> This page is for NVIDIA NIM for Batched Molecular Dynamics, version 1.1.0 (Latest) (default).
> For other versions, use one of these documentation indexes:
> - 1.1.0 (Latest) (default): https://docs.nvidia.com/nim/alchemi/alchemi-bmd/1.1.0/llms.txt
> - 1.0.0: https://docs.nvidia.com/nim/alchemi/alchemi-bmd/1.0.0/llms.txt

> For clean Markdown of any page, append .md to the page URL.
> For a complete documentation index, see https://docs.nvidia.com/nim/alchemi/llms.txt.
> For AI client integration (Claude Code, Cursor, etc.), connect to the MCP server at https://docs.nvidia.com/nim/alchemi/_mcp/server.

# Overview

The NVIDIA NIM for Batched Molecular Dynamics provides a high-performance engine for batched molecular
dynamics (MD) simulations, enabling researchers to study the time evolution of atomic
systems using machine learning interatomic potentials (MLIPs). Through supporting multiple
MLIP models including MACE, AIMNet2, and TensorNet optimized for NVIDIA GPUs, this NIM
enables researchers to perform MD simulations at scale with near-Density Functional Theory (DFT) accuracy at a
fraction of the computational cost.

Molecular dynamics is fundamental to understanding material properties, chemical reactions,
protein dynamics, and many other phenomena in computational chemistry and materials science.

The NIM is primarily recommended for the following use cases:

- **High-throughput MD simulations:** Ideal for studying thermodynamic properties,
phase transitions, and transport phenomena across many systems.
- **Materials science research:** Investigate thermal conductivity, diffusion coefficients,
melting points, and other temperature-dependent properties.
- **Drug discovery workflows:** Study ligand binding, conformational sampling, and
free energy calculations with GPU-accelerated MD.
- **Reaction dynamics:** Explore reactive systems and transition states with accurate
MLIP-driven dynamics.

## Key Features of NIM for BMD

The NIM for BMD includes the following key features:

- **Multiple MLIP Models:** Supports various machine learning interatomic potentials including
MACE, AIMNet2 (with NSE variant for organic molecular systems), and TensorNet models, allowing
users to choose the best model for their specific application.
- **Dynamic Batching:** Optimizes GPU utilization by dynamically batching concurrent
MD requests based on available GPU memory, maximizing throughput and efficiency.
- **Multiple Ensembles:** Supports NVE (microcanonical), NVT (canonical with Langevin thermostat),
and NPT (isothermal-isobaric with Monte Carlo barostat) ensembles for diverse simulation needs.
- **Trajectory Output:** Returns complete trajectory snapshots including positions, velocities,
energies, and optionally stress tensors for post-processing and analysis.
- **Restart Capability:** Simulations can be seamlessly restarted from any saved state,
enabling long simulations to be run in chunks and facilitating checkpoint and restart workflows.
- **Electric Field Support:** For AIMNet2 models, external electric fields can be applied
to study field-dependent phenomena.
- **Flexible Configuration:** Per-request simulation parameters allow fine-grained control
over temperature, timestep, ensemble, and barostat settings.

## Supported Ensembles

The NIM supports the following simulation ensembles:

| Ensemble | Settings | Description |
| :------- | :------- | :---------- |
| NVE | `nvt=False`, `npt=False` | Microcanonical - constant energy |
| NVT | `nvt=True`, `npt=False` | Canonical - Langevin thermostat |
| NPT | `nvt=True`, `npt=True` | Isothermal-isobaric - MC barostat |

## Supported Models

**Default model:** At startup, the container downloads the MACE-MPA-0 medium
checkpoint from the
[MACE-MPA-0 repository on Hugging Face](https://huggingface.co/mace-foundations/mace-mpa-0).
No manual model download or mount is required for the default launch flow.

NIM for BMD supports multiple MLIP model architectures. The
`ALCHEMI_NIM_MODEL_TYPE` environment variable selects the architecture; the
specific model variant is determined by the mounted checkpoint, or by the
default MACE-MPA-0 checkpoint when no custom model is configured.

The following table lists the supported models and their capabilities:

| `ALCHEMI_NIM_MODEL_TYPE` | Supported Models | Periodic | Non-periodic | Model Provisioning | Reference |
|---|---|---|---|---|---|
| `mace` | MACE-MPA-0 (default); user-provided checkpoints from the MACE-MP model family | Yes | No | Default MACE-MPA-0 downloaded from Hugging Face at startup; other supported checkpoints are user-provided | [MACE-MPA-0](https://arxiv.org/abs/2401.00096), [MACE](https://arxiv.org/abs/2206.07697) |
| `tensornet` | TensorNet-PES-MatPES-PBE-2025.2, TensorNet-PES-MatPES-r2SCAN-2025.2 | Yes | No | User-provided | [MatPES](https://arxiv.org/abs/2503.04070), [TensorNet](https://arxiv.org/abs/2306.06482) |
| `aimnet2` | AIMNet2 v2-format checkpoints | Yes | Yes | User-provided | [AIMNet2](https://doi.org/10.1039/D4SC08572H) |

TensorNet models are [MatPES](https://arxiv.org/abs/2503.04070) potentials
distributed by [materialyze on Hugging Face](https://huggingface.co/materialyze)
(PyG version, not DGL).

AIMNet2 supports periodic and non-periodic systems. The backend starts in
periodic mode by default (`ALCHEMI_NIM_PBC=true`). User-provided models must be
v2-format `.pt` checkpoints with `format_version` set to `2`; refer to the
[AIMNetCentral model-format guide](https://isayevlab.github.io/aimnetcentral/model_format/).

For user-provided models, refer to [Model Configuration](/nim/alchemi/alchemi-bmd/configuration#model-configuration) for mounting instructions. For download hints and container launch arguments by model type, refer to [Custom Models](/nim/alchemi/alchemi-bmd/custom-models).


## References

* [MACE architecture](https://arxiv.org/abs/2206.07697)
* [MACE-MP-0](https://arxiv.org/abs/2401.00096)
* [AIMNet2](https://doi.org/10.1039/D4SC08572H)
* [AIMNet2-NSE](https://doi.org/10.26434/chemrxiv-2025-kdg6n)
* [TensorNet architecture](https://arxiv.org/abs/2306.06482)
* [TensorNet-MatPES](https://arxiv.org/abs/2503.04070)

## Advantages of NIMs

NIMs offer a simple and easy-to-deploy route for self-hosted AI applications. Two major
advantages that NIMs offer for system administrators and developers are:

- Increased productivity — NIMs allow developers to build generative AI applications quickly,
in minutes rather than weeks, by providing a standardized way to add AI capabilities to their applications.

- Simplified deployment — NIMs provide containers that can be easily deployed on various platforms,
including clouds, data centers, or workstations, making it convenient for developers to test, deploy, and scale their applications.

The NIM for BMD provides fast, accurate models behind a consistent
API for molecular dynamics simulations. As part of the broader NVIDIA
NIM ecosystem, NIM for BMD can be used in conjunction with other
NIMs such as the BGR (Batched Geometry Relaxation) NIM to build comprehensive
atomistic modeling pipelines.