NIM for ALCHEMI Batched Geometry Relaxation

Type: Microservice Tags: NVIDIA, NIM, ALCHEMI, batched geometry relaxation, BGR, molecular simulation, materials science, MLIP, MACE, AIMNet2, TensorNet Related: NVIDIA-NIM, NIM-for-ALCHEMI-Batched-Molecular-Dynamics, NVIDIA-BioNeMo, NVIDIA-Clara, NVIDIA-AI-Enterprise, cuEquivariance, TensorRT, Triton-Inference-Server, NVIDIA-CUDA, NGC Sources: https://docs.nvidia.com/nim/alchemi/alchemi-bgr/latest/overview.html Last Updated: 2026-04-29

Summary

NVIDIA NIM for ALCHEMI Batched Geometry Relaxation (BGR) is a NIM microservice for high-throughput atomistic structure relaxation in computational chemistry and materials science. Current NVIDIA docs describe it as using machine learning interatomic potentials (MLIPs) such as MACE, AIMNet2, and TensorNet models optimized for NVIDIA GPUs.

Detail

Purpose

Batched geometry relaxation optimizes atom positions and, optionally, unit-cell parameters so candidate molecular or materials structures can be evaluated before downstream simulation, screening, or discovery workflows.

Current scope

  • Supports periodic materials and isolated molecules.
  • Provides optional cell optimization and DFT-D3(BJ) dispersion corrections.
  • Uses dynamic batching to estimate and adjust batch size based on available GPU memory and structure size.
  • Implements the FIRE2 optimizer on GPU for accelerated relaxation.
  • Allows per-request force and pressure convergence criteria.
  • Supports MLIP model types selected through ALCHEMI_NIM_MODEL_TYPE, including MACE, TensorNet, and AIMNet2.
  • Includes MACE-MPA-0 as a bundled/default model path when the container runs with an NGC API key.

NVIDIA context

BGR is the relaxation step in NVIDIA’s ALCHEMI atomistic-modeling NIM pair. It naturally connects with NIM-for-ALCHEMI-Batched-Molecular-Dynamics for simulation after relaxation and with cuEquivariance because MACE-style MLIP models use geometry-aware neural network operations.

Connections

Source Excerpts

  • NVIDIA docs describe BGR as a high-performance engine for batched geometry relaxation across periodic materials and isolated molecules.
  • The current docs list MACE, AIMNet2, and TensorNet as supported MLIP model families.

Resources