Hot Topics: Machine Learning for Weather Forecasting & Heatwave Prediction Medium Range & Beyond

29th May 2025 Vicky Lucas Hot Topics: Machine Learning for Weather Forecasting & Heatwave Prediction to Medium Range & Beyond A Guide for Meteorologists – Burn After Reading Why read this report? Written for operational or research meteorologists everywhere, this is a guide on how machine learning is being used in weather prediction, with a focus on the challenge of extreme heat. Cutting through the hype – it’s an approachable review on what is happening and what to watch. If you are wondering where things are heading, keep reading – some of the fieriest issues are still unresolved. 1. Degrees of change: introduction ........................................................................................... 2 2. Mercury rising: chronology of machine learning weather models ....................................... 3 3. On the horizon: ML in heatwave forecasting & beyond medium range ................................ 7 4. The heat is on: latest developments ...................................................................................... 9 5. Flashpoint: enduring challenges .......................................................................................... 10 6. Glossary: light relief ............................................................................................................. 11 Hot Topics: Machine Learning for Weather Forecasting & Heatwave Prediction Medium Range & Beyond 29th May 2025 UBC Climate Solutions Research Collective – Solutions Scholars Program 2024/25 Vicky Lucas 1. Degrees of change: introduction Machine learning (ML) is reshaping meteorological modelling, offering new ways to emulate, improve or replace a range of weather and climate modelling (Chantry et al., 2021). In just a few years techniques have moved from focused applications, such as radar nowcasting, to full global models using machine learning for weather prediction (MLWP) that rival, and in some cases outperform, traditional numerical weather prediction (NWP) systems. The shift is not just in who builds these models, often technology companies, but in how forecasts are produced. MLWP does not solve physical equations, but learns patterns in data, based on techniques used in other machine learning applications such as recommendation engines, image generation and Large Language Models (LLMs). Many MLWP systems are trained on historical and constructed weather datasets, very often reanalysis – and especially ERA5 (Climate Data Store). MLWP model training requires significant data, time and energy, but once trained, operational ML models can produce forecasts in seconds or minutes on minimal hardware compared with NWP supercomputers. This speed opens up possibilities for larger ensembles, rapid updates, and hybrid workflows. Something to bear in mind is that this is a domain in flux, no architecture clearly prevails. Groups are experimenting with configurations, progressively leapfrogging their own and others’ achievements. While many MLWP models appear in peer-reviewed journals, some remain as pre-prints – including references in this report (e.g., arXiv), and some make open-source versions available. As developments continue, the plethora of global general MLWP models can be divided by their training regimes (Shi et al., 2025 – arXiv): • Deterministic predictive learning, as used in FourCastNet, uses supervised training – labelled data where each input comes with a corresponding output – a learnt relationship. These models minimise forecast error, are fast to run, but accumulate errors over longer lead times. • Probabilistic generative modelling, such as GenCast, uses diffusion to generate ensemble forecasts, offering uncertainty modelling without explicit perturbations of NWP ensembles. Generative models learn the distribution of the training data, then generates new samples. • Foundation models, such as Aurora, pre-trained on extremely large and diverse datasets to fulfill a broad range of tasks, transferring learning to other datasets either with fine-tuning on new data or prompting e.g., from weather to air quality modelling. This broad training delineation is a framework to understand ML in meteorology. In some instances, deep learning, a subset of ML, would be a more accurate term, but ML is used here for simplicity. A glossary of ML terms is included at the end of this report – when new terms are used a hyperlink to the glossary is provided. This report follows how MLWP is shifting from research to real-world use. It starts by tracing the recent proliferation of global ML models, introducing some of the notable participants and a little on how they work. Next, is an exploration on what is happening in forecasting heatwaves and extending the forecasting window beyond the medium range and into sub-seasonal to seasonal (S2S). The final sections look ahead to where the field is going – concluding with unresolved tensions and challenges. Hot Topics: Machine Learning for Weather Forecasting & Heatwave Prediction Medium Range & Beyond 29th May 2025 UBC Climate Solutions Research Collective – Solutions Scholars Program 2024/25 Vicky Lucas 2. Mercury rising: chronology of machine learning weather models The rapid development of general global machine learning weather models reflects technology company efforts in meteorological modelling and capitalising on techniques, such as used in Large Language Models e.g., ChatGPT (transformers) or image generation (diffusion). Since 2022 MLWP models have made headlines on outperforming respected operational NWP. These ML models were optimised for minimising average errors, but later models have switched to generative modelling and two types of this approach, GenCast and Aurora, are discussed in more detail later. A timeline of general global ML models 2022 to 2025 is shown in Figure 1, and a glossary of ML terms used is provided at the end of this report. FourCastNet (Pathak et al., 2022 – arXiv) was the first general-purpose deterministic MLWP model based on a neural network that rivaled the accuracy of operational NWP. FourCastNet was trained on 10TB of data, emulating atmospheric dynamics learnt from ERA5 reanalysis. It can be run globally in seconds at 0.25° resolution. The breakthrough was twofold. Firstly, using the Fourier Neural Operator approach (Li et al., 2021 – arXiv) which learns mapping between entire spatial fields (function space) rather than individual and neighbouring grid points. Its Fourier architecture captures both lowand high-frequency components, efficiently representing global and local structures. Second, the integration of a Vision Transformer enables the model to learn long-range spatial dependencies. The Vision Transformer was a landmark paper on self-attention, which allows networks to prioritize relevant information when learning; published by Google in 2017 and used in LLMs such as ChatGPT – it is one of the most cited papers this century (Pearson et al., 2025). Together, these innovations enabled fast learning, high resolution, and efficient computation. GraphCast (Lam et al., 2023) appeared in pre-prints soon after, later published in Science, applying Graph Neural Networks to global forecasting. GNNs represent the atmosphere as an icosahedral mesh, enabling uniform spatial resolution. Nodes (grid points) are connected by edges which define the flow of information between all nodes – allowing the model to learn teleconnections regardless of distance or input grid regularity. Forecasting over 1,000 variables up to 10 days ahead, GraphCast achieved greater accuracy than ECMWF IFS conventional NWP model – including at 10 day lead times for the top 2% of hot days over land (Lam et al., 2023). Previous models had taken their names from their architecture, but Pangu-Weather changed that – named after a primordial mythical being who separated heaven and earth and became geographic features. This deterministic model introduced two new approaches (Bi et al., 2023). Firstly, it used a three-dimensional approach, using vertical structure via pressure level data. Secondly it trained separate models for 1, 3, 6 and 24-hour lead times, which reduced errors over long lead times. Figure 1: timeline of selected ML weather models Hot Topics: Machine Learning for Weather Forecasting & Heatwave Prediction Medium Range & Beyond 29th May 2025 UBC Climate Solutions Research Collective – Solutions Scholars Program 2024/25 Vicky Lucas Throughout these advances ECMWF, the European Centre for Medium-Range Weather Forecasts – both a research institute and operational service – has been central in the application of ML. The ECMWF operational NWP, the IFS HRES, is often used as the benchmark model comparison, or the ensemble IFS ENS. ECMWF developed its own deterministic graph-based (GNN) system, AIFS (Lang et al., 2024 – arXiv), and an ensemble AIFS-ENS based on diffusion (Figure 1) (Lang et al., 2024). Diffusion is the same approach used for image generation and in a later MLWP model, GenCast – features in the next section of this report, where diffusion will be discussed in more detail. The AIFSENS was made operational in early 2025, alongside the longstanding conventional NWP the IFS. ECMWF displays third-party MLWP outputs in real-time and evaluates model performance. Figure 2 shows comparative performance through the root mean squared error of two-metre air temperature for the northern extra-tropics, for winter 2024/25 – showing that Aurora consistently achieves the greatest accuracy on that broad metric – outperforming the operational NWP, IFS. A range of charts and performance scores are available from the ECMWF chart catalogue. Figure 2: RMSE 2m temperature, winter 24/25 northern extra-tropics - consistently best performance from Aurora IFS (ECWMF operational NWP) (red) and five MLWP models: AIFS (ECMWF) (brown), FourCastNet (green), GraphCast (cyan), Pangu-Weather (dark blue) & Aurora (orange) – available from ECMWF Hot Topics: Machine Learning for Weather Forecasting & Heatwave Prediction Medium Range & Beyond 29th May 2025 UBC Climate Solutions Research Collective – Solutions Scholars Program 2024/25 Vicky Lucas ClimaX (Nguyen et al., 2023 – arXiv) was the first foundation model, trained on truly vast and heterogeneous datasets including ERA5 and the climate model output of CMIP6. The approach introduced a shared transformer encoder pre-trained, and which can be fine-tuned to more specific tasks. ClimaX is capable of global and regional forecasts, also sub-seasonal and climate projections, and downscaling. Similarly, Prithvi WxC (Schmude et al., 2024 – arXiv) is a later foundation model, a collaboration between IBM and NASA. Another first, Aardvark, has recently been published in Nature (Allen et al., 2025). The system has two notable features, firstly it can run on a desktop computer in minutes (Turner, 2025). Secondly, whilst it has been trained on ERA5, it does not need any NWP products to run, it only requires observations as input such as satellite and weather station data. This end-to-end system, fully observation-driven, has been an aspiration in the field (McNally et al., 2024). Aardvark shows strong performance against operational NWP, even with a coarse 1.5 degree output and using just 10% of inputs to existing NWP systems (Turner, 2025). These features make Aardvark valuable for applications like disaster preparedness in low-resource settings. The details: GenCast & Aurora To illustrate recent architectures, two models are detailed here, GenCast and Aurora (Figure 1). GenCast is a probabilistic generative model for medium range 15 day forecasting (Price et al., 2025) while Aurora is a foundation model which can be tuned to multiple applications including forecasting air quality and tropical cyclone tracks (Bodnar et al., 2024 – arXiv). GenCast is a generative model – in training it learns the underlying data distribution and then produces new realistic samples. This generative function comes from using a diffusion process, based on noise. In training, to predict conditions 12 hours ahead (T+12), GenCast takes two input fields from ERA5, effectively T-12 and T+0. Rather than forecasting T+12 directly, it computes the residual (i.e., the change between T–12 and T+0) and adds scaled noise to create the first guess. GenCast then iteratively removes noise over multiple steps. By perturbing the tendency rather than the absolute state, it preserves coherent atmospheric structures. GenCast was pre-trained at 1° resolution on 40 years of ERA5 data, then fine-tuned at 0.25°. In forecast mode, known as inference in ML, GenCast predicts each T+12 using the two most recent states, continuing until it reaches T+360 (15 days). This type of rolling prediction is called autoregression in ML. Generating an ensemble is straightforward due to the diffusion approach; each ensemble member starts from a different scaled noise sample. 50 ensemble members were chosen to match ECMWF IFS ENS – a lot more could be produced. While diffusion models are computationally intensive—each step involves multiple denoising passes—GenCast can produce a full 15-day forecast in minutes per member. Its Graph Neural Network (GNN), adapted from GraphCast (discussed earlier), captures and supports spatial dependencies across time steps. GenCast outperformed IFS ENS, including extreme temperatures, the 99.9th and 99.99th percentiles, out to 15 days (Price et al., 2025). In contrast, Aurora is a foundation model – the term foundation model came from Stanford in 2021 – where the model is a common basis for many task specific adaptations. Aurora is pre-trained on more than one million hours of global Earth system data, including weather reanalyses, operational forecasts, and the CMIP6 climate simulations. It contains 1.3 billion parameters and uses a 3D Swin Transformer, an architecture adapted from computer vision. The Swin, shifted window, transformer improves efficiency by dividing data into local windows and then shifting them between layers to capture broader spatial and vertical dependencies. Aurora also employs perceiver-based encoders Hot Topics: Machine Learning for Weather Forecasting & Heatwave Prediction Medium Range & Beyond 29th May 2025 UBC Climate Solutions Research Collective – Solutions Scholars Program 2024/25 Vicky Lucas and decoders which allow the model to ingest varying types and shapes of data, transforming them to a common internal format to combine different input data. These architectural choices allow Aurora to model complex Earth system dynamics across domains. After pretraining, it can be fine-tuned for tasks like 10-day global weather forecasting, 5-day air quality prediction, ocean wave modelling, and tropical cyclone tracking. Aurora (Bodnar et al., 2024 – arXiv) produces 10-day global forecasts in under a minute and 5-day air quality predictions in seconds. Summary From the fast run times of FourCastNet to the flexible tuning of Aurora and the ensemble uncertainty quantification of GenCast, MLWP is approaching, and by some metrics exceeding NWP performance, all with significantly reduced operational computational cost. The enablers of that progress include: • Open access to high-resolution datasets such as ERA5, • Cross-disciplinary architecture design (e.g. transformers, diffusion), • Collaboration between research institutes, operational centres, and tech companies, • Access to MLWP on GitHub and Hugging Face – for sharing data, code and documentation. For further reading, Chen et al. (2023) offers a good overview of methods and architectures, while Zhang et al. (2025) provides an up-to-date review of knowledge, methods, trends and challenges. Waqas et al. (2024) offers a systematic review of AI integrated with NWP. For more on MLWP comparative performance, Google Research has a scorecard of several models at WeatherBench2. Hot Topics: Machine Learning for Weather Forecasting & Heatwave Prediction Medium Range & Beyond 29th May 2025 UBC Climate Solutions Research Collective – Solutions Scholars Program 2024/25 Vicky Lucas 3. On the horizon: ML in heatwave forecasting & beyond medium range General MLWP at medium range & S2S MLWP is extending skill into the medium and sub-seasonal range (15–46 days). Two deterministic general purpose transformer models, FuXi and FengWu have moved to architectures with ensemble generation for forecasts out to six weeks in S2S versions. FuXi-S2S demonstrates improved skill over ECMWF IFS S2S for several metrics, including 2m temperature and Madden-Julian Oscillation (MJO) prediction up to day 36 (Chen et al., 2024). FengWu-W2S uses ocean–atmosphere–land coupling constraints for physical consistency (Ling et al., 2024). The results for 2m temperature show comparable but greater skill than ECMWF S2S and FuXi-S2S at three to six weeks. FengWu-W2S can predict the MJO to 37 days, slightly longer than ECMWF S2S or FuXi-S2S, and predicts the North Atlantic Oscillation with greater skill than ECMWF S2S at four to six weeks. Climate ML models are also being tested in climate applications. NeuralGCM (Kochkov et al., 2024) is a hybrid model trained on ERA5 that performs well for 10–15 day forecasts and can simulate climate metrics over multiple decades when forced with prescribed sea surface temperatures. The model is not coupled to land or ocean and does not include greenhouse gases or aerosols. NeuralGCM does not show a clear trend of increasing error when initialised further into the future from the training data. ‘Our results provide strong evidence for the disputed hypothesis that learning to predict shortterm weather is an effective way to tune parameterisations for climate.’ (Kochkov et al., 2024). ML improving local & regional NWP for extreme heat Machine learning is being used to improve heatwave forecasting at local scales and to extend skill in forecasting surface extremes. At the urban level, ML has been used for bias correction and downscaling. A recent London study (Blunn et al., 2024) used ML to combine high resolution NWP output with citizen science weather station observations during heatwaves. ML reduced mean absolute temperature error by 11%, identifying latent heat flux as the most important predictor of temperature bias. In regional modelling, ML can support sensitivity analysis of NWP physical schemes. One example from Australia (Reddy et al., 2023) used ML to identify, of twenty four parameters, the two key drivers of modelled surface temperature and relative humidity predictions: the shortwave radiation scattering parameter and saturated soil water content multiplier. These two studies show complementary uses of ML in improving conventional NWP. MLWP extending lead times for extreme heat prediction Papers from 2022 to 2025 exploring the use of ML approaches for extending the lead times of heatwave predictions is detailed in Table 1, and several discussed below. The table also includes recent papers on explainable AI (XAI) discussed in the next section, and two review papers. Among the more recent and longer-range examples is the Weirich-Benet et al. (2023) study, which applied both linear regression and random forests to forecast summer heatwaves in Central Europe at lead times of 1–6 weeks. Key predictors included 500 hPa geopotential height, soil moisture, and sea surface temperatures. While model performance declined with increasing lead time, the machine learning outputs outperformed persistence and climatology. Beyond two weeks, their forecasts were as skilful as the ECMWF sub seasonal ensemble-mean hindcast for the region. Lopez-Gomez et al. (2023) explored ML configurations trained specifically on extreme temperatures — rather than all temperatures. This adjustment led to improved performance over persistence and comparable skill with ECMWF S2S after day 14. The MLWP could forecast out-of-sample events, but struggled with extremely hot days, above the 95th percentile. Hot Topics: Machine Learning for Weather Forecasting & Heatwave Prediction Medium Range & Beyond 29th May 2025 UBC Climate Solutions Research Collective – Solutions Scholars Program 2024/25 Vicky Lucas In a case study of the North American Pacific Northwest 2021 heatwave, the anomalous temperatures were only predicted 2 to 5 days prior, although at about a month ahead the MLWP was similar in structure to the ECMWF ensemble. Duan et al. (2025), apply the NeuralGCM, discussed above, to explore the Pacific Northwest heatwave, which it replicates. Xie et al. (2024), used a Convolutional Neural Network – an approach that finds local spatial correlations in gridded data – to forecast heatwaves in China up to 30 days ahead. The method filtered out high-frequency signals (<10 days) and isolated the 10–90 day low-frequency background state, using inputs from NCEP/NCAR reanalysis and polar-orbiting satellite data. While forecast skill was lower up to 20 days, the model outperformed both the China Meteorological Administration NWP and ECMWF S2S ensemble mean between days 20 and 30 when averaged across the country. Table 1: summary of papers on ML approaches applied to heat extremes from 2022 to 2025 Reference & link Data or model Area Technique Lead time or purpose Ennis et al., 2025 GraphCast, FuXi, Pangu, GEFS USA MLWP – NWP comparison Up to 20 days Lovo et al., 2025 CESM (NCAR) France Multiple ML models, also XAI study XAI Shafiq et al., 2025 5 years weather observations Lahore XAI study, LSTM best performance 1-3 days, XAI Camps-Valls et al., 2025 Multiple Global Review of methods Review paper Duan et al., 2025 ERA5 Pacific Northwest NeuralGCM and ensemble comparison E3SM NWP 6 days Xie et al., 2024 NCEP/NCAR reanalysis China CNN with filtering Up to 30 days Weirich Benet et al., 2023 ECMWF Europe Linear and Random Forest 1-6 weeks Lopez Gomez et al., 2023 ERA5 Global Multiple ML models Up to 28 days Salcedo-Sanz et al., 2023 Multiple Global Review of methods Review paper MLWP for extremes – open questions Despite encouraging case studies, recent reviews caution that MLWP is not ready for operational forecasting of extremes. Olivetti & Messori (2024) highlight three key limitations: (i) most models are tuned for average forecast skill, not extremes; (ii) architectures are not optimised for the limited data available on rare events; and (iii) assumptions about error distributions are often too simplistic. They also note that leading global MLWP lack validation for extreme event prediction. Salcedo-Sanz et al. (2023) provide a history and broader critique, focused on extreme events and a literature review of heatwaves. They also emphasise the limited data available for training on rare events and argue that most current ML systems lack integration with physical climate knowledge to achieve reliable predictions. The authors call for the use of multiple reanalysis datasets, improved model transparency through explainable AI (XAI), and greater focus on compound and concurrent extremes. Similarly, the review paper of Camps-Valls et al., (2025) discusses the hurdle of limited training data, and of deploying understandable models – needed for gaining trust. Many authors underscore that MLWP requires consistent and transparent development for extremes. Hot Topics: Machine Learning for Weather Forecasting & Heatwave Prediction Medium Range & Beyond 29th May 2025 UBC Climate Solutions Research Collective – Solutions Scholars Program 2024/25 Vicky Lucas 4. The heat is on: latest developments As MLWP matures, researchers are responding to critiques — including the validation of extreme event forecasts. Recent work evaluated three global MLWP models (FourCastNet, Pangu-Weather, GraphCast) alongside ECMWF IFS HRES for events including the 2021 Pacific Northwest heatwave (Pasche et al., 2025). While ML models captured the broad structure of the event, they underperformed when spatially and temporally aggregated. FourCastNet had the largest errors; Pangu-Weather tended to overestimate affected areas for a given threshold. At lead times under a week, the IFS HRES remained the most accurate. Notably, MLWP have few variables – lacking surface relative humidity – and the authors indicated the need for sea surface temperature and soil moisture inputs for longer lead times. A separate study examined 60 U.S. heatwaves and found GraphCast consistently outperformed Pangu-Weather and the US GEFS ensemble across most regions with ERA5 used as ground truth (Ennis et al., 2025 – arXiv). The open-source nature of some MLWP models allows for this independent analysis and verification – which is likely to continue – assisted by the WeatherBench2 (Rasp et al., 2023 – arXiv) database of extreme event cases, (also ClimateBench (Watson-Parris et al., 2022)) and ECMWF adding MLWP forecast evaluation to its database of extreme cases (Magnusson, 2023). Explainability remains a major concern. While studies by Lovo et al. (2025), Shafiq et al. (2025), and Wei et al. (2025 – arXiv) have applied explainable AI (XAI) tools to heat forecasting, these often rely on simplified models or components. As Lovo notes, Convolutional Neural Networks – the earlier predictive approach for MLWP suited to gridded data and used e.g., by GraphCast – ‘are black boxes even when using XAI tools’. More transparent methods, hybrid models combining ML with physical constraints, are important for scientific trust and likely to be so for public trust. A potential trend is the integration of large language models (LLMs) to support forecast interpretation and decision-making. WildfireGPT, developed as a proof-of-concept, combined extreme weather projections, literature, and user prompts to generate geospatially contextualized response guidance (Xie et al., 2024 – arXiv). Designed to be user-centric and visually integrated, it outperformed generic LLMs in ten case studies. A similar ‘HeatwaveGPT’ could help bridge the gap between technical NWP, ML outputs, operational needs and actions, and public-facing messaging. Looking ahead, the field is still evolving and diversifying. Aardvark, capable of generating forecasts directly from observations, represents a shift in architecture – able to include some measurements too complex for NWP assimilation (McNally et al., 2024). And new systems regularly appear such as the open source WeatherMesh-3 (Du et al., 2025) by an atmospheric sensing company, WindBorne. The S2S timescales and beyond will be a domain of interest; just as MLWP can outperform NWP at short to medium range (1 to 15 days) seasonal predictions will improve for MLWP – including global seasonal or decadal predictions such as ACE2 (Watt-Meyer et al., 2024, Kent et al., 2025 – arXiv). Another recent development is from NVIDIA – Earth-2 is their commercial digital twin which can be used for kilometre scale weather and climate predictions, it is focused on risk and is a business offering (NVIDIA, 2025). Establishing business models could be a next step for other organisations that have invested resources over the last few years. As validation continues, explainability evolves and new models appear, the importance of ML in meteorological prediction looks set to remain. Progress could be incremental or radical depending on performance and resourcing pressures – but going too far too fast could undermine trust and acceptance – from science and from the public.