Introduction 

The idea of being able to make long range predictions, months, seasons or even years ahead involves a complex combination of initial conditions in the atmosphere, ocean, land and cryosphere, a good representation of internal variability and teleconnections across the globe and an accurate prescription of external boundary forcing. Here we take a look at the current state of this science and in spite of the advice attributed to Niels Bohr, that “Prediction is very difficult, especially about the future”, we make what will perhaps turn out to be a foolhardy attempt to predict some of the future trends in the field. 

Scientific Progress 

Daily weather appears to have an effective predictability limit of a couple of weeks (Charney 1966; Lorenz 1969; Selz et al., 2019; Zhang et al 2019). However, it’s important to note that this is not due to any randomness in the weather, in fact there is no randomness in weather prediction processes at all. They are the exact opposite and are inherently deterministic: if you start the weather forecast from exactly the same conditions then you get exactly the same forecast. Instead, it’s the inherent sensitivity of atmospheric evolution to small changes (that are both impractical to measure and inevitably present in prediction models) that alters the long-term evolution of the atmosphere and limits long range predictability. At this point we could throw up our hands and say that long range forecasts are a lost cause. This may even be the case for some situations and predictions at one month range for the midlatitudes for example are notoriously difficult and may be limited to sporadic windows of opportunity (Vitart et al., 2017; Kent et al., 2022). However, even these apparent inherent limits on the predictability of daily weather processes are being challenged (Vonich and Hakim, 2024). It is also well established that some processes in the earth system already provide much longer range predictability from the ocean (Msadek et al., 2010; Shukla 1998), from the land (Koster et al., 2004; Seo et al., 2018) or sometimes even from the atmosphere (Scaife et al. 2022), as they have much longer timescales than daily weather. In these cases, like the effects of systematic climate change, they can impart long range predictability of average weather and other weather statistics at forecast ranges of months or even years ahead (Smith et al, 2020; Meehl et al., 2021).  

It’s important to realise that some of these processes are not fully understood (Kushnir et al 2019, Schmidt 2024), or are not perfectly represented in the computer models we use for prediction (Anstey et al., 2022; Scaife and Smith 2018; Sun et al., 2015) and are therefore not fully functioning in guiding our predictions. Because of this, we see continued improvements in the understanding (Patrizio et al., 2025), simulation (Takahashi et al., 1999) and skill of long-range predictions of climate variability (Scaife et al., 2014) and their impacts on surface weather and climate. Given all this, my first prediction for 2050 is that provided the research is funded, then our knowledge of climate predictability and the skill of long-range predictions will far exceed our current horizons. 

Machine Learning and Artificial Intelligence 

Meteorology is currently undergoing a revolution. Despite initial scepticism, so called ‘artificial intelligence’ algorithms are now starting to exceed basic aspects of the performance of physics-based computer weather forecasting models. Machine learning (ML) is perhaps a more representative term than Artificial Intelligence (AI) for these new methods, but either way, these empirical models, sometimes trained only on observational reanalysis datasets, are breaking new ground in terms of the skill of weather forecasts and the cost and speed at which forecasts can be produced. They offer great promise of increased skill, cheaper forecast production and reduced energy consumption (Keisler et al., 2022; Bi et al., 2023).  

This new methodology and the rapid progress it is making, has surprised many in the field. I’ve observed a kind of three stage process in the response to this revolution. In stage one, on first hearing about new ML breakthroughs, scientists are often naturally initially very sceptical. Later on in stage two, when the initial results are verified and reproduced, we tend to question how far reaching the new results will be and in particular whether they will affect our own work. Then finally, there is a third acceptance stage where we finally admit that actually, this does radically change the state of play. I think this is where we are with ML and weather, while ML and climate is still at stage two, where rapid progress is being made but the full implications are not yet demonstrated. Despite this, it is now clear that not only can very skilful short to medium range weather forecasts be produced using ML models trained on observational reanalysis data but also that skilful subseasonal (Chen et al., 2024) and even seasonal forecasts (Kent et al., 2025) are possible. Other studies have also shown that decadal prediction models can be successfully emulated using neural networks (Toms et al., 2021). Work is now accelerating rapidly on the production of ML climate models (e.g. Watt-Meier et al., 2024; Kochkov et al., 2024), in some cases using purely observational reanalysis data. I stress this cumbersome term as it was not immediately obvious a priori whether the observational record was long enough to achieve this. After all, there are only a few tens of seasons in the modern observational record and the key to successful ML predictions is to provide it with a comprehensive training dataset which can be used to tune the millions of parameters in the underlying neural networks.  

There are of course potential pitfalls with these methods. They currently rely mainly on reanalysis data, itself containing traditional, physics-based model information, and the promise of bias free predictions may be overstated. Nevertheless, it does appear that training models on shorter timescale evolution of the atmosphere is sufficient to make successful long range climate simulations (e.g. Watt-Meier et al 2024). Perhaps this is all less surprising when we acknowledge that neural networks can represent any mathematical function (Ismailov 2023, Hecht-Nielsen 1987) and when we see that the results of tests in atmospheric conditions far from those used in the training data reproduce well known atmospheric response patterns (Hakim and Masanam, 2024). These experiments confirm that ML models can encode, at least to some degree, the underlying physical mechanisms that govern climate variability and seasonal to decadal prediction. Although there are always unknown wild cards, for example the advent of quantum computing which could in principle revolutionise the production of forecasts again, neural network methods are likely to drive a revolution in seasonal to decadal prediction as they are already doing in weather forecasting. 

I am confident that our physics-based models will continue to be necessary for process based meteorological research. They will be important benchmarks for validating and comparing with ML models, as well as production of additional training data due to the slow accumulation of additional observational samples. They will also likely be needed to verify unusual or out of sample predictions which increase under climate change and may be difficult for ML models. Nevertheless, the revolution in ML methods for weather and climate simulation and prediction leads me to a second prediction for 2050: that machine learning methods will accelerate scientific research and operational production of seasonal to decadal predictions by allowing rapid generation of numerical experiments with ensembles of unprecedented size.  

Climate Change 

At the current rate of warming we will have long exceeded 1.5 degrees of global warming by 2050 and the Earth will be around 2 degrees warmer than preindustrial levels. The occurrence of extreme weather and climate events under such levels of global change will be commonplace and clear to all. Indeed, the likelihood of unprecedented extremes such as the 40 degree heatwave in the UK in July 2022 is already exponentially increasing (Kay et al., 2025).  

Unprecedented events invariably occur when climate variability constructively interferes with the ever-growing climate change signal. Compounded with an ever-growing global population, this means that in the future, seasonal, interannual and decadal predictions of climate variability are going to become all the more important for socio-economic resilience and disaster risk reduction. So whether it is predictions of safe passage routes through the disrupted, and at times ice free Arctic Ocean, predictions of intense winter flooding across northern Europe, or predictions of unprecedented drought in the tropics due to El Niño events; the demand for accurate and reliable seasonal and decadal climate predictions of the coming months and years will be intense. Seasonal to decadal predictions will provide the first warnings of impending unprecedented heat waves, droughts, storms and floods and the first indicators of a multitude of subsequent impacts and consequences. 

To serve these needs, the meteorological community needs to step up to the challenge of providing simple, frequently updated climate forecasts with interpretation from the expert meteorologist that are full accessibility by the non-expert layperson. Efforts towards this are already underway (e.g. Kumar et al., (2025); Buontempo et al., (2022); Vera (2024)) but openly available, clear and simple interfaces to skilful and reliable forecasts will be needed if long range prediction is going to fulfil its potential and its promise to society. 

There are a few other climate aspects of seasonal and decadal predictions that I can’t leave out of this short perspective: 

Although they have not yet been observed in the modern era, potential nonlinear ‘tipping’ points in the climate system that could lead to dramatic and effectively irreversible shifts in regional climate have been identified (Lenton et al., 2008). To this end, it is important that we run very large ensembles of predictions and examine the least likely but most impactful future possibilities. Sudden changes in behaviour like the recent disruption to the stratospheric quasi-biennial oscillation after decades of regular cycling (Osprey et al., 2016) illustrate that such effects could perhaps also occur within tropospheric climate and then affect surface weather. Being alert to these possibilities through deeper understanding of climate dynamics, model limitations and deeper examination of unusual forecast ensemble members during the climate forecast process is needed to address these remote but very impactful possibilities. 

There is also a body of work on ‘geoengineering’ which argues that to try to counteract the worst impacts of climate change we could counter climate change with physical interventions, for example to cause global cooling or to avoid the worst extremes (Reynolds, 2019). This controversial area is fraught with ethical problems and perhaps it should never be applied, but if it were to go ahead, there would be a need for real time counterfactuals to establish over many cases and many climate events whether the interventions were having the desired effect. In this case it would be seasonal to decadal ensemble predictions, with and without the geoengineering changes, that would be needed to provide the required statistics to confidently establish the impacts.   

In addition to all of this, there is also increased public awareness of climate variability. The general education of society beyond just appreciating that the climate is warming, to include awareness of when and where climate variability is likely to result in the greatest hazards is needed to reap the benefit of seasonal to decadal predictions and allow anticipation of extreme and record-breaking events. We have made a start to this; who has not now heard of the jet stream? or sudden stratospheric warmings? or the El Niño phenomenon? But having a wider appreciation of the impacts of such phenomena is needed to intelligently interpret forecasts. 

This all leads to my final prediction, that by 2050, seasonal to decadal predictions will be in greater demand and in greater use than ever before in an age with more frequent weather and climate disasters. 

References 

Anstey, J.A., Simpson, I.R., Richter, J.H., Naoe, H., Taguchi, M., Serva, F., Gray, L.J., Butchart, N., Hamilton, K., Osprey, S., Bellprat, O., Braesicke, P., Bushell, A.C., Cagnazzo, C., Chen, C.-C., Chun, H.-Y., Garcia, R.R., Holt, L., Kawatani, Y., Kerzenmacher, T., Kim, Y.-H., Lott, F., McLandress, C., Scinocca, J., Stockdale, T.N., Versick, S., Watanabe, S., Yoshida, K. and Yukimoto, S. (2022). Teleconnections of the Quasi-Biennial Oscillation in a multi-model ensemble of QBO-resolving models. Q.J.R. Meteorol. Soc, 148: 1568-1592. https://doi.org/10.1002/qj.4048 

Bi, K., Xie, L., Zhang, H., Chen, X., Gu, X., & Tian, Q. (2023). Accurate medium-range global weather forecasting with 3D neural networks. Nature, 619(7970), 533–538. https://doi.org/10.1038/s41586-023-06185-3 

Buontempo, C., and Coauthors, (2022). The Copernicus Climate Change Service: Climate Science in Action. Bull. Amer. Meteor. Soc., 103, E2669–E2687, https://doi.org/10.1175/BAMS-D-21-0315.1. 

Charney, J. G., Fleagle, R. G., Lally, V. E., Riehl, H., & Wark, D. Q. (1966). The feasibility of a global observation and analysis experiment. Bull. Amer. Met. Soc., 47, 200–220. 

Chen, L., Zhong, X., Li, H. et al. (2024). A machine learning model that outperforms conventional global subseasonal forecast models. Nat Commun 15, 6425 (2024). https://doi.org/10.1038/s41467-024-50714-1 

Hakim, G. J., and S. Masanam, (2024). Dynamical Tests of a Deep Learning Weather Prediction Model. Artif. Intell. Earth Syst., 3, e230090, https://doi.org/10.1175/AIES-D-23-0090.1. 

Hecht-Nielsen, R. (1987). Kolmogorov’s Mapping Neural Network Existence Theorem. Proceedings of the IEEE First International Conference on Neural Networks. III: 11–13. 

Ismailov V.E. (2023). A three layer neural network can represent any multivariate function. Journal of Mathematical Analysis and Applications, 523, 1, 127096, ISSN 0022-247X, https://doi.org/10.1016/j.jmaa.2023.127096. 

Kay G., et al. (2025). Rapidly increasing chance of record UK summer temperatures. Atm. Sci. Lett., in press. 

Keisler, R. (2022). Forecasting Global Weather with Graph Neural Networks (Version 1). https://doi.org/10.48550/ARXIV.2202.07575 

Kent et al., (2022). What potential for improving sub-seasonal predictions of the winter NAO? Atm. Sci. Lett., e1146, doi.org/10.1002/asl.1146 

Kent et al. (2025). Skilful global seasonal predictions from a machine learning weather model trained on reanalysis data. npj Clim. Atm. Sci., submitted 

Koster R.D. et al., (2004). Regions of Strong Coupling Between Soil Moisture and Precipitation. Science 305,1138-1140.DOI:10.1126/science.1100217 

Kumar et al., (2024). World Meteorological Organization (WMO) Accredited Infrastructure to Support Operational Climate Prediction. Bull. Amer. Met. Soc., doi.org/10.1175/BAMS-D-23-0284.1 

Kushnir Y. et al. (2019). Near Term Climate Prediction. Nat. Clim. Ch., 9, 94-101. 

Lenton, T.M., et al., (2008).Tipping elements in the Earth’s climate system, Proc. Natl. Acad. Sci. U.S.A. 105 (6) 1786-1793, https://doi.org/10.1073/pnas.0705414105. 

Lorenz, E. N. (1969). Atmospheric predictability as revealed by naturally occurring analogues. J. Atm. Sci., 26, 636–646. 

Meehl G. et al, 2021. Initialised Earth System Prediction from Subseasonal to Decadal Timescales. Nat. Rev. Earth. Environ., 2, 340-357. 

Osprey S., et al. (2016). An unexpected disruption of the atmospheric quasi-biennial oscillation. Science, 0.1126/science.aah4156. 

Patrizio, C.R., et al. (2025). Ocean-atmosphere feedbacks key to NAO decadal predictability. npj Clim Atmos Sci 8, 146. https://doi.org/10.1038/s41612-025-01027-7 

Reynolds J.L. (2019). Solar geoengineering to reduce climate change: a review of governance proposals, Proc. Roy. Soc. A, 475, 2229, 10.1098/rspa.2019.0255 

Scaife A.A., et al. (2014). Predictability of the Quasi-Biennial Oscillation and its Northern Winter Teleconnection on Seasonal to Decadal Timescales. Geophys. Res. Lett., 41, 1752-1758. 

Scaife A.A. et al. (2022). Long range predictability of extratropical climate and the length of day. Nat. Geosci., 10, 10.1038/s41561-022-01037-7. 

Selz, T., et al. (2019). Estimating the Intrinsic Limit of Predictability Using a Stochastic Convection Scheme. J. Atmos. Sci., 76, 757–765, https://doi.org/10.1175/JAS-D-17-0373.1.  

Seo E., et al., (2018). Impact of soil moisture initialization on boreal summer subseasonal forecasts: mid-latitude surface air temperature and heat wave events Clim. Dyn., doi:10.1007/s00382-018-4221-4. 

Shukla, J. (1998). Predictability in the Midst of Chaos: A Scientific Basis for Climate Forecasting. Science 282, 728-731. DOI:10.1126/science.282.5389.728 

Smith D. et al., (2020). North Atlantic climate far more predictable than models imply. Nature, 583, 796-800. 

Sun, C., Li, J. & Jin, FF. A delayed oscillator model for the quasi-periodic multidecadal variability of the NAO. Clim Dyn 45, 2083–2099 (2015). https://doi.org/10.1007/s00382-014-2459-z  

Takahashi M. (1999). Simulation of the Quasi-Biennial Oscillation in a general circulation model. Geophys. Res. Lett., Vol. 26, Issue 9, 1307-1310. 

Vera, C.  (2024). Climate Services in South America. Oxford Research Encyclopedia of Climate Science.  https://oxfordre.com/climatescience/view/10.1093/acrefore/9780190228620.001.0001/acrefore-9780190228620-e-963. 

Vitart, F., et al., (2017). The Subseasonal to Seasonal (S2S) Prediction Project Database. Bull. Amer. Meteor. Soc., 98, 163–173, https://doi.org/10.1175/BAMS-D-16-0017.1. 

Kent et al. (2025). Skilful global seasonal predictions from a machine learning weather model trained on reanalysis data. npj Clim. Atm. Sci., submitted 

Msadek, R., K. W. Dixon, T. L. Delworth, and W. Hurlin (2010), Assessing the predictability of the Atlantic meridional overturning circulation and associated fingerprints, Geophys. Res. Lett., 37, L19608, doi:10.1029/2010GL044517. 

Toms, B. A., Barnes, E. A., & Hurrell, J. W. (2021). Assessing decadal predictability in an Earth-system model using explainable neural networks. Geophys. Res. Lett., 48, e2021GL093842. https://doi.org/10.1029/2021GL093842  

Vonich, P. T., & Hakim, G. J. (2024). Predictability limit of the 2021 Pacific Northwest heatwave from deep-learning sensitivity analysis. Geophysical Research Letters, 51, e2024GL110651. https://doi.org/10.1029/2024GL110651  

Zhang, F., Y. Q. Sun, L. Magnusson, R. Buizza, S. Lin, J. Chen, and K. Emanuel (2019). What Is the Predictability Limit of Midlatitude Weather? J. Atmos. Sci., 76, 1077–1091, https://doi.org/10.1175/JAS-D-18-0269.1.