A Breakthrough Year for AI in Weather Forecasting: Insights and Opportunities
New Crossroad in Weather Prediction
Numerical weather prediction (NWP) has long relied on physics-based models to generate operational weather forecasts. In other words, the benefits of weather forecasting you have seen is the direct beneficiary of Moore’s law. Much of the number crunching for what you see on TV weather forecasting has come from increasing processor speed. To sustain advancements in weather forecasting, computation must scale to petascale levels, as exemplified by IBM's Roadrunner in 2007. However, future progress requires not just larger computational facilities but innovative approaches. Machine learning (ML) offers the most practical solution, with cutting-edge techniques from Google, Microsoft, and NVIDIA revisiting and advancing 20th-century ideas through deep learning and generative AI. This transition not only provides a new frontier of weather forecasting, but entirely new opportunities.
Fairly recent development, ML models for weather forecasting
NWP models lie at the basis of today’s weather forecasts. However, in the past few years, a series of breakthroughs in AI research has been translated to weather forecasting. The academic study of using machine learning approaches for weathering forecasting is a relatively new space. Academic papers in weather and climate modelling referencing machine learning techniques such as neural networks or decision trees started to really hit publish in 2016.
Source: https://gmd.copernicus.org/articles/16/6433/2023/
The situation changed rapidly, between February 2022 and April 2023.
In a series of papers, predominantly from large technology companies such as NVIDIA, Huawei and Google DeepMind, rapid progress was made in the quality of ML-based weather forecasts. These ML-based weather forecasts first approached the skill of the European Centre for Medium-Range Weather Forecasts (ECMWF) IFS model, then matched IFS skill, and then claimed to surpass IFS. Your local weather person relays the ECMWF to you during the nightly broadcast, which is generally considered the gold standard in weather forecasting. As discussed in this blog post, the ML models only required a single GPU, took less than a minute, and consumed a tiny fraction of the energy required for an IFS forecast. https://www.ecmwf.int/en/about/media-centre/science-blog/2023/rise-machine-learning-weather-forecasting
The GPU based compute was also orders of magnitude cheaper than traditional CPU’s.
Source: https://resources.nvidia.com/en-us-energy-efficiency/faster-weather-prediction
2024 Developments
The hits kept coming in 2024. Seminal papers published in Nature on Google's GenCast, and Neural General Circulation Model (NeuralGCM) continue to define a new frontier for climate and weather prediction. While GenCast is a major breakthrough in pure ML-based weather prediction, NeuralGCM combines physics and ML to deliver orders of magnitude improvements in compute over traditional NWP. Please see Appendix for more technical details.
These innovations not only highlight the potential of AI but also underline its relevance across industries. At CLAI Ventures, we are thrilled to invest in this rapidly evolving field, with a recent investment in Brightband, a pioneer in AI-based weather forecasting. On a side note, join us for a lively conversation with the founder Julian Green on Thursday, January 23rd 4:00 pm PT via Zoom. Click the link below to register!
Transformative Impacts of AI-Weather Innovations
By infusing traditional physics based models with AI ML tools and techniques, we have containerized what was once an unmovable amount of compute. In other words, weather forecasts at usable resolutions can be developed for entirely new use cases. Weather forecasting and prediction with significantly reduced computational demand can be built into modern software architecture. Brightband very much embodies our thesis: exploring how innovative tools can be developed and the transformative value they can unlock.
The practical implications for access to high resolution and accurate weather prediction is vast, spanning numerous industries:
Energy Trading: Accurate wind and solar forecasts drive optimized trading strategies in energy markets.
Utilities: Grid operators use precise forecasts to balance renewable energy supply with demand.
Transportation: Airlines and logistics firms leverage real-time weather data to optimize routes and schedules.
Agriculture: Farmers make data-driven decisions on irrigation, pest control, and planting schedules.
Insurance: Risk models informed by AI weather predictions enhance underwriting for natural disaster coverage.
Government and Defense: Local governments and military planners utilize forecasts for disaster preparedness and strategic operations.
At CLAI Ventures, we believe that AI-driven advancements in weather forecasting will play a transformative role in addressing the warming and more volatile climate. Our investment in Brightband reflects our commitment to technologies that leverage state-of-the-art models like GenCast and NeuralGCM. These tools are not just about better forecasts; they empower industries to make smarter decisions, mitigate risks, and adapt to an increasingly unpredictable world.
The journey of AI-weather has only just begun, and we are thrilled to support innovations at this critical intersection of technology and climate impact!
Appendix
GenCast: Revolutionizing Global Forecasting
GenCast, a probabilistic weather model developed by Google, represents a significant leap in operational weather forecasting. With the ability to generate 15-day ensemble forecasts at a 0.25° resolution, it outperforms ECMWF’s ensemble forecasts (ENS), the leading operational system, on 97.2% of evaluated metrics.
Key Features:
Extreme Weather Prediction: GenCast excels in predicting tropical cyclone tracks, wind power production, and other extreme weather phenomena that traditional models often fail to capture with high accuracy.
Efficiency: Generating a single 15-day forecast takes just 8 minutes on a Cloud TPUv5 device, making it highly efficient compared to traditional systems that require hours or even days.
Global Accuracy: Its skill in global forecasting provides significant advantages for industries like energy trading, aviation, and agriculture, where long-term precision is critical.
Technical Innovation:
Implemented as a conditional diffusion model and uses a neural network with an encoder, processor, and decoder architecture
Generates weather forecasts by iteratively refining a candidate state initialized with noise
Trained on 40 years of ERA5 reanalysis data (1979-2018)
A notable example demonstrated the model's capability by predicting Typhoon Hagibis trajectories with varying levels of uncertainty based on lead time.
GenCast outperforms traditional ECMWF’s ENS in skill, calibration, and prediction of extreme events. It generates high-resolution, realistic weather trajectories and captures spatial-temporal dependencies. While GenCast operates at a lower resolution (0.25°) compared to updated ENS (0.1°), scaling and efficiency techniques like distillation and fine-tuning with operational data could enhance its performance. The study highlights GenCast’s potential to revolutionize weather forecasting through generative AI, offering greater accuracy and accessibility for critical decision-making applications.
Source: Kochkov, D., Yuval, J., Langmore, I. et al. Neural general circulation models for weather and climate. Nature 632, 1060–1066 (2024). https://doi.org/10.1038/s41586-024-07744-y
NeuralGCM: Bridging Physics and AI
The Neural General Circulation Model (NeuralGCM) merges physics-based principles with machine learning, achieving a balance between interpretability and computational efficiency. Unlike pure ML models, NeuralGCM incorporates a differentiable dynamical core that solves governing equations for atmospheric dynamics.
Key Features:
Versatility: NeuralGCM performs competitively across different timescales, from short-term weather forecasts (1 to 15 days) to multi-decade climate projections.
Emergent Phenomena: It accurately simulates realistic tropical cyclone patterns and their trajectories, showcasing its ability to handle complex climate behaviors.
Efficiency: By hybridizing physics and AI, NeuralGCM achieves orders of magnitude computational savings compared to traditional GCMs, making it scalable for real-time applications.
Technical Innovation:
Integrates the dynamical core of a traditional NWP model with local ML-based parameterizations
A neural network parameterizes unresolved physical processes like cloud formation, precipitation, and subgrid-scale dynamics.
The differentiable dynamical core allows for extended backpropagation, enabling the model to learn interactions between physical processes and atmospheric dynamics over hundreds of simulation steps.
Source: Price, I., Sanchez-Gonzalez, A., Alet, F. et al. Probabilistic weather forecasting with machine learning. Nature 637, 84–90 (2025). https://doi.org/10.1038/s41586-024-08252-9