GPU Sampling MetaD Package

• GPU Sampling MetaD (GSM) is a fully GPU-accelerated toolkit for enhanced sampling in MD, integrating metadynamics with machine learning potentials for high-precision simulations.
• It employs an entirely GPU-resident workflow with GPUMD and PyTorch, eliminating CPU-GPU data transfer bottlenecks and scaling effectively to millions of atoms.
• Benchmarking shows an order-of-magnitude speedup over CPU-based methods, with demonstrated applications in peptide torsion simulations, surface catalysis, and GaN phase transitions.

The GPU Sampling MetaD (GSM) package is a fully GPU-accelerated computational toolkit for enhanced sampling in molecular dynamics (MD), integrating metadynamics (MetaD) algorithms with machine learning potentials (MLPs) for high-precision, large-scale simulations on single or distributed GPUs. Designed to circumvent the limitations of CPU-based enhanced sampling and leverage complete GPU parallelism, GSM provides an efficient, modular, and extensible framework for simulating rare events in atomic systems comprising up to millions of atoms, demonstrated through multiple classical applications and comprehensive benchmarking against mainstream tools (Zhang et al., 8 Oct 2025).

1. Architecture and Workflow Integration

GSM operates through a complete GPU-resident pipeline, eliminating the bottlenecks of data transfer between CPU and GPU. The package is tightly coupled to GPUMD (GPU Molecular Dynamics code) and utilizes advanced MLPs to define potential energy surfaces, enabling high-accuracy force computations over very large molecular systems. The end-user interface is implemented in Python via PyTorch, which provides both automatic differentiation and GPU scheduling. Users write “GSM scripts” that specify collective variables (CVs), biasing protocols, and simulation parameters. These scripts are compiled into TorchScript representations before entering the GPUMD workflow. Within GPUMD, GSM uses an interface to automatically parse simulation data, convert standard MD runs into MetaD simulations, and manage runtime modification of the simulation loop. An accompanying analysis module (“GPU-Sampling Trajectory Analysis”) supports post hoc free-energy reconstruction and trajectory diagnostics.

2. Metadynamics Methods and Bias Application

The underlying enhanced sampling paradigm is metadynamics. At each iteration $t$, a bias potential $V_B$ is deposited along user-defined CVs $s_t$, modifying the total interaction potential:

$U_p(s_t) = U_g(s_t) + V_B(s; t)$

where $U_g(s_t)$ is the physical interaction potential (generally parameterized by MLPs). Sampling evolves according to the biased dynamics. To reconstruct the unbiased free-energy surface (FES), GSM employs reweighting:

$$F(s) = -\log \left\langle \exp{\left[\beta \left(V_B(s) - V_B(s(\mathbf{R}))\right)\right]} \right\rangle_{\text{biased}}$$

Here, the angular brackets denote averages over the trajectory generated under biased dynamics. Bias potentials are managed in a modular class structure (MultiCVMetad), with user-overridable methods controlling CV computation and bias update schedules.

3. Machine Learning Potentials and GPU Optimization

MLPs are central to GSM’s capacity for high-accuracy force fields. They replace traditional empirical force fields or ab initio methods by providing neural network-based models for interaction energies and forces. These models require intensive floating-point operations, which GSM executes entirely on the GPU, benefiting from hundreds of Tensor Cores and high memory bandwidth. The PyTorch framework allows direct construction and evaluation of CVs and bias gradients in GPU memory space with support for mixed precision and vectorized operations.

GSM’s GPU kernel scheduling ensures that both MD propagation (integration of atomic trajectories) and bias depositions are performed without leaving GPU memory. By retaining the entire simulation state—including atomic positions, velocities, bias grids, and CV trajectories—on the GPU, GSM achieves throughput comparable to pure GPUMD and avoids the latency introduced by frequent CPU-GPU synchronizations typical in CPU-based MetaD tools.

4. Performance and Benchmarking

Empirical tests reveal that GSM achieves more than a tenfold speedup relative to established CPU-GPU hybrid approaches such as Plumed integrated with LAMMPS. Across system sizes from ${\sim}10^4$ to ${\sim}10^6$ atoms, the simulation maintains near-native GPUMD efficiency even when using expensive MLPs. Consumer GPUs (e.g., NVIDIA RTX4090) suffice for these scales, indicating accessibility for typical academic and industrial users. This efficiency critically extends the accessible timescales and system sizes for MetaD and rare event sampling protocols.

5. Representative Applications

GSM’s versatility is demonstrated through detailed simulations:

• Alanine Dipeptide Torsion: A 20 ns well-tempered metadynamics simulation, with CVs defined by two dihedral angles, recovers the FES with metastable basins corresponding to distinct peptide conformations (right-handed and left-handed α-helix, β-sheet). The reconstructed surface matches known literature, validating both bias force accuracy and statistical integrity.
• Water Dissociation on Rutile (110): By defining a CV as the O–H bond distance, GSM reconstructs the free-energy profile for surface-catalyzed water dissociation and computes an adsorption barrier (~21.48 kJ/mol) consistent with prior first-principles estimates, reflecting quasi-DFT precision enabled by MLPs.
• B4–B1 Phase Transition in GaN: GSM uses multiple system sizes (27,000 up to 2.2 million atoms) to capture multi-site nucleation and solid-state phase transition kinetics in gallium nitride, overcoming finite-size artifacts and elucidating the atomic mechanism of phase change in realistic device-scale crystals.

6. Extensibility and Future Directions

Planned extensions highlighted in (Zhang et al., 8 Oct 2025) include introducing more advanced machine-learning-based CV definitions, optimizing bias update schedules, and further streamlining workflows—particularly for dataset organization and use of pretrained MLPs. GSM’s architecture supports rapid uptake of improved MLP architectures or GPU acceleration frameworks, promising further gains in accuracy and scalability. Its flexible Python-based scripting reduces setup complexity, enabling wider adoption for challenging domains such as long-chain protein folding, complex surface catalysis, and large crystalline systems.

A plausible implication is that, as GSM algorithms and PyTorch-native CVs evolve, integration with real-time analysis, active learning, or adaptive sampling protocols may emerge, given the foundational support for modular, GPU-resident pipelines and bias-driven rare event exploration.

7. Summary and Significance

The GSM package advances the state of GPU-accelerated metadynamics sampling, enabling high-precision, fully GPU-resident workflows for rare event MD simulations. By leveraging MLPs, efficient bias force calculation, and a PyTorch-centric interface, GSM achieves order-of-magnitude speedups over previous approaches and scales to millions of atoms on commodity hardware. Its deployment in classical MD applications demonstrates robust quantitative accuracy and comprehensive sampling capability. The open, scriptable nature and planned future enhancements suggest continued expansion of the GSM framework’s capabilities within both molecular and device-level simulation domains (Zhang et al., 8 Oct 2025).