Molecular dynamics simulations with OpenMM in Serverless AI

Molecular dynamics simulations help you study how atoms and molecules move over time. They are commonly used to explore protein stability, conformational changes and ligand binding. You can run these simulations on Nebius AI Cloud using OpenMM, a Python-based simulation engine with GPU acceleration. By packaging OpenMM in a Docker image and running it as a Serverless AI job, you only pay for the seconds the GPU is in use and avoid keeping infrastructure running between simulations. In this tutorial, you will run a protein simulation and inspect its results. The tutorial walks you through the following steps:

Build the simulation image locally with Docker and push it to Container Registry.
Configure an Object Storage bucket for storing simulation results.
Run a CUDA®-accelerated protein simulation as a Serverless AI job that reads a protein structure and writes the simulation results to the Object Storage bucket.

Costs

Nebius AI Cloud charges you for the following billing items:

Compute virtual machines (VMs)
Boot disks attached to the VMs
Used space in Standard storage in an Object Storage bucket

Container Registry is provided free of charge.

Prerequisites

Web console
CLI

Make sure you are in a group that has at least the editor role within your tenant or project; for example, the default editors group. You can check this in the Administration → IAM section of the web console.
Install Docker on your local machine. To do so, you can install Docker Desktop. If your local machine uses a different CPU architecture than AMD64 — for example, an Apple Silicon Mac or a Windows-on-ARM device — Docker Desktop builds the AMD64 image under emulation.
Install the Nebius AI Cloud CLI. It is required to authenticate against Container Registry and to push the Docker image to it.

Make sure you are in a group that has at least the editor role within your tenant or project; for example, the default editors group. You can check this in the Administration → IAM section of the web console.
Install and configure the Nebius AI Cloud CLI and the AWS CLI. Here are the instructions, depending on what tools you have already installed:
- How to install both CLIs
- How to install the Nebius AI Cloud CLI only
- How to install and configure the AWS CLI only
When configuring the AWS CLI, make sure to create a service account and its access key. The Serverless AI job uses these credentials to write results to Object Storage.
Install Docker on your local machine. To do so, you can install Docker Desktop. If your local machine uses a different CPU architecture than AMD64 — for example, an Apple Silicon Mac or a Windows-on-ARM device — Docker Desktop builds the AMD64 image under emulation.

Steps

Prepare infrastructure

Locate all resources in the same project.

Create an Object Storage bucket to store simulation results. Serverless AI deletes the container disk after the job completes. The bucket preserves the simulation artifacts generated by the job: the processed structure, trajectory, logs, metadata and plots.
- Web console
- CLI
1. In the web console, go to Storage → Object Storage.
2. Click Create bucket.
3. Specify the openmm-simulation name for the bucket.
4. In the Bucket size field, select Unlimited.
5. Click Create bucket.
nebius storage bucket create --name openmm-simulation
In Container Registry, create a registry to host the Docker image with OpenMM. Hosting the image in Container Registry keeps image pulls inside the same region as your Serverless AI job, which reduces the image pull latency and improves the job start time. Serverless AI doesn’t bill for start time, so you can use Docker Hub or configure a private registry if the added pull latency doesn’t matter for your workflow.
- Web console
- CLI
1. In the web console, go to Storage → Container Registry.
2. Click Create registry.
3. Specify the openmm name for the registry.
4. Click Create registry.
Create the registry:
nebius registry create --name openmm

Build and push the image

Configure the Docker credential helper for Container Registry so that docker push can authenticate through the Nebius AI Cloud CLI:
```
nebius registry configure-helper
```
Clone the Serverless AI Cookbook repository and switch to the openmm-simulation directory:
```
git clone https://github.com/nebius/serverless-ai-cookbook.git && \
cd serverless-ai-cookbook/life-science/openmm-simulation
```
The sim/ folder contains Python code with the OpenMM simulation settings. This tutorial uses the code as-is and lets you adjust the target protein and the simulation step count when you create the Serverless AI job.
You can explore the code and change the simulation settings — for example, by editing the OpenMM classes in sim/utils.py and sim/run.py. For more details, see OpenMM Python API.

Export the region and the registry path to environment variables:

export NB_REGION_ID=<region_ID>
export NB_REGISTRY_PATH=$(nebius registry list \
  --format json \
  | jq -r '.items[] | select(.metadata.name=="openmm") | .metadata.id' \
  | cut -d- -f2)

For the list of regions, see Regions.

Verify that Docker is available and the daemon is running:
```
docker version
docker ps
```
If docker ps returns a table of containers (can be empty), the daemon is running. If the daemon is not running, launch it.
Build the image, targeting linux/amd64:
```
docker build --platform linux/amd64 \
  -t cr.$NB_REGION_ID.nebius.cloud/$NB_REGISTRY_PATH/openmm-serverless:v1 .
```
Serverless AI runs Docker images on the linux/amd64 architecture. Using the --platform linux/amd64 parameter produces an AMD64 image regardless of the host architecture of your local machine. The resulting image runs natively in Serverless AI with no runtime penalty. For more information, see Docker official documentation on multi-platform builds.

Push the image to the registry:

docker push cr.$NB_REGION_ID.nebius.cloud/$NB_REGISTRY_PATH/openmm-serverless:v1

Run the simulation as a Serverless AI job

Create a CUDA-accelerated job that runs a short simulation on the 1UBQ protein structure:

Web console
CLI

In the web console, go to Storage → Container Registry, then open the openmm registry and the openmm-serverless image.
In the v1 tag row, click → Deploy as job. The Create job page opens with Image path already filled in.
Specify the remaining job parameters:
- Name: openmm-1ubq.
- Entrypoint command: bash -c "python -m sim.run --protein-id 1UBQ --steps 1000 && cp -r /openmm/results/. /mnt/data/".
- Job timeout in hours: 4.
- Computing resources: With GPU.
- Available platform: NVIDIA® L40S PCIe with Intel Ice Lake.
- Preset: 1 GPU — 8 CPUs — 32 GiB RAM.
- Mount volumes: Bucket.
- Mount path: /mnt/data. After that, click Attach bucket and then select the openmm-simulation bucket.
In the entrypoint command, /openmm/results/ is the location of simulation results inside the Docker container. The command runs the simulation and copies the results to the mounted bucket at /mnt/data. You can adjust --protein-id and --steps to target a different protein or change the simulation length.
Click Create.
Follow the job logs to watch the simulation progress:
1. In the sidebar, go to AI Services → Jobs.
2. Next to the job, click View logs. Alternatively, select the job that you want to view the logs for and switch to the Logs tab.

Create the environment variables with your AWS CLI credentials and the Object Storage bucket settings. The Serverless AI job reads these and uploads the results to the bucket through the AWS API:

export AWS_ACCESS_KEY_ID=$(aws configure get aws_access_key_id)
export AWS_SECRET_ACCESS_KEY=$(aws configure get aws_secret_access_key)
export AWS_DEFAULT_REGION=$NB_REGION_ID
export S3_ENDPOINT_URL=https://storage.${NB_REGION_ID}.nebius.cloud
export S3_BUCKET=openmm-simulation
export S3_PREFIX=openmm

Set the subnet ID where the job will run:
```
export NB_SUBNET_ID=<subnet_ID>
```
See How to get a subnet ID.

Create the job.

nebius ai job create \
  --name openmm-1ubq \
  --image cr.$NB_REGION_ID.nebius.cloud/$NB_REGISTRY_PATH/openmm-serverless:v1 \
  --platform gpu-l40s-a \
  --preset 1gpu-8vcpu-32gb \
  --timeout 4h \
  --args "--protein-id 1UBQ --steps 1000" \
  --env "AWS_ACCESS_KEY_ID=$AWS_ACCESS_KEY_ID" \
  --env "AWS_SECRET_ACCESS_KEY=$AWS_SECRET_ACCESS_KEY" \
  --env "AWS_DEFAULT_REGION=$AWS_DEFAULT_REGION" \
  --env "S3_ENDPOINT_URL=$S3_ENDPOINT_URL" \
  --env "S3_BUCKET=$S3_BUCKET" \
  --env "S3_PREFIX=$S3_PREFIX" \
  --subnet-id $NB_SUBNET_ID

You can reuse this command to start several simulation jobs in parallel: for each job, change the --name and --args values — for example, --protein-id 2PTC --steps 2000.

Follow the job logs to watch the simulation progress:
```
nebius ai job logs <job_ID> --follow
```

A healthy CUDA-accelerated job prints the following line near the start of the logs:

Using OpenMM platform: CUDA

If the line shows CPU or Reference instead, the GPU runtime didn’t initialize and the simulation will be far slower than expected.

View results in the bucket

Web console
CLI

The job uploads the simulation results to the bucket in the folder named <protein_id>_<timestamp>/ — for example, 1UBQ_20260512_192514/.To view results:

In the web console, go to Storage → Object Storage.
Open the openmm-simulation bucket.
Open the output folder of the latest job.

The job uploads the simulation results to the bucket through the AWS S3-compatible API at s3://<bucket>/<prefix>/<protein_id>_<timestamp>/ — for example, s3://openmm-simulation/openmm/1UBQ_20260512_192514/.To view results:

List the job outputs in the bucket:

aws s3 ls "s3://$S3_BUCKET/$S3_PREFIX/" \
  --endpoint-url "$S3_ENDPOINT_URL"

Download a specific job output locally:

aws s3 sync "s3://$S3_BUCKET/$S3_PREFIX/<output>/" "./results/<output>/" \
  --endpoint-url "$S3_ENDPOINT_URL"

A completed simulation contains the following artifacts:

<protein>_processed.pdb — the processed input structure.
<protein>_trajectory.dcd — the trajectory of the simulation.
<protein>_simulation.log — the simulation log.
<protein>_metadata.txt — the simulation metadata.
plots/ — visualizations generated from the trajectory.

To inspect the simulation locally, open the downloaded <protein_id>_processed.pdb and <protein_id>_trajectory.dcd files in a molecular visualization tool of your choice, for example, PyMOL.

PyMOL rendering of the 1UBQ protein structure

How to delete the created resources

A Serverless AI job stops billing the moment it completes, but the bucket keeps accruing charges until you remove it. The container registry is free of charge; delete it if you no longer need the image. The service account and its access key live at the tenant level, don’t accrue charges and can be reused for future workloads. If you don’t need the bucket, delete it, so Nebius AI Cloud doesn’t charge for it.

​Costs

​Prerequisites

​Steps

​Prepare infrastructure

​Build and push the image

​Run the simulation as a Serverless AI job

​View results in the bucket

​How to delete the created resources

Costs

Prerequisites

Steps

Prepare infrastructure

Build and push the image

Run the simulation as a Serverless AI job

View results in the bucket

How to delete the created resources