AI4 Climate: Harnessing artificial intelligence to transform climate science
AI4 Climate explores and applies cutting-edge Artificial Intelligence (AI) and Machine Learning (ML) techniques to advance climate science and deliver improved climate information more efficiently.
AI4 Climate is funded by the UK Government’s Department for Science, Innovation and Technology (DSIT) through the International Science Partnerships Fund (ISPF) and sits within the Met Office’s National Capability AI (NCAI) Programme.
Vision and purpose
AI4 Climate aims to integrate AI/ML methods alongside traditional physics-based approaches to enhance climate modelling and projection. The programme seeks to:
-
deliver more accurate, locally relevant climate projections faster and more cost-effectively to support timely decision-making
-
support rapid responses to climate-related questions
-
enable global collaboration through open science and shared innovation
Research areas:
AI Climate downscaling
To support faster and more locally relevant climate projections, this work package applies AI techniques to emulate high-resolution regional climate simulations. These simulations are essential for understanding climate risks and informing adaptation strategies but are often limited by computational cost. AI4 Climate addresses this by learning relationships between global and local climate patterns, enabling broader and more efficient downscaling. The team is developing benchmark datasets and evaluation metrics for the UK and international regions, alongside researching an open-access AI model adaptable to diverse climates, advancing the goal of scalable, cost-effective climate information.
Data-driven climate modelling
In pursuit of rapid and flexible climate insights, this work package explores models trained directly on observational and simulation data. These data-driven approaches offer low-cost alternatives to traditional climate models, especially valuable in resource-constrained settings. Applications include seasonal-to-decadal prediction and emulation of Earth system components. The work includes exploring the development of deployment frameworks, adapting existing models for climate applications, and co-developing capabilities with international partners, supporting AI4 Climate’s mission to accelerate climate science through innovative AI/ML integration.
Hybrid physical model /AI approach
This work package exemplifies the fusion of AI with physics-based modelling, replacing uncertain components of traditional models with ML. By training AI systems on high-resolution simulations, the team is aiming to develop hybrid models that retain trusted physical dynamics while improving accuracy and efficiency. These models feedback within the simulation itself, offering a transformative approach to climate modelling. The work supports AI4 Climate’s vision by enhancing model fidelity and enabling faster responses to climate questions, while building capacity in partner institutions.
Dataset creation and curation
Central to enabling AI/ML innovation is the availability of high-quality training data. This work package delivers curated, sustainable datasets for training, testing, and validating ML-based climate models. It includes global 10km Numerical Weather Prediction forecasts, ensemble datasets, and tools for cataloguing and code refactoring. These resources underpin multiple AI4 Climate objectives, including downscaling, hybrid modelling, and urban-scale prediction, ensuring that data infrastructure supports scalable and reproducible climate science.
K-Scale simulations for AI training
This work led by the University of Leeds generates kilometre-scale simulations over large domains to improve representation of fine-scale climate processes. Using advanced modelling frameworks like the Unified Model and LFRic (named after Lewis Fry Richardson), the team produces high-resolution datasets that serve as training foundations for AI models. The simulations support hybrid modelling strategies and contribute to international collaborations, reinforcing AI4 Climate’s aim to accelerate predictive capability and reduce long-term computational costs.
Urban scale modelling
To address climate challenges in cities, this work package develops ML-based downscaling techniques that translate coarse-resolution model outputs into detailed urban-scale climate information. The team is building pipelines for prediction and evaluation, leveraging new observational networks and collaborating with partners in cities like London, Paris, Delhi, and Singapore. The project also aims to deliver a common software framework and benchmarking tools, with a focus on spatial transferability and global applicability. This work directly supports AI4 Climate’s goals of rapid, locally relevant projections and international collaboration, particularly in Official Development Assistance (ODA) regions.
Technical development of evaluation tools
To ensure trust and transparency in AI-driven climate science, this work package modernizes legacy workflows into a unified Climate Model Evaluation Workflow (CMEW). Built around the Earth System Model Evaluation Tool (ESMValTool), the framework supports diagnostics tailored to emerging ML models and integrates third-party observations via a generic Application Programming Interface (API). By delivering open-source tools and scalable workflows, the project empowers scientists to assess model performance and internal consistency, advancing AI4 Climate’s commitment to open science and responsible innovation.
Evaluation tools
In collaboration with the University of Reading, this work develops interoperable tools to evaluate both physics-based and ML climate models. The focus is on building portable Python-based workflows and extending ESMValTool to support new metrics and data formats. These tools help operational teams and researchers assess model accuracy and coherence, supporting AI4 Climate’s objectives of reproducibility, innovation, and global impact.
Machine learning for seasonal to decadal predictions
This work explores models trained to produce initialised climate predictions from months to a few years ahead for adaptation to climate variability and change. We have demonstrated that global seasonal predictions are possible using these new methods, and we are investigating how skilful they are and whether these new methods produce any unwanted side effects. The work involves:
-
investigating levels of skill in initialised predictions
-
designing hybrid systems using the best combinations of physics based and AI methods.
Global impact
Aligned with ISPF objectives, AI4 Climate:
-
supports international partnerships by co-developing AI tools with global research institutions and ODA partners
-
advances sustainable development through accessible, low-cost climate modelling tools
-
enables transformative technology by applying AI/ML to climate science
-
strengthens UK leadership in science, technology, and innovation
-
promotes open science and responsible data sharing
-
contributes to a resilient planet by improving understanding of extreme weather and climate risks
-
builds tomorrow’s talent by developing AI/ML skills across the UK and partner countries.
Partnerships
AI4 Climate collaborates with institutions including:
-
University of Leeds
-
University of Reading
-
University of Bristol
-
Centre National de Recherches Météorologiques (CNRM) France
-
University of the Witwatersrand, South Africa
-
The Allen Institute, USA.
This webpage was published in October 2025.