ML Coupling Workshop

3rd - 4th September 2025, Cambridge UK.

Programme

Wednesday 3rd September

9:00-9:30 Arrival and Coffee
9:30-9:45 Welcome
9:45-11:15 Talks Session 1
11:15-11:45 Break
11:45-12:45 Talks Session 2
12:45-14:00 Lunch
14:00-15:30 Talks Session 3
15:30-16:00 Break
16:00-16:45 Panel Discussion
16:45-17:00 Introduction to breakout groups
17:00-19:00 Poster session with drinks and an informal dinner

Thursday 4th September

9:00-9:15 Arrival and Coffee
9:15-9:30 Get into breakout groups
9:30-10:30 Breakout session 1
  • Coupling Interfaces
  • Hardware
  • Differentiable Models and Online Training
10:30-11:00 Feedback Session 1
11:00-11:30 Break
11:30-12:30 Breakout session 2
  • Stability and Uncertainty
  • Machine learning architectures
  • Research to Operations
12:30-13:00 Feedback Session 2
13:00-14:00 Lunch

Talks and Speakers

Improving physical models of the atmosphere using ML

Cyril Morcrette - Met Office

Cyril leads a team improving the way that clouds and radiation are represented in weather forecasts and climate simulations.

Abstract: The Met Office develops and maintains a unified modelling framework used for weather forecasting and climate applications. Simulations using this model can be used to feed product and deliver advice and warnings, they can also be used to provide the next generation of km-scale high-fidelity datasets for ML training. As a result, these physical simulations need to be the best they can be, and ML can be used to improve certain parametrization schemes and interactions between processes. An overview of projects aiming to improve the physical model by embedding ML techniques will be provided. We will also provide an overview of the software development that allow us to fuse ML techniques with our physical model is a manner which is robust to changes in the modelling landscape.

A Tale of Two Couplings

Hannah Christensen - University of Oxford

Associate professor and head of the Atmospheric Processes group at the University of Oxford. Researching uncertainty quantification of parameterisation schemes in numerical models and leading the Model Uncertainty-Model Intercomparison Project (MUMIP). She is also exploring the application of machine learning methods in the domain.

Abstract: Earth System Prediction is changing before our eyes, with Machine Learnt (ML) models now able to out-perform traditional dynamical models for many tasks. In this presentation I will discuss how we can bring together traditional dynamical codebases with ML emulators to advance Earth System Prediction. I will give two examples. In the first, we replace one specific, but uncertain, small-scale process with an ML emulator. In the second, we attempt the more ambitious task of replacing a whole Earth-system component with an ML emulator. In both projects, the ML emulator is trained on observational data. In both projects, the emulator is probabilistic, to represent aleatoric (random) uncertainties captured in these training datasets. However, the challenges involved with the coupling are very different between the two projects, necessitating very different solutions.

The use of turbulence surrogate models in plasma integrated modelling

William Hornsby - UKAEA

Scientific software engineer specialising in modern HPC architectures and machine learning for simulation in fusion applications.

Abstract: Plasma micro-turbulence is one of the dominant transport mechanisms of heat from the core of a fusion power plant. Direct numerical calculation of the micro-instabilities that form turbulence is computationally expensive and is a significant bottleneck in integrated plasma modelling, in which the many physical processes are coupled to predict reactor-level behaviour and to optimise operational scenarios of fusion power plants. The considerable number of geometric and thermodynamic parameters, the interactions that influence the turbulence and the resolutions needed to accurately resolve these turbulent modes, makes direct numerical simulation for parameter space exploration computationally extremely challenging. However, this makes it suitable for surrogate modelling, where speed ups of up to 105 are possible making rapid scenario development a possibility. In this talk the integrated plasma modelling use-case will be introduced as well as the turbulence surrogate modelling efforts at UKAEA, including how the models are integrated into larger workflows.

Development of a ML-enhanced version of the ICON Earth System Model

Julien Savre - DLR

Research Scientist.

Abstract: Despite continuous improvements in the performance of Earth System Models (ESMs), systematic errors and modeling uncertainties still limit their ability to produce reliable climate projections. Whereas part of the observed biases can be explained by uncertainties associated with the representation of subgrid processes in the atmosphere (like convective clouds or turbulence) through so-called parameterizations, the benefits we can expect from further improving and complexifying these parameterizations remain small due to inherent limitations related to the coarse spatial resolutions employed by the models. In recent years, machine learning (ML) has emerged as a powerful tool to develop surrogate parameterizations informed with high-resolution model outputs or Earth observations, promising to significantly reduce long-standing biases and enhance the projection capabilities of current ESMs (Eyring et al., 2024, https://doi.org/10.1038/s41561-024-01527-w). This talk will report on recent achievements made towards the development of a hybrid, machine learning enhanced (MLe) version of the ICOsahedral Non-hydrostatic model (ICON-MLe) equipped with various ML-based parameterizations for cloud fraction, radiation or convection, and preliminary results obtained in AMIP-like simulations will be presented. The talk will also highlight the necessity to recalibrate hybrid ESMs using fast automated techniques (Grundner et al., 2025, https://doi.org/10.48550/arXiv.2505.04358), as well as the importance of developing methods that can be easily shared across models.

Active learning for Emulation

Christopher Sprague - Alan Turing Institute

Senior Research Associate at the Alan Turing Institute.

Abstract: High-fidelity simulations are essential across science and engineering, but their computational cost often limits how widely they can be used. Emulators – machine-learned surrogates of simulators – offer orders-of-magnitude speedups, but require substantial (and often expensive) training data from simulators. This talk will introduce AutoEmulate, the Alan Turing Institute’s open-source framework for automated emulation, and show how we integrate active learning to reduce the amount of simulation data needed. The talk will outline different types of active learning, explain why the stream-based setting is especially relevant, and share how we have built these methods into AutoEmulate for practical use.

Hybrid machine learning and data assimilation in weather forecasting

Alan Geer - ECMWF

Principal Scientist at ECMWF.

Abstract: Traditional numerical weather prediction (NWP) takes observations of the earth system, combines these observations with physical models using data assimilation in order to create "initial conditions", and then runs a physically-based model from these initial conditions to forecast future weather. Data-driven forecasts, based on machine learning, are now in some respects more accurate than those made by physical models. Attempts are also being made to replace the data assimilation process to create an "end-to-end" data-driven forecasting system that directly converts observations into weather forecasts. However, there are strong arguments for retaining the physical approach, including the Bayesian perspective that the most accurate posterior knowledge of the atmosphere should come from the combination of observationally-based knowledge with prior knowledge, meaning physical equations and other knowledge embedded in numerical models. But the quality of data-driven forecasts suggests that existing physical forecast models have substantial errors and need to be improved. This motivates a hybrid physical-empirical approach to weather forecasting, using empirical components to improve or augment the physical approach. One approach is to learn systematic error corrections that can be applied to the physical forecast model every few hours, or at the timestep level. A more "granular" hybrid approach targets the most uncertain components of a forecasting system, which are replaced or augmented by data-driven components, including more focused systematic error corrections, while other components may remain entirely physical. Both approaches are being explored inside the physically-based Integrated Forecasting System (IFS) of the European Centre for Medium-range Weather Forecasts. The error correction approach is in testing and significantly improves the quality of physically-based forecasts, bringing them closer in quality to data-driven equivalents, while retaining most benefits of the physical approach. A granular hybrid is already used operationally in the forward modelling of satellite radiance observations in sea ice areas, where no sufficiently accurate physical models and state information were previously available. This opens up a new area for physically-based NWP. In both cases, a major question is how best to train and maintain the empirical components, especially when the surrounding physical model versions are upgraded. This likely involves a combination of pre-training offline and fine-tuning online. The aim is to continue to optimise the machine learning components within the data assimilation system for weather forecasting, most likely decoupled from the process of obtaining initial conditions. From a technical point of view, the empirical components have been implemented in Python-based machine learning packages for the offline training and then rewritten in Fortran and C++ when used in NWP systems. As solutions to these issues become more developed, they may define the architecture of hybrid environmental prediction systems for decades to come.

ML-Emulator for Cloud Microphysics in ICON in a Realistic Climate Model Experiment

Caroline Arnold - Helmholtz-Zentrum Hereon

Research AI Consultant.

Abstract: As the spatial resolution of general circulation models (GCMs) increases and storms and clouds can be resolved, the underlying cloud microphysics still need to be parameterised. This is a known to be a major source of uncertainty in climate and weather simulations. The established parameterisations use bulk moment schemes, where the conversion of cloud and rain droplets is approximated through empirical relationships. Particle-based superdroplet simulations would provide a more accurate representation, but are typically not feasable for use in GCMs.

We couple SuperdropNet [1], an ML emulator for warm rain cloud microphysics trained on superdroplet simulations, to ICON [2]. Previously, we validated the coupled model in an idealised cloud microphysics test case and showed that SuperdropNet runs stable and provides reasonable precipitation patterns [3].

Now we move towards a realistic climate model experiment with 10 km horizontal resolution. We use historical greenhouse gas forcing and observed sea-surface temperatures as boundary conditions (a so called AMIP experiment) to compare if and how our hybrid model simulates more realistic precipitation compared to the widely used empirical two moment bulk scheme parameterisation.

Coupling SuperdropNet to ICON is achieved using FTorch. We are able to run ICON on 128 nodes on the CPU partition of the HPC system Levante with minimal overhead. Conditions beyond the training data range of SuperdropNet lead to negative feedback loops and impact the long-term stability of the coupled simulation. We implement physics-based constraints that improve stability. We run the hybrid model and the reference simulation for seven days and present first results. Furthermore, we test an autoregressive rollout of SuperdropNet that allows for longer GCM time steps and investigate how this impacts stability and results.

REFERENCES:

[1] Sharma, S., and Greenberg, D.: "SuperdropNet: a Stable and Accurate Machine Learning Proxy for Droplet-based Cloud Microphysics." JAMES, 2025.
[2] 10.35089/WDCC/IconRelease01
[3] Arnold, C., Sharma, S., Weigel, T., and Greenberg, D. S.: Efficient and stable coupling of the SuperdropNet deep-learning-based cloud microphysics (v0.1.0) with the ICON climate and weather model (v2.6.5), Geosci. Model Dev., 17, 4017-4029, 10.5194/gmd-17-4017-2024, 2024.

Differentiable programming for scientific computing with Enzyme and Julia

Valentin Churavy - University of Augsburg

Milan Klower - University of Oxford

Valentin is a PostDoc and RSE who has previously worked at JuliaLab
Milan is a NERC research fellow and developer of SpeedyWeather.jl.

To couple large-scale scientific codes with machine-learning approaches, we are required to obtain derivatives from scientific programs. These scientific programs are often written in a very different style compared to machine learning and thus have different requirements for the automatic differentiation framework. For example, Python+JAX requires a functional style, with no array mutation allowed, and supports only limited control flow. In contrast, scientific simulations such as climate models are typically written with both array mutation and complicated control flow, particularly in the context of multi-physics. This talk is going to introduce the different approaches for obtaining derivatives from computer programs and why compiler-enabled automatic differentiation à la Enzyme leads to true differentiable programming. We motivate the necessity of differentiable programming with examples from climate modeling, discussing current and ongoing projects using Oceananigans and SpeedyWeather..