The Institute of Computing for Climate Science (ICCS) at the University of Cambridge is a cross-disciplinary centre for computational science, data science and machine learning, focusing on the application of these disciplines in climate science research. ICCS will host the 29th European workshop on Automatic Differentiation (EuroAD29) on 29 and 30 September 2026. What's the link here – what is automatic differentiation and why is it relevant for the climate community?
Senior Research Software Engineer Dr Joe Wallwork explains.
What is AD and what does it enable?
Automatic differentiation (AD, also known as Algorithmic Differentiation) is the name given to a family of technologies which take a piece of scientific code and compute derivatives of mathematical expressions that it involves. This allows scientists to get more out of their application software: as well as estimating values for quantities of interest, the ability to compute derivatives unlocks the door to optimisation techniques, sensitivity analysis, and uncertainty quantification.
Furthermore, the computation of derivatives is also key to the back-propagation operation which lies at the heart of modern machine learning methods. Clearly, the ability to compute derivatives of code is a key enabling technology. Derivatives of code can also be computed by hand or with numerical approximations but AD allows for the computation of exact derivatives, while avoiding lengthy, tedious and error-prone manual calculations.
Relevance for climate modelling
As mentioned above, AD can be used to compute derivatives for optimisation techniques. Climate modelling is based on Earth System Models (ESMs) – large-scale scientific models, which attempt to capture all of Earth's physical processes and the myriad ways in which they interact over millennia. Given the scale of the challenge, ESMs tend to involve many unknown parameters, representing unknown science or processes occurring on scales smaller than the model resolution, for example. Effective tuning of such parameters (i.e. model calibration) is key to achieving model accuracy and AD provides the ability to do this in an automated fashion.
Differentiable programming
Differentiable programming is an increasingly discussed topic in climate research. Traditional AD tools have typically been standalone pieces of software which are either embedded into an application or used to generate derivative code, which is then included in the build system. The process of applying AD tools to climate models can be time-consuming and less automatic than one might hope, given the large-scale nature of such codebases. Differentiable programming is a different approach, where the model itself is written in a fully differentiable framework. In such a framework, the challenge of model calibration becomes more tractable.
A popular approach for differentiable programming is to use the JAX framework, although its adoption for a climate model would likely require rewriting from scratch – arguably an even bigger undertaking than applying a traditional AD tool. Another approach is posed by Enzyme, which performs AD on the low-level LLVM language and allows the values to be exposed in several modern programming languages, reducing the extent to which the model needs to be modified. Work is currently underway at ICCS to provide Fortran bindings for Enzyme, with the goal of facilitating differentiable programming in the many climate models that are written in Fortran.
EuroAD29
The EuroAD workshop series has run regularly since 2005, providing a venue for researchers and software engineers who develop theory and software related to AD to share ideas, discoveries and inventions. The topics covered include new methods, software developments, inter-comparison studies, and – in recent years – AD in machine learning.
The 29th edition in Cambridge will continue this tradition, but also include breakout sessions in which we will discuss current open challenges and how we can work together to address them. Attendees from all career stages are welcome, particularly PhD students, early career researchers and software engineers.
Registration for EuroAD29 is now open. We look forward to seeing you there.