FourCastNet

What Is NVIDIA FourCastNet?

NVIDIA FourCastNet is a global, data-driven weather-forecasting model developed for NVIDIA's Earth-2 climate and weather-computing ecosystem. Instead of generating text, images, or other consumer-facing content, it predicts numerical weather fields such as wind, temperature, pressure, humidity, and geopotential height.

The model takes a representation of the current atmospheric and surface state on a global latitude-longitude grid. It then produces the predicted state six hours later. By feeding the output back into the model, users can create longer autoregressive forecasts and simulations. This makes FourCastNet useful for research and operational experimentation where many forecast scenarios are needed.

FourCastNet was originally introduced in 2022 and remains available through NVIDIA's current Earth-2 model and NIM deployment ecosystem. NVIDIA's current materials describe an SFNO-based implementation, while deployment documentation also refers to FourCastNet v2 SFNO checkpoint variants. Users should therefore check the specific model card, checkpoint, and container documentation associated with their deployment.

How the Model Works

FourCastNet uses a Spherical Fourier Neural Operator, or SFNO. A neural operator learns how an entire spatial field changes rather than treating each location as an unrelated prediction. The Fourier component represents broad spatial patterns in the frequency domain, while the spherical formulation is suited to global geographic data.

In practical terms, the model works with a very large weather-state tensor covering the Earth. Its documented input grid contains 721 latitude points and 1,440 longitude points at 0.25-degree resolution. The input includes 73 surface and atmospheric variables together with datetime metadata. The datetime is important because atmospheric conditions vary with the calendar and time of year.

• Geographic coverage: Global latitude-longitude grid.
• Grid resolution: 0.25 degrees.
• Grid dimensions: 721 latitude points by 1,440 longitude points.
• Weather variables: 73 surface and atmospheric variables.
• Input representation: NumPy tensor plus datetime information.
• Standard forecast step: Six hours.

This is fundamentally different from sending a question to a language model. FourCastNet does not interpret an informal location-and-date request by itself. A user or surrounding application must prepare compatible gridded input data, manage units and normalization, and interpret the resulting numerical fields.

Outputs and Forecast Horizon

The output is a four-dimensional numerical tensor containing predictions for the same 73 weather variables on the same global grid. One inference step represents the atmospheric state six hours after the supplied input state. The deployment can expose selected variables or complete output archives, depending on configuration.

Longer forecasts are produced by running repeated steps. This autoregressive approach makes it possible to simulate weather beyond the first six-hour interval, but forecast errors can accumulate as predicted states are reused as later inputs. NVIDIA documentation describes predictive stability for more than one year of simulated time; that statement should be understood as a provider-documented model behavior, not a guarantee that long-range forecasts will maintain short-range accuracy.

FourCastNet has no documented context window or maximum text-output-token limit because it is not a language model. Its relevant limits are the fixed grid, the documented variable set, the input-data preparation requirements, available GPU memory, and the simulation settings supported by the particular deployment.

Training Data and Primary Purpose

The NVIDIA model card identifies ERA5 reanalysis data from 1979 through 2017 as the training data and ERA5 data from 2018 as the evaluation data. ERA5 is a historical atmospheric reanalysis dataset that combines observations with weather-model calculations to provide consistent estimates of atmospheric, land, and ocean-related conditions.

FourCastNet's primary purpose is fast global weather-state prediction. Its speed is especially relevant when users need many forecast paths rather than one expensive simulation. Example applications include:

• Rapid global weather forecasting and research into AI-based numerical-weather prediction.
• Large-ensemble simulations for estimating forecast uncertainty.
• Wind and renewable-energy analysis.
• Extreme-weather and climate-risk research.
• Exploration of atmospheric dynamics and alternative forecasting methods.
• Generation of inputs for downstream scientific or operational workflows.

The model is not a complete weather-information product for ordinary end users. Turning its tensor output into a local forecast, map, alert, or business decision requires additional processing, visualization, validation, and often comparison with observations or other forecast systems.

Deployment in NVIDIA Earth-2

FourCastNet fits into NVIDIA Earth-2 as an open or downloadable scientific model that can also be served through NVIDIA's inference infrastructure. NVIDIA provides a model card, deployment materials, and a FourCastNet NIM service. NIM, or NVIDIA Inference Microservices, packages supported models for deployment on NVIDIA GPU infrastructure and provides an interface for inference.

The FourCastNet inference service supports configurable simulation length, ensemble size, and noise amplitude according to the supplied deployment research. An ensemble runs multiple related forecast paths, which can help researchers study uncertainty rather than relying on a single deterministic trajectory. The exact options and hardware requirements depend on the NIM version and deployment configuration.

This positioning gives users two broad ways to work with the model: download and integrate the model into a compatible scientific environment, or use the NVIDIA-hosted or self-managed NIM pathway where available. Neither option turns FourCastNet into a general conversational API. Its interface is intended for structured weather tensors and numerical results, not prompts, chat messages, function calls, or token streams.

Supported Inputs, Outputs, and Capabilities

FourCastNet accepts numerical weather-state data and datetime metadata. Its documented input and output are scientific arrays rather than text, images, audio, or video. The model therefore has no conventional multimodal chat capability, image understanding, speech recognition, code generation, or text generation.

The distinction between structured numerical output and a language-model JSON mode matters. FourCastNet can return machine-readable weather data through its inference interface, but it does not generate arbitrary JSON objects from natural-language instructions.

Main Strengths and Trade-offs

FourCastNet's principal strength is computational efficiency for global weather simulation. A learned model can produce forecast states quickly once the required input is prepared and the appropriate GPU environment is available. That efficiency is valuable for ensembles, sensitivity studies, and applications that need many scenarios.

Its global coverage and 0.25-degree grid also make it suitable for broad atmospheric analysis rather than only a single local station. The 73-variable output contains substantially more information than a simple temperature forecast, allowing downstream users to examine multiple aspects of the predicted state.

There are important trade-offs. FourCastNet is specialized: it cannot answer general questions, write code, summarize documents, search the web, or replace a conventional weather-data service without an application around it. It also depends on gridded inputs that have been prepared in the expected format. A user seeking a forecast for one city cannot necessarily provide only a city name and receive a result.

Repeated autoregressive steps introduce error accumulation, and model predictions should not be confused with direct observations. Hardware, software, model-version, and licensing requirements can also make deployment more involved than using a hosted consumer weather application. These limitations are particularly relevant when forecast outputs will support safety-critical or regulated decisions.

Pricing and Access

No public token-based input or output price is listed for FourCastNet. The downloadable model and NIM access are subject to NVIDIA's access, infrastructure, and licensing terms, while hosted inference or self-managed deployment costs may depend on the selected NVIDIA service and GPU environment.

Consequently, FourCastNet should not be evaluated using the pricing model of a text API. The financial cost is more likely to come from GPU infrastructure, storage, data preparation, hosting, and operational support than from a published per-token rate. Prospective users should verify the terms for the exact Earth-2 model package or NIM release they plan to use.

Speed and Cost: What the Editorial Scores Mean

The supplied model assessment assigns FourCastNet an editorial speed score of 9 out of 10 and a cost score of 8 out of 10. These are subjective evaluations of its efficiency relative to conventional numerical-weather-prediction workflows, not NVIDIA-published benchmark scores and not prices.

The scores reflect the model's suitability for producing many numerical forecast states efficiently. They do not mean that every deployment will be inexpensive or fast. Actual performance depends on the GPU, batch and ensemble settings, input pipeline, forecast horizon, storage, and whether the model is self-hosted or accessed through a managed service.

When to Choose FourCastNet

FourCastNet is a strong candidate when the task requires rapid global weather-field simulation, repeated forecast scenarios, or experimentation with learned alternatives to conventional numerical-weather models. It is particularly appropriate for researchers, climate-risk teams, renewable-energy analysts, and developers building scientific workflows who already have access to compatible weather data and NVIDIA GPU infrastructure.

Choose another option when the requirement is a simple location-based forecast, a polished consumer weather experience, a text-based explanation, or a model that accepts unstructured prompts. A conventional numerical weather-prediction system or established weather-data provider may be more appropriate when physical-modeling provenance, operational validation, or ready-to-use local products is the priority. A general-purpose language model is more suitable for documentation, coding, or conversational analysis, but it is not a substitute for FourCastNet's numerical atmospheric output.

FourCastNet is also not the best choice when the user cannot prepare the required 73-variable global input tensor or lacks the hardware and deployment environment needed to run the model. Its value comes from specialized, high-throughput weather simulation rather than broad task coverage.

Bottom Line

NVIDIA FourCastNet is a specialized global weather model built around Spherical Fourier Neural Operators. It predicts 73 atmospheric and surface variables on a 0.25-degree grid in six-hour steps and can be iterated for longer simulations or used in ensemble workflows. Its placement in NVIDIA Earth-2 and FourCastNet NIM makes it relevant to scientific and operational experimentation, especially where rapid scenario generation matters.

The model's advantages are also its boundaries. It produces numerical weather fields, not conversation or media; it requires carefully prepared gridded inputs; and its deployment cost depends on infrastructure rather than a simple public token price. For users with the right data and GPU environment, FourCastNet offers a focused way to explore fast AI-based global forecasting. For users seeking an immediately usable local forecast or general AI assistant, a different type of system will be more appropriate.