Is this page helpful?

Release Notes#

Release 1.1.0#

Summary#

This maintenance release includes software updates, minor bug fixes, and documentation improvements for NVIDIA NIM for Batched Molecular Dynamics.

Updates#

  • Model loading and configuration: Updated support for MACE-MP, AIMNet2 v2-format checkpoints, and TensorNet 2025.2 models.

  • Periodic boundary conditions: Improved periodic and non-periodic configuration handling, including AIMNet2 operation in both modes.

  • Trajectory output: Added atomic masses to molecular dynamics trajectory snapshots.

  • DFT-D3 configuration: Corrected configuration handling and public environment variable names.

  • Validation and readiness: Improved request validation and service readiness behavior.

  • Documentation: Corrected and streamlined model setup, API, trajectory-analysis, and example documentation.

Release 1.0.0#

Summary#

This is the first release of NVIDIA NIM for Batched Molecular Dynamics.

Features#

  • Multiple MLIP Models: Support for MACE, AIMNet2 (including NSE variant), and TensorNet models.

  • Multiple Ensembles: NVE (microcanonical), NVT (Langevin thermostat), and NPT (Monte Carlo barostat) ensembles.

  • Trajectory Output: Returns complete trajectory with positions, velocities, energies, and optionally stress tensors.

  • Restart Capability: Seamless restart from any saved state using the istep and md_time parameters.

  • Electric Field Support: External electric field application for AIMNet2 models.

  • Flexible Barostat: Isotropic and anisotropic cell shape changes for NPT.

  • Configurable Output: Adjustable save_interval for trajectory snapshot frequency.

Limitations#

  • The MACE-MPA-0 model downloads automatically from Hugging Face when the container starts. You must provide AIMNet2 and TensorNet models externally using a volume mount.

  • Periodic boundary conditions (PBC) mode is fixed at container startup. You cannot mix periodic and non-periodic structures in the same container instance.

  • Each request processes a single structure. For high-throughput screening of many structures, use the Batched Geometry Relaxation (BGR) NIM.

  • Very large structures can exceed GPU memory limits. Use the /v1/status endpoint to check max_system_size for your configuration.

  • Electric field (efield) is only supported for AIMNet2 models.