Earth-2 Correction Diffusion NIM User Guide

A guide for generating an AI weather forecast using NVIDIA Earth2Studio CorrDiff NIM 1.0.0

Important Prerequisites

In order to deploy CorrDiff NIM, you need to be added to a namespace in the NRP Nautilus hypercluster. You also need kubectl and kubelogin from Kubernetes for pulling the container from the NIM.

If you are currently not in a namespace, please contact your admin. To install kubectl and gain access to the GPU clusters, follow the instructions HERE.

NGC API Key

To acquire the NIM, we need to create an NVIDIA NGC account and generate a personal API Key

Create an account HERE.

After your account is setup, click on your profile on the top right corner, then go to Account Settings. Once you are in the Account Settings section, scroll down to find API Keys and click on Generate API Key.

In the API Keys section, you can see all the personal keys you have created. Click on Generate Personal Key and add key details. Make sure you select all the services under Key Permission before it is generated.

Copy and paste the key that only you have access to. NGC does not save your key values for you; if you ever lose your key, you can always generate a new API Key.

Caution

Your personal key should be kept private at all times.

With your API key values, go to your terminal and create a secret using kubectl for your key:

kubectl create secret docker-registry ngc-secret-<USERNAME> \
  --docker-server=nvcr.io \
  --docker-username='$oauthtoken' \
  --docker-password= 'YOUR API KEY' \
  -n sdsu-shen-climate-lab

Replace 'YOUR API KEY' with your actual key values.

Then create the ngc-api-key

kubectl create secret docker ngc-api-key-<USERNAME> \
  --from-literal=NGC_API_KEY= 'YOUR API KEY' \
  -n sdsu-shen-climate-lab

Replace 'YOUR API KEY' with your actual key values and replace with your username.

Check that your secret and api key exist:

kubectl get secrets -n <YOUR NAMESPACE>

If they are properly stored in your system, you should see:

NAME                            TYPE                             DATA   AGE
ngc-secret-<USERNAME>            kubernetes.io/dockerconfigjson   1      2m
ngc-api-key-<USERNAME>           Opaque                           1      2m

Download and edit the corrdiff-nim-deployment.yaml file by replacing with your username and with the title of your namespace HERE. Please store the file in a path where you run your Kubernetes.

---

Launching the NIM

To launch the NIM, run the following in your terminal:

kubectl create -f corrdiff-nim-deployment-<YOUR NAME>.yaml -n <YOUR NAMESPACE>

Then run:

kubectl get pods -n <YOUR NAMESPACE> -w | grep corrdiff

Tip

It takes about 5-10 mins for the Kubernetes cluster to download the images(~26GB), keep checking the live status of the container in watch mode. The container is ready when the status shows RUNNING. To check your live connection logs, run kubectl logs -f deployment/corrdiff-nim-laurahu -n sdsu-shen-climate-lab in your terminal.

---

Required Installations in JupyterHub

Install earth2studio in your kernel

pip install earth2studio

Verify installation in your Python3 kernel

import earth2studio
print(earth2studio.__version__)

Make sure your Python version is >3.10.

Install submodules NVIDIA Earth-2 Correction Diffusion in NIM to ensure that data loads safely:

pip install earth2studio[corrdiff]
pip install earth2studio[data]

---

Run CorrDiff NIM

Create a designated folder in your JupyterHub called corrdiff

Download the files named corrdiff_output_lat.npy, corrdiff_output_lon.npy from this GitHub repository, and upload them to the folder. These files tell us how the output is mapped onto CONUS.

Download CorrDiffHurrHeleVis.ipynb, this is the ipynb file that deploys NIM to generate downscaled outputs.

About the ipynb Script

The script includes API key validation and NIM health check; it is a prerequisite to ensure that the status of both checkpoints is good before running inferences. The runtime for your inference depends on your sample size and step size. Users are recommended to limit the inference runtime by adjusting the timeout.

Note that CorrDiff NIM only generates raw tensor outputs; users should handle metadata post-processing. The sample script includes channel-specific ensemble mean and other post-processing strategies, which are suitable for hurricane tracking. Visit the CorrDiff model card for more information on inputs and outputs.

---

Troubleshooting

Common issues encountered when deploying CorrDiff NIM v1.0.0 and their solutions can be found HERE.

---

Delete Deployment

To free up space for other users, one should always delete their NIM container once they are done deploying it. To delete your deployment, run the following commands in your terminal.

kubectl delete deployment corrdiff-nim-<YOUR NAME> -n <YOUR LAB NAMESPACE>
kubectl delete service corrdiff-nim-service-<YOUR NAME> -n <YOUR LAB NAMESPACE>

References

• NVIDIA NIM
• Model Card
• Earth2Studio

Note

The new corrdiff version 1.1.0 is out on 1/20/2026. This version requires an AI Enterprise subscription.