[Paper] Enabling AI Deep Potentials for Ab Initio-quality Molecular Dynamics Simulations in GROMACS

Source: arXiv - 2602.02234v1

Overview

The paper demonstrates how to bring state‑of‑the‑art AI‑driven “deep potentials”—neural‑network models that reproduce ab‑initio quantum‑chemical accuracy—into GROMACS, one of the most widely used molecular‑dynamics (MD) engines. By tightly coupling GROMACS with the DeePMD‑kit library, the authors enable fast, production‑level simulations of complex biomolecular systems while keeping the computational cost far below that of traditional density‑functional theory (DFT).

Key Contributions

• Seamless integration of DeePMD‑kit’s C++/CUDA backend with GROMACS, exposing AI deep potentials as native “Neural‑Network Potentials” (NNPs).
• Support for multiple model families (attention‑based DPA2 and graph‑neural‑network‑based DPA3) and various deep‑learning frameworks, all callable from a single GROMACS executable.
• Comprehensive performance evaluation on four protein‑in‑water benchmarks (1YRF, 1UBQ, 3LZM, 2PTC) using NVIDIA A100 and GH200 GPUs.
• Quantitative throughput comparison: DPA2 achieves up to 4.23× (A100) and 3.18× (GH200) higher simulation speed than DPA3.
• In‑depth profiling of GPU kernel launches, memory footprints, and domain‑decomposed inference, pinpointing the main bottlenecks for future optimization.

Methodology

1. Model Selection – The authors chose two recent deep‑potential architectures:
   • DPA2: an attention‑mechanism model that aggregates atomic environments via learned attention weights.
   • DPA3: a graph‑neural‑network (GNN) model that treats atoms as nodes and bonds as edges.
2. Software Coupling – DeePMD‑kit already provides high‑performance inference kernels (C++/CUDA). The team wrapped these kernels in a GROMACS‑compatible API, allowing GROMACS to request energies and forces from the neural models during each MD step.
3. Benchmark Setup – Four realistic protein‑in‑water systems (ranging from ~10 k to ~50 k atoms) were simulated under NVT conditions. Each system was run on both NVIDIA A100 and NVIDIA GH200 GPUs, measuring wall‑clock time per MD step, GPU memory usage, and kernel‑level statistics.
4. Profiling & Analysis – NVIDIA Nsight and custom timers captured kernel launch overhead, occupancy, and data movement. The authors compared the two models across the same hardware and workloads to isolate algorithmic vs. implementation effects.

Results & Findings

• Memory Footprint: DPA3 required ~30 % more GPU memory due to larger intermediate tensors in its GNN layers.
• Kernel‑Launch Overhead: A sizable fraction of total runtime (≈15‑20 %) stemmed from frequent small kernel launches, especially for DPA3.
• Domain‑Decomposed Inference: Splitting the simulation box across MPI ranks reduced per‑rank workload but introduced extra data‑exchange overhead; the net effect was modestly beneficial for DPA2 but detrimental for DPA3.

Overall, the attention‑based DPA2 proved more GPU‑friendly, delivering higher throughput while consuming less memory.

Practical Implications

• Accelerated High‑Fidelity MD: Researchers can now run ab‑initio‑quality MD for proteins and solvated systems at speeds comparable to classical force fields, opening doors to longer timescales and larger ensembles without sacrificing quantum accuracy.
• Plug‑and‑Play Workflow: Since the integration lives inside the standard GROMACS binary, existing pipelines (e.g., GROMACS‑based preprocessing, analysis, and visualization tools) require minimal changes—just a flag to enable the NNP.
• GPU‑Centric Deployments: The performance gains on A100/GH200 mean that cloud GPU instances or on‑prem HPC clusters can be leveraged for production runs, reducing the total cost of ownership versus running DFT‑based MD on CPU clusters.
• Model‑Agnostic Future: By abstracting the DL backend, developers can swap in newer deep‑potential families (e.g., transformer‑based or equivariant networks) without rewriting GROMACS code, fostering rapid adoption of emerging AI potentials.

Limitations & Future Work

• Scalability to Very Large Systems: The study focused on systems up to ~50 k atoms; scaling to millions of atoms may expose additional communication bottlenecks not captured here.
• Kernel‑Launch Overhead: Reducing the number of small kernel launches (e.g., via kernel fusion or batched inference) is a priority for further speedups.
• Model Generalization: While DPA2 and DPA3 were trained on specific chemical spaces, their transferability to exotic materials or extreme thermodynamic conditions remains to be validated.
• Multi‑GPU & Multi‑Node Optimization: Future work will explore more aggressive domain decomposition and overlapping communication/computation to fully exploit multi‑GPU clusters.

Authors

• Andong Hu
• Luca Pennati
• Stefano Markidis
• Ivy Peng

Paper Information

• arXiv ID: 2602.02234v1
• Categories: cs.DC, physics.chem-ph, physics.comp-ph
• Published: February 2, 2026
• PDF: Download PDF