Machine Learning Weather Prediction: Beyond Physics Models?

Current AI models rely on physical models in two main ways. First, they usually require physically generated initial conditions: observations are assimilated into numerical models to produce an atmospheric analysis. Early direct-from-observation systems such as Aardvark Weather suggest this step may eventually be bypassed. Second, most AI models are trained on ERA5 reanalysis, itself generated using physical models and observations. The central question is whether physical-model data will remain necessary as AI systems increasingly generate data used to train other AI systems.

AI-generated analysis and reanalysis

We will investigate whether ML systems can generate reanalysis-quality data suitable for training forecasting models. Systems such as Aardvark Weather could provide a cheaper alternative to conventional reanalysis. We will build AI-generated reanalysis datasets, train forecasting models on them, and assess forecast skill, physical consistency and downstream utility. We will test whether errors or blind spots introduced by one model become amplified by the next.

Virtual observations

We will study whether ML can create “virtual weather stations”. A real station could be deployed briefly while an ML model learns to predict its measurements from remote-sensing data, then reconstruct station-like observations over much longer periods. These synthetic observations could support training and evaluation. We will assess cascades in which virtual observations feed AI analysis and forecast models.

AI forecasts as training augmentation

The Aurora model has shown that forecast trajectories from strong physical models can improve an AI model when used as pre-training data, augmenting the historical record with realistic weather states. We will test whether forecasts from AI systems can play the same role. This raises a question familiar from LLMs and generative modelling: does repeated training on model-generated data lead to “model collapse”, or can suitable methods make synthetic forecast data beneficial? We will establish when AI-generated training data helps, when it degrades performance, and how to avoid collapse.

Adversarial robustness and security

As an additional work area, we will investigate the sensitivity of AI weather models to adversarial examples: small, deliberately chosen perturbations that cause large forecast errors. Starting with Otter Weather, we will compare worst-case perturbations with random noise, examine impacts on extremes such as tropical cyclones, and test whether attacks transfer between models. We will then study end-to-end systems such as Aardvark Weather, asking whether manipulation of a small subset of satellite or ground observations could materially alter forecasts. We will also consider multi-step autoregressive forecasts, deterministic versus generative models, attacks on predictive uncertainty, and physically plausible perturbations within sensor-error limits.

Summary

The project will establish how far AI weather forecasting can become self-sustaining: learning directly from observations, generating analysis products and synthetic training data, and remaining robust when inputs are imperfect or deliberately manipulated. It will clarify the future roles of observations, physical models and AI-generated data, and the opportunities and risks of increasingly AI-native weather prediction.

Student Desirable Background:

An ideal background would involve exposure to the following topics: probabilistic modelling, statistics, and deep learning. The candidate should have a keen interest in environmental science. Knowledge of diffusion, transformers, graph neural networks, and neural processes would be a great bonus.

By submitting an application for a University of Cambridge project, you consent to your documents being shared with the relevant supervisory team at the partner institution. If you have questions about this process aidecisionscdt@manchester.ac.uk before applying.

How to apply

Please apply under the University of Manchester application portal and select 'PhD in Artificial Intelligence' (OAA Application form).

When applying, you must include the project’s full name, your supervisor’s name, funding status, academic history, and two referees’ contact details. Applications with missing documents will not be processed. Late or incomplete submissions will not be considered, and the university is not responsible for missed deadlines.

After you have applied you will be asked to upload the following supporting documents:

• Final Transcript and certificates of all awarded university level qualifications
• Interim Transcript of any university level qualifications in progress
• CV
• Supporting statement: A mandatory one- to two-page statement explaining your motivation for postgraduate research, why you chose Manchester, relevant academic or work experience, key research findings, and the techniques or skills you've gained. Applications will be put on hold if this is not provided.
• Contact details for two referees (Please ensure referee email addresses are official university or workplace accounts, as we may need to verify the references.)
• English Language Certificate (if applicable)

Equality, diversity and inclusion is fundamental to the success of The University of Manchester, and is at the heart of all of our activities. We know that diversity strengthens our research community, leading to enhanced research creativity, productivity and quality, and societal and economic impact. (Equality, diversity and inclusion | The University of Manchester)

We actively encourage applicants from diverse career paths and backgrounds and from all sections of the community, regardless of age, disability, ethnicity, gender, gender expression, sexual orientation and transgender status.

We also support applications from those returning from a career break or other roles. We are dedicated to supporting work-life balance and offer flexible working arrangements to accommodate individual needs. Our selection process is free from bias, and we are committed to ensuring fair and equal opportunities for all applicants.

We are dedicated to supporting work-life balance and offer flexible working arrangements to accommodate individual needs, including a part-time study option.

Join us and help grow the UK’s national capability in artificial intelligence, part of the UK’s modern industrial strategy. Everyone is welcome if you are passionate about uncertainty quantification, decision making with humans in the loop and decision making for machine learning systems and their applications in Science and Engineering, and looking to develop your research and innovation skills at the doctoral level to further your career in the AI sector.

Funding Notes

The UKRI AI CDT in Decision Making for Complex Systems is part of the TechFirst offering stipends of £30,000 for students eligible for home fee status. The start date is September 2027.

The Project will be based in University of Cambridge.

References

Aardvark Weather / end-to-end forecasting from observations
Allen, A., Markou, S., Tebbutt, W. et al. End-to-end data-driven weather prediction. Nature 641, 1172–1179 (2025). https://doi.org/10.1038/s41586-025-08897-0.
This is the key reference for the argument that AI forecasting can bypass conventional real-time analysis/data-assimilation pipelines.

Aurora / heterogeneous pre-training and forecast data as training data
Bodnar, C., Bruinsma, W. P., Lucic, A. et al. A foundation model for the Earth system. Nature 641, 1180–1187 (2025). https://doi.org/10.1038/s41586-025-09005-y.
Particularly relevant to the section on pre-training with multiple datasets, including operational forecast data, and then fine-tuning for downstream Earth-system tasks.

Otter Weather / inexpensive AI forecasting model
Diaconu, C., Scholz, J., Shysheya, A., Markou, S., Mukhopadhyay, P., Cranmer, M. & Turner, R. E. Otter Weather: Skillful and Computationally Efficient Medium-Range Weather Forecasting. arXiv:2606.26421 (2026).
This is the natural model reference for the proposed initial adversarial-robustness experiments.

GraphDOP / forecasting trained and initialised directly from observations
Alexe, M., Boucher, E., Lean, P. et al. GraphDOP: Towards skilful data-driven medium-range weather forecasts learnt and initialised directly from observations. arXiv:2412.15687 (2024).
GraphDOP is particularly relevant to the proposal because it is trained and initialised from observations without physics-based reanalysis inputs.

AIFS-DOP / newer observation-only forecasting result
Pinnington, E., Lean, P., Alexe, M. et al. AIFS-DOP: End-to-End Medium-Range Weather Prediction from Observations Alone with Machine Learning. arXiv:2606.19093 (2026).
The newer version of GraphDOP, trained on 40 years of gridded observations without NWP reanalysis/model data and reported as competitive with IFS on several observation-verified scores.

AI forecasting models have been trained.

Model collapse from recursively generated data
Shumailov, I., Shumaylov, Z., Zhao, Y. et al. AI models collapse when trained on recursively generated data. Nature 631, 755–759 (2024). https://doi.org/10.1038/s41586-024-07566-y.
This is probably the cleanest citation for the “model collapse” motivation in your synthetic-training-data section.

Foundational adversarial examples reference
Goodfellow, I. J., Shlens, J. & Szegedy, C. Explaining and Harnessing Adversarial Examples. International Conference on Learning Representations (2015).