB2 QM/MM+DPRc method is faster by a factor of 251.…”

“…The network parameters are optimized using an active learning approach described in detail elsewhere . The parameter optimizations were performed with the DP-GEN software, and the DP Compress algorithm , was applied to the trained models to improve computational performance and reduce the memory requirements. The active learning procedure involves 3 stages: training, exploration, and labeling.…”

Micron-scale heterogeneous catalysis with Bayesian force fields from first principles and active learning

“…1) The speed of our model exceeds both the non-reactive DeePMD benchmark on simple thermal dynamics in homogeneous systems by 30× for bulk water [26] and 4× for bulk copper [27], the SNAP benchmark by 1.7× for bulk carbon [25], and the reactive force field ReaxFF by at least 5×. ReaxFF benchmark was on 1M atoms on one V100 GPU node, and does not scale linearly with the system size.…”

“…The SNAP model was employed on bulk carbon system with peak performance 6.21 M atoms•steps/s/node [25]. Likewise, DeePMD achieved 0.3 M atoms•steps/s/node for homogeneous bulk systems -water and copper -on the OLCF Summit machine [26], with the performance recently improved to 2.0 M atoms•steps/s/node [27]. Other models, like graph neural networks, have not been shown to approach such scales due to difficulty in parallelization.…”

“…The strong scaling results are shown in Figure 7. In the top panel, the speed of FLARE is compared to the peak performance of SNAP [25] and DeePMD [27]. FLARE outperforms SNAP and DeePMD by factors of 1.7 and 5.3, respectively, when in the regime of linear scaling, achieving a peak performance of 10.5 M atom-steps/s/node.…”