Abstract:We present the GPU version of DeePMD-kit, which, upon training a deep neural network model using ab initio data, can drive extremely large-scale molecular dynamics (MD) simulation with ab initio accuracy. Our tests show that the GPU version is 7 times faster than the CPU version with the same power consumption. The code can scale up to the entire Summit supercomputer. For a copper system of 113, 246, 208 atoms, the code can perform one nanosecond MD simulation per day, reaching a peak performance of 86 PFLOPS (43% of the peak). Such unprecedented ability to perform MD simulation with ab initio accuracy opens up the possibility of studying many important issues in materials and molecules, such as heterogeneous catalysis, electrochemical cells, irradiation damage, crack propagation, and biochemical reactions.

What problem does this paper attempt to address?

The paper attempts to address the problem of achieving ab initio level accuracy in large-scale molecular dynamics (MD) simulations while improving computational efficiency. Specifically, the paper introduces a method based on deep neural networks (Deep Potential, DP), which achieves efficient and high-precision molecular dynamics simulations of systems containing millions or even billions of atoms on supercomputers through optimization and parallelization. ### Main Issues 1. **Computational Efficiency**: Traditional ab initio molecular dynamics (AIMD) methods, due to their high computational complexity, can only handle systems with hundreds of atoms and are limited in time scale. While empirical force fields (EFF) can handle larger systems, their accuracy is low and cannot meet practical needs. 2. **System Scale**: For many practical problems, such as heterogeneous catalysis, electrochemical batteries, irradiation damage, crack propagation, and biochemical reactions, systems containing thousands to millions of atoms need to be handled. Existing methods find it difficult to handle such large-scale systems while ensuring accuracy. ### Solutions 1. **Deep Potential Model (DP)**: A deep potential model was constructed using the deep learning framework TensorFlow. This model can accurately describe many-body atomic interactions and can generate high-precision potential energy surfaces by training on ab initio data. 2. **GPU Optimization**: GPU optimization was performed on DeePMD-kit, including the optimization of custom TensorFlow operators and the refactoring of standard TensorFlow operators, enabling the code to run efficiently on modern heterogeneous supercomputers (such as Summit). 3. **Parallelization**: Efficient parallelization strategies were adopted, allowing DeePMD-kit to run in large-scale parallel environments, fully utilizing the computational resources of supercomputers. ### Main Contributions 1. **High-Performance Computing**: Through optimization and parallelization, DeePMD-kit achieved a peak performance of 86 PFLOPS on the Summit supercomputer, accounting for 43% of Summit's peak performance. 2. **Large-Scale Simulations**: It can simulate 113 million atoms in a copper system, completing 1 nanosecond of MD simulation per day; in a water system, it can simulate 4 million molecules, completing 110 MD steps per second. 3. **Performance Improvement**: Compared to the CPU version, the GPU version is 39 times faster with the same number of nodes and 7 times faster with the same power consumption. ### Conclusion Through deep learning and GPU optimization, DeePMD-kit successfully achieves efficient and high-precision molecular dynamics simulations on large-scale systems, providing new tools and methods for studying various important problems in materials and molecular research.

86 PFLOPS Deep Potential Molecular Dynamics simulation of 100 million atoms with ab initio accuracy

86 PFLOPS Deep Potential Molecular Dynamics simulation of 100 million atoms with ab initio accuracy

Pushing the limit of molecular dynamics with ab initio accuracy to 100 million atoms with machine learning

Scaling Molecular Dynamics with ab initio Accuracy to 149 Nanoseconds per Day

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

Molecular dynamics simulations with many-body potentials on multiple GPUs - The implementation, package and performance.

Deep Potential Molecular Dynamics: a scalable model with the accuracy of quantum mechanics

Efficient Deep Molecular Dynamic Model Training on Heterogeneous System

High-speed and low-power molecular dynamics processing unit (MDPU) with ab initio accuracy

Molecular dynamics simulation of macromolecules using graphics processing unit

GPU acceleration of ab initio simulations of large-scale identical particles based on path integral molecular dynamics

Algorithms of GPU-enabled reactive force field (ReaxFF) molecular dynamics

DeePMD-kit: A deep learning package for many-body potential energy representation and molecular dynamics

Robust and effective ab initio molecular dynamics simulations on the GPU cloud infrastructure using the Schrödinger Materials Science Suite

Towards Electronic Structure-Based Ab-Initio Molecular Dynamics Simulations with Hundreds of Millions of Atoms

Fast plane wave density functional theory molecular dynamics calculations on multi-GPU machines

TensorMD: Scalable Tensor-Diagram based Machine Learning Interatomic Potential on Heterogeneous Many-Core Processors

Graphics Processing Unit Acceleration and Parallelization of GENESIS for Large-Scale Molecular Dynamics Simulations

Training One DeePMD Model in Minutes: a Step Towards Online Learning.