Periodic Reporting for period 1 - AI4PEX (Artificial Intelligence and Machine Learning for Enhanced Representation of Processes and Extremes in Earth System Models)
Okres sprawozdawczy: 2024-04-01 do 2025-09-30
The progression of global warming poses challenges that demand urgent, science-driven solutions. Earth system models (ESMs), vital for predicting climate change, possess inherent uncertainties in their predictions. The EU-funded AI4PEX project's primary aim is to address those uncertainties by enhancing ESMs. The project will address key contributors to uncertainties by employing advanced machine learning and AI. By fusing observations with these cutting-edge technologies, AI4PEX seeks to learn and accurately model the complex processes which degrade our confidence in climate predictions. Using a multidisciplinary approach, the project aims towards a breakthrough in Earth system model accuracy, crucial for anticipating future climate extremes and their societal impacts.
AI4PEX integrates developments on machine learning (ML)-enhanced observations, modeling, uncertainty quantification, model evaluation and extremes, to advance knowledge and understanding of the Earth system and modeling in order to improve the representation of processes underpinning key feedbacks and uncertainties.
Beta versions of ML-enhanced observational data streams strengthen process understanding and model constraints. A database of low-level cloud patterns, robustly distinguishing stratocumulus-to-cumulus regimes, provides cloud classes for later model evaluation and parameterisation. An ML approach was expanded to reconstruct 4D fields of temperature and dissolved inorganic carbon from EO and in situ data, delivering beta products for ocean heat and carbon transport. AI4PEX also delivered beta products for global information-content and persistence metrics, advances toward empirically constrained global energy flux products, a neural-network river-discharge and routing prototype, and groundwork for dataset of biomass dynamics.
Research focused on ML-enhanced process representations in atmosphere, ocean and land models. For the atmosphere, the ICON-A-MLe configuration coupled an equation discovery cloud cover scheme with automatic tuning, producing 20-year AMIP runs that reduced cloud and top-of-atmosphere radiation biases, while a scale-aware NN cloud scheme was implemented in HadGEM3-GC5.0/UKESM alongside emulators. For the ocean, mesoscale eddy closures in NEMO were benchmarked and an emulator for surface chlorophyll and carbon fluxes was developed. For the land, hybrid JSBACH parameterisations of stomatal conductance and photosynthesis reduce carbon and water-flux biases.
AI4PEX focused on developing uncertainty quantification (UQ) for ML and hybrid modelling. Methodological advances included enhanced UQ in variational data assimilation, with stochastic priors, neural state estimation under irregular sampling, and differentiable data assimilation (DA) frameworks. Activities developed and tested ML emulators and stochastic parameterisations, established a benchmarking framework for hybrid ESMs, and investigated transfer of NN emulators from offline to online coupling. In the ocean and land, emulation and variational inference approaches were extended to biogeochemistry and hybrid soil-carbon models.
AI4PEX advanced ML-based evaluation techniques, including dynamic mode decomposition and physics-aware Koopman operators. ML-enhanced representations of the atmosphere have been partially analysed for 25-year simulations on the HadGEM3-GC5.0 model, noting global consistency but regional biases. Substantial improvements were made for the ML-enhanced ICON-A-MLe model, showing bias reduction in cloud cover and radiative fluxes.
AI4PEX made progress in detecting and attributing changes in climate extremes to their dynamical and thermodynamical drivers. A causal representation-learning framework is developed for dynamical adjustment and factorial analysis of temperature and precipitation trends, enabling robust separation of circulation-driven and thermodynamic components. Research has examined dynamical contributions to summer warming using multiple decomposition approaches, and refined probabilistic extreme event attribution by incorporating regional aerosol covariates. The application of automated differentiable architecture search to discover high-performing convolution NNs for precipitation downscaling of extreme events, showed that ML methods improve the representation of a global model compared to simple interpolation.
Beta versions of ML-enhanced observational data streams strengthen process understanding and model constraints. A database of low-level cloud patterns, robustly distinguishing stratocumulus-to-cumulus regimes, provides cloud classes for later model evaluation and parameterisation. An ML approach was expanded to reconstruct 4D fields of temperature and dissolved inorganic carbon from EO and in situ data, delivering beta products for ocean heat and carbon transport. AI4PEX also delivered beta products for global information-content and persistence metrics, advances toward empirically constrained global energy flux products, a neural-network river-discharge and routing prototype, and groundwork for dataset of biomass dynamics.
Research focused on ML-enhanced process representations in atmosphere, ocean and land models. For the atmosphere, the ICON-A-MLe configuration coupled an equation discovery cloud cover scheme with automatic tuning, producing 20-year AMIP runs that reduced cloud and top-of-atmosphere radiation biases, while a scale-aware NN cloud scheme was implemented in HadGEM3-GC5.0/UKESM alongside emulators. For the ocean, mesoscale eddy closures in NEMO were benchmarked and an emulator for surface chlorophyll and carbon fluxes was developed. For the land, hybrid JSBACH parameterisations of stomatal conductance and photosynthesis reduce carbon and water-flux biases.
AI4PEX focused on developing uncertainty quantification (UQ) for ML and hybrid modelling. Methodological advances included enhanced UQ in variational data assimilation, with stochastic priors, neural state estimation under irregular sampling, and differentiable data assimilation (DA) frameworks. Activities developed and tested ML emulators and stochastic parameterisations, established a benchmarking framework for hybrid ESMs, and investigated transfer of NN emulators from offline to online coupling. In the ocean and land, emulation and variational inference approaches were extended to biogeochemistry and hybrid soil-carbon models.
AI4PEX advanced ML-based evaluation techniques, including dynamic mode decomposition and physics-aware Koopman operators. ML-enhanced representations of the atmosphere have been partially analysed for 25-year simulations on the HadGEM3-GC5.0 model, noting global consistency but regional biases. Substantial improvements were made for the ML-enhanced ICON-A-MLe model, showing bias reduction in cloud cover and radiative fluxes.
AI4PEX made progress in detecting and attributing changes in climate extremes to their dynamical and thermodynamical drivers. A causal representation-learning framework is developed for dynamical adjustment and factorial analysis of temperature and precipitation trends, enabling robust separation of circulation-driven and thermodynamic components. Research has examined dynamical contributions to summer warming using multiple decomposition approaches, and refined probabilistic extreme event attribution by incorporating regional aerosol covariates. The application of automated differentiable architecture search to discover high-performing convolution NNs for precipitation downscaling of extreme events, showed that ML methods improve the representation of a global model compared to simple interpolation.
Current AI4PEX results show a strong potential to transform Earth system modeling and analysis of extremes. ML-based parameterisations for atmospheric processes reduce biases in cloud cover and radiative fluxes, as well as in the representation of extreme precipitation, in three European ESMs, thereby directly targeting the feedbacks that dominate uncertainty in climate sensitivity and impact. A diverse set of ML-based UQ approaches provides novel schemes to constrain parametric and structural uncertainties across components, while causal frameworks and methods for rigorous decomposition of circulation vs thermodynamic drivers of extremes offer sound basis for detection, attribution and model weighting. At the same time, advances in ML-driven reconstruction of EO datasets deliver novel and enhanced observational constraints that will underpin improved understanding of Earth system processes and support ESM development. Collectively, these developments contribute to reducing uncertainty in key feedbacks and provide more reliable projections, with full impact contingent on continued resolution of key scientific and technical challenges, such as:
- Supporting the transferability of ML components into process-based models and across modelling centres, minimising tooling incompatibilities for online ML-inference within ESM code infrastructures.
- Online learning and adaptive calibration are identified as both promising and technically demanding. AI4PEX laid theoretical and practical groundwork for online-learning, yet porting approaches to domain level or fully coupled ESMs remains a conceptual and structural challenge.
- Assessing and developing confidence in hybrid frameworks. Causal regularisation and causal representation learning already make a proof of concept for improving robustness to distribution shifts, attribution schemes and data biases. Future needs concern taking these tools to operational ESM model development and benchmark frameworks.
- Consolidation of datasets, benchmarks and protocols are prerequisites for broad impact.
- Supporting the transferability of ML components into process-based models and across modelling centres, minimising tooling incompatibilities for online ML-inference within ESM code infrastructures.
- Online learning and adaptive calibration are identified as both promising and technically demanding. AI4PEX laid theoretical and practical groundwork for online-learning, yet porting approaches to domain level or fully coupled ESMs remains a conceptual and structural challenge.
- Assessing and developing confidence in hybrid frameworks. Causal regularisation and causal representation learning already make a proof of concept for improving robustness to distribution shifts, attribution schemes and data biases. Future needs concern taking these tools to operational ESM model development and benchmark frameworks.
- Consolidation of datasets, benchmarks and protocols are prerequisites for broad impact.