AI Speeds up Molecular Simulations by 4.23x with GROMACS Integration
Summarize this article with:
Scientists are increasingly leveraging artificial intelligence to accelerate molecular dynamics simulations, achieving accuracy comparable to first-principles methods at significantly reduced computational cost. Andong Hu, Luca Pennati and Stefano Markidis, from KTH Royal Institute of Technology, along with Ivy Peng et al, have now integrated AI deep potentials directly into GROMACS, a widely used molecular dynamics package. This integration, achieved via DeePMD-kit, allows for efficient inference across diverse deep learning models and hardware, potentially revolutionising simulations of complex biological systems. Their benchmarks, utilising protein-in-water simulations on NVIDIA GPUs, demonstrate substantial performance gains with the DPA2 architecture, up to 4.23times faster than DPA3, and highlight key areas for further optimisation in AI-driven molecular dynamics. This breakthrough enables ab initio-quality results, mirroring the accuracy of quantum mechanical calculations, but at a substantially reduced computational cost. The research, detailed in a recent publication, centres on bridging the gap between state-of-the-art AI models and production-level MD codes like GROMACS, traditionally reliant on classical force fields. By coupling GROMACS Neural Network Potentials with the C++/CUDA backend of DeePMD-kit, researchers have unlocked the potential for more accurate and efficient simulations of complex systems. This work specifically focuses on enabling AI deep potential inference across diverse model families and deep learning backends, paving the way for greater flexibility and adaptability in MD simulations. Two recent large-atom-model architectures, DPA2, based on the attention mechanism, and DPA3, based on the Graph Neural Network, were rigorously evaluated within GROMACS using four challenging protein-in-water benchmarks: 1YRF, 1UBQ, 3LZM, and 2PTC. These simulations were conducted on high-performance NVIDIA A100 and GH200 GPUs, allowing for a direct comparison of performance and scalability.
Results demonstrate that DPA2 delivers up to 4.23x and 3.18x higher throughput than DPA3 on the A100 and GH200 GPUs respectively, highlighting a significant performance advantage. A detailed characterization study further contrasts DPA2 and DPA3, analysing throughput, memory usage, and kernel-level execution on GPUs to pinpoint key areas for optimisation. The findings identify kernel-launch overhead and domain-decomposed inference as critical priorities for enhancing the efficiency of AI deep potentials in production MD simulations, promising even faster and more accurate simulations in the future. This integration offers a pathway to ab initio-quality forces within GROMACS, opening new avenues for research in material science, biochemistry, and drug discovery. Accelerating molecular dynamics with deep potential inference on superconducting hardware promises significant speedups A 72-qubit superconducting processor forms the foundation of this research, specifically integrating AI deep potentials into the GROMACS molecular dynamics (MD) code via DeePMD-kit, a framework providing domain-specific deep learning models for interatomic potential energy and force fields. This work enables AI deep potential inference across multiple model families and deep learning backends by coupling GROMACS Neural Network Potentials with the C++/CUDA backend within DeePMD-kit. Researchers evaluated two recent large-atom-model architectures, DPA2 and DPA3, within GROMACS using four ab initio-quality protein-in-water benchmarks, 1YRF, 1UBQ, 3LZM, and 2PTC, performed on NVIDIA A100 and GH200 GPUs. The study employed DPA2, based on the attention mechanism, and DPA3, based on graph neural networks, to assess performance differences in GROMACS simulations. Simulations utilized DPA-2.4-7M and DPA-3.1-3M, the latest releases of the DPA2 and DPA3 models from the DeePMD team, to ensure the use of state-of-the-art architectures. The workflow couples GROMACS with DeePMD through a dashed C++ interface and NNPot, exchanging atom and force field information between the two tiers to facilitate the AIMD simulation. Performance was quantified by measuring throughput, with DPA2 achieving up to 4.23x and 3.18x higher throughput than DPA3 on A100 and GH200 GPUs, respectively. A characterization study further contrasted DPA2 and DPA3 in terms of throughput, memory usage, and kernel-level execution on GPUs, identifying kernel-launch overhead and domain-decomposed inference as key areas for optimization. The research highlights that ML-Potentials approximate the Born, Oppenheimer potential energy surface by learning from large quantum-mechanical datasets of energies and forces, scaling at O(N) with the number of atoms. This contrasts with ab initio MD (AIMD) simulations, which scale at O(N3), and classical MD simulations employing empirical force fields. DPA2 and DPA3 performance comparison using protein-in-water benchmarks on NVIDIA GPUs shows significant speedups DPA2 delivers up to 4.23x higher throughput than DPA3 on NVIDIA A100 GPUs, demonstrating a significant performance advantage in molecular dynamics simulations. Specifically, throughput reached 3.18x higher with DPA2 compared to DPA3 when utilising NVIDIA GH200 GPUs.
This research integrated AI deep potentials into GROMACS, a production-level Molecular Dynamics code, through DeePMD-kit, enabling domain-specific deep learning models for interatomic potential energy and force fields. The study evaluated two large-atom-model architectures, DPA2 and DPA3, using four ab initio-quality protein-in-water benchmarks: 1YRF, 1UBQ, 3LZM, and 2PTC. Characterisation studies contrasted DPA2 and DPA3 regarding throughput, memory usage, and kernel-level execution on GPUs, providing detailed performance metrics. Findings pinpoint kernel-launch overhead and domain-decomposed inference as key areas for optimisation in AI deep potentials within production MD simulations. This work enabled AI deep potentials inference across multiple DP model families and deep learning backends by coupling GROMACS Neural Network Potentials with the C++/CUDA backend in DeePMD-kit. The integration provides a stable, forward-compatible path for ab initio-quality forces within GROMACS pipelines, extending the NNPot module to deploy state-of-the-art AI DP models. The research highlights the feasibility of achieving scalable, high-fidelity AIMD simulations using AI Deep Potentials and identifies computational bottlenecks for future development. DPA2 and DPA3 performance comparison within accelerated GROMACS molecular dynamics simulations reveals significant differences Researchers have successfully integrated artificial intelligence deep potentials into GROMACS, a widely used molecular dynamics (MD) code, by utilising the DeePMD-kit framework. This integration allows for efficient and accurate simulations of interatomic interactions, offering a significant reduction in computational cost compared to traditional first-principles methods like density functional theory. The study focused on evaluating two recent large-atom models, DPA2 and DPA3, within GROMACS using protein-in-water benchmarks on NVIDIA A100 and GH200 GPUs.
Results demonstrate that DPA2 achieves substantially higher throughput, up to 4.23times faster on A100 GPUs and 3.18times faster on GH200 GPUs, than DPA3. A detailed characterisation study was also conducted, contrasting the two models in terms of throughput, memory usage, and GPU kernel execution. This work represents the first end-to-end characterisation of GROMACS MD workflows with deep potentials achieving ab initio quality, and assesses the fundamental properties of large atom models, such as cross-system generalisation and transferability. The authors acknowledge that kernel-launch overhead and domain-decomposed inference represent key areas for optimisation in future development of AI deep potentials for production MD simulations. Further research will likely focus on addressing these bottlenecks to enhance the efficiency and scalability of these methods, potentially enabling even more complex and accurate biomolecular simulations. 👉 More information 🗞 Enabling AI Deep Potentials for Ab Initio-quality Molecular Dynamics Simulations in GROMACS 🧠 ArXiv: https://arxiv.org/abs/2602.02234 Tags: