How DTN accelerates operational weather prediction using NVIDIA Earth-2 on AWS

This post was contributed by Satheesh Maheswaran, Samuel Lillo, Stefan Weissenberger, Tim Brown, Taylor Gowan, Kyle Keys, Colin Bridger and Doug Chenevert.

Weather prediction is increasingly critical for business operations, with impacts on safety, efficiency and profitability.  It’s pivotal across many industries: agriculture, aviation, energy, and maritime operations, which rely on weather forecasts for events that directly affect their operations.  While traditional physics-based models have served well for atmospheric simulations, artificial intelligence (AI) now presents an opportunity to transform our capabilities.

In this post we’ll take you through the process of integrating an innovative AI weather forecasting model created in collaboration with DTN and NVIDIA for cyclone tracking in a production environment. DTN successfully integrated this model into their production environment in June 2025. This new approach delivers multiple benefits: more accuracy and reliable forecasts; improved operational margins; reduced business risk; and greater opportunities for organizational growth.

DTN is a global data and technology company helping leaders in energy, agriculture, and weather to make faster, smarter decisions to boost margins, growth, and resilience. This includes DTN harnessing artificial intelligence to create decision-grade weather solutions.  These solutions support critical business decisions, especially during high-impact weather events.  In addition to more than 25 public, private and proprietary weather models, DTN is using NVIDIA Earth-2 on AWS to build scalable, resilient and secure infrastructure.  In 2025, DTN developed a patent-pending AI model that predicts cyclone tracks as an ensemble to give customers faster decisioning at scale for weather impacts and risk assessments.

NVIDIA Earth-2

NVIDIA Earth-2 is an enterprise platform that helps teams understand the impacts of weather and climate. It provides a complete software stack for accelerating traditional physics-based weather forecasts, developing and running AI weather models, and creating visualizations of weather and climate data so that organizations can run forecasts faster, at higher resolution, and turn them into interactive products.

Two NVIDIA open-source Python libraries offer a comprehensive toolbox for developing AI-powered solutions: NVIDIA PhysicsNeMo and Earth2Studio. NVIDIA PhysicsNeMo is used to train AI physics models. It comprises optimized large scale training pipelines and architectures for several leading AI weather models. Earth2Studio provides the building blocks to develop and deploy AI weather forecasting pipelines, including pre-trained models and easy integration to live weather conditions for initializing predictions.

Earth-2 facilitates the development of models for many different applications, including forecasting global and regional weather, increasing the resolution of forecasts, and downstream impact modeling. Once trained, these AI models can generate forecasts orders of magnitude faster than traditional physics-based forecasts. This means we can use Earth-2 to develop or augment solutions to better understand extreme weather by rapidly generating thousands of predictions.

Typical inference workloads can require as little as a single GPU  making it practical to run multiple inferences within minutes, rather than relying on large clusters to produce ensembles of traditional numerical weather predictions.

DTN used Earth2Studio to deploy FourCastNet, a global AI weather model, on AWS. They can now run a medium-range 25-km resolution weather forecast in seconds on an NVIDIA GPU like the ones provided by the Amazon Elastic Compute Cloud (Amazon EC2) G6e, P5, or P6 instance families. Earth2Studio’s modular workflows enable solution developers to experiment with and evaluate models of their choosing. For instance, we can validate the same inference pipeline using different AI models. We can extend inference pipelines with downstream models like the CorrDiff AI super-resolution model to provide higher-resolution regional forecasts.

Figure 1 Example weather forecast pipeline built with Earth2Studio.  The current weather conditions are received from an online repository, for example from the Registry of Open Data on AWS.  Two AI models are combined to create a high-resolution weather forecast.  The resulting data can then be analyzed for business impact.

NVIDIA Earth-2 is designed to enable development of solutions, that derive AI-powered weather and climate insights. This enables organizations to develop solutions for building resilience against extreme weather, which is becoming more frequent and intense due to climate change.

How to deploy a solution architecture for production

DTN’s production deployment of NVIDIA Earth-2 on AWS uses a cloud-native architecture designed for scale, reliability, and operational efficiency. This solution combines containerized workloads with orchestration to deliver consistent ensemble forecasts regardless of how volatile the weather is.

Figure 2 shows an example architecture for deploying an Earth-2 AI workflow.  At the foundation of our architecture are two custom Docker containers: a GPU-optimized Earth2Studio inference container built on NVIDIA PhysicsNeMo image, and a lightweight Python utilities container for data preparation and results aggregation.  We store these containers in Amaon Elastic Container Registry (Amazon ECR), to ensure consistent deployment across the production environment.

Figure 2 AWS architecture for NVIDIA Earth-2 weather prediction, orchestrating containerized CPU/GPU workloads through step functions, AWS Batch and Lambda with data flowing through Amazon S3

The workflow orchestration is handled by AWS Step Functions, which coordinates a three-phase process:

1. Data Preparation (Phase 1): AWS Lambda functions trigger the Python utilities container to generate processing dates and format initial condition data
2. Weather Model Inference (Phase 2):  AWS Batch deploys the Earth2Studio container across GPU-enabled EC2 instances to execute the FourCastNet model inference, generating ensemble members for each forecast
3. Results Aggregation (Phase 3): The Python utilities container processes ensemble outputs, calculating statistical measures like standard deviations to quantify forecast uncertainty

AWS Batch manages the compute resources, intelligently scheduling Earth2Studio jobs across a fleet of GPU-enabled instances. The implementation supports both single-date processing for targeted forecasts and batch processing for operational runs across extended time periods.

Amazon Simple Storage Service (Amazon S3) serves as the central data repository throughout the pipeline, with Amazon Simple Notification Service (Amazon SNS) notifications integrated at key completion points to trigger downstream processes and alerts. Our solution includes custom Amazon S3 mounting using s3fs within containers for efficient data access, reducing transfer times for the large meteorological datasets.

For system resilience, we implemented comprehensive error handling using the Step Functions retry mechanisms and dead-letter queues. The workflow downloads the input observational data only once and makes it available as cache directories for Earth2 models across multiple job runs to optimize performance.

We defined the entire infrastructure as code using AWS Cloud Development Kit (AWS CDK), enabling consistent deployment across environments and helping disaster recovery scenarios. This infrastructure-as-code approach also supports a continuous improvement process, allowing rapid incorporation of refinements to the Earth-2 models as NVIDIA releases updates.

We’ve made the complete reference implementation of this architecture available as open-source code. This provides the flexibility to scale during extreme weather events when forecast demand peaks while keeping operational reliability throughout the year.

How DTN implemented the solution in a production environment

The implementation required careful expansion of existing high-performance computing infrastructure to accommodate Earth2Studio. That meant configuring real-time boundary conditions, optimizing job scheduling, and establishing comprehensive monitoring and alerting systems. With help from AWS and NVIDIA, DTN leveraged their operational expertise and completed the entire implementation process in approximately four months from initial kick-off to production deployment.

To validate system performance, DTN constructed rigorous test cases around significant historical weather events, including Hurricanes Milton and Helene, and Storm Eowyn in the North Atlantic. In Figure 3 we show hurricane tracks for Milton and Helene.  We also include the ensemble tracks for hurricane Lee and Helene, demonstrating the value in ensemble forecasts using a globally trained model. The results are all inference-based on the trained model, initialized with a collection of initial conditions to provide the ensemble forecasts.

Figure 3 Hurricane tracks for Milton and Helene.  Ensemble tracks for hurricanes Lee and Helene are also shown.  The results show good agreement with BestTrack data from IB tracks.

Production results and business impact

The enhanced forecast system is one way that DTN keeps its commitment to continuous innovation for its customers. By integrating critical business data and advanced forecasting capabilities, the system enables customers to accelerate growth while improving operational margins and minimizing risk exposure.

According to Renny Vandewege, General Manager of Weather and Climate Intelligence at DTN, “Our innovative techniques go beyond traditional weather intelligence. By integrating critical business data, market insights, and supply chain information into a comprehensive decision support platform, we empower our customers to make smarter, more informed operational decisions. Our customizable dashboards and automated tools leverage our Decision-Grade Data and transform that data into actionable intelligence, tailored to each industry’s unique needs.”

Future roadmap and conclusions

The integration of AI weather modelling into the DTN’s Forecast system enables continuous learning and iteration to maximize model strengths, and a comprehensive validation and verification structure monitors model outputs and leverages data into DTN decision support systems. This now supports additional AI-based modelling insights, including forecast ensembles for confidence metrics and extended time horizons from hourly to sub-seasonal scales.

The rapid evolution of AI in weather forecasting, documented in both scientific literature and mainstream media, positions DTN to leverage emerging capabilities to support their global community.  The collaboration between DTN, AWS, and NVIDIA has demonstrated progress in applying AI forecasts for excellent production-grade outcomes. NVIDIA Earth-2 enables DTN to stay on the bleeding edge of AI innovation by providing the flexibility to explore state of the art models from the AI community while also expanding the use of AI for various applications for their diverse customers.