GROMACS (with ML)

GROMACS 2026: NNP-Interface & Hybrid ML Dynamics Review

The GROMACS 2026 release cycle marks a transition from experimental offloading to a production-ready GMX_ML NNP-interface. This framework allows for the direct embedding of Neural Network Potentials (NNP) into the MD integration step, supporting architectures such as DeepPot-SE, Allegro, and MACE 📑. The implementation leverages the LibTorch and TensorFlow C++ APIs to treat ML models as native force providers 🧠.

Integration & Active Learning Workflows

GROMACS has standardized the Active Learning (AL) loop, particularly through tight integration with the DeepMD-kit ecosystem. This enables closed-loop model refinement where predictive deviations trigger autonomous data collection and retraining cycles 📑.

• Hybrid Force Mixing: Supports the concurrent application of classical force fields and ML potentials (ML/MM), facilitating multi-scale modeling with energy conservation guarantees 📑.
• PIMD Support: Path Integral Molecular Dynamics is now accelerated via ML potentials, allowing for the inclusion of nuclear quantum effects at a fraction of the traditional cost 📑.
• Inference Latency: Benchmarks on NVIDIA H100/B200 hardware demonstrate that the GMX_ML interface adds less than 5% overhead to the total step time for optimized tensor models 📑.

⠠⠉⠗⠑⠁⠞⠑⠙⠀⠃⠽⠀⠠⠁⠊⠞⠕⠉⠕⠗⠑⠲⠉⠕⠍

Numerical Integrity & Performance Scaling

The GROMACS 2026 core maintains strict adherence to physical constraints while scaling across thousands of GPU nodes using a unified MPI/OpenMP domain decomposition strategy 📑.

• Virial Stress Accuracy: Accurate calculation of the virial tensor within the NNP-interface enables stable NPT ensemble simulations, though accuracy remains dependent on the model's derivative quality 🧠.
• CUDA Graph Optimization: Implementation of CUDA Graphs for ML inference calls reduces CPU-side launch overhead, a critical factor for small-to-medium scale systems 🧠.

Evaluation Guidance

Technical evaluators should prioritize the validation of the Virial stress tensor accuracy, as this is the primary failure point for ML-driven NPT simulations. It is recommended to enable CUDA Graph optimizations to mitigate kernel launch latency. For long-duration trajectories, monitoring of energy drift is essential to verify the numerical stability of the specific NNP architecture deployed 🌑.