Swedish Joint Project for Next-Generation Modeling of Cloud Feedbacks and Climate Change using AI

Project Leader and Principal Investigator: Dr V. T. J. Phillips, Department of Physical Geography and Ecosystem Science, Lund University, Lund, Sweden

Co-Investigator: M. Ohlsson, Department of Astronomy and Theoretical Physics,

Lund University, Lund, Sweden

Co-Investigators: T. Landelius, SMHI, Norrköping, Sweden

Stakeholders: S. Montin, Energiforsk, Sweden; M. Billstein, Vattenfall, Sweden

Project: Dnr 2020-03406 (“AI i klimatets tjänst” call)

Reporting Period: 2020–2025

Overview

The project is entitled “Next-Generation Modeling of Cloud Feedbacks and Climate Change using AI: Implications for Alternative Energy”. It was led by Lund University and funded in Autumn 2020 by Sweden’s Innovation Agency (Vinnova). The project lasted about 4 years (November 2021-2025, including a year of no-cost extension).

This joint project involved four institutions in Sweden:

· SMHI, in Norrkoping;

· Energiforsk;

· Vattenfall;

· Department of Physical Geography and Ecosystem Science, Lund University, in Sweden.

Clouds play a key role in controlling climate sensitivity, and traditional global models use very coarse approximations. The project aimed to develop new AI-based technology to improve representations of clouds in climate models. While full implementation in a global climate model was not achieved due to numerical and computational challenges, the project successfully developed AI schemes for cloud prediction, created extensive training datasets, and generated methodological insights that will support future work on cloud-radiation feedbacks and improved climate predictions.

Conventional climate models must treat clouds and related physical processes statistically since they are too small spatially to resolve on the numerical grid. From Sherwood et al. (2014).

AI Approach with Deep Learning

The plan was to apply a neural network to predict cloud properties from large-scale conditions resolved by the global model. Any neural network consists of layers of neurons, each receiving a signal and sending another signal to other neurons in the next layer. The strength of the connections adjusts during learning. A neural network can be trained with specimen datasets to predict an output from given inputs. Once trained, the outputs are predicted by the combination of input values.

In the project, detailed simulation of an observed storm provided the initial training dataset. The neural network, once trained, was to be applied in every grid-box of the global model. The type of neural network to be applied is a ‘deep learning’ convolutional scheme that involves both time and space with a memory of past cloud conditions (Bai et al. 2018).

In short, this AI scheme is applied in every grid-box of the global model to treat the unresolved small-scale processes such as clouds, and potentially also, radiation.

Neural networks involve layers of neurons, each being connected to other neurons in the next layer.

Cloud Modelling

The aerosol-cloud model (AC) predicts the microphysical, dynamical and radiative properties of clouds (Waman et al. 2023). AC includes at least 7 chemical species of aerosol: primary biological aerosols, non-biological insoluble organics, soluble organics, sea-salt, ammonium sulphate, mineral dust and soot. Interstitial and immersed/embedded components of each aerosol species are predicted in cloud and precipitation. The model predicts active IN and cloud condensation nuclei (CCN) concentrations from the chemistry, sizes and loading of aerosols.

An observed case of a storm in Oklahoma, USA, is simulated by AC and after verifying accuracy of the simulation, the dataset is used to train a ‘deep learning’ form of the neural network.

A typical meteorological research aircraft for sampling clouds in field campaigns such as MC3E.

Aircraft Measurements

An observed case of clouds was used to validate the AC model. Aircraft flew through clouds measuring ice and drop concentrations while radar and ground-based measurements monitored macrophysical quantities (e.g., surface precipitation). National Center for Atmospheric Research in USA helped to provide the data from past field campaigns:

· MC3E (funded by DoE/NASA): Warmer-based clouds in precipitating deep convective

systems were sampled by aircraft with probes (e.g. optical probes with lasers to image and count individual hydrometeors) during May 2011 over Oklahoma;

Comparison of the 3D detailed simulation (not the AI scheme) of a storm observed in MC3E with coincident aircraft observations, for the ice concentration. Generally, ice concentrations are the most challenging type of cloud property to predict correctly, as many processes of ice initiation must be represented. From simulations described by Patade et al. (2022).

Progress in Project

A postdoctoral scholar worked on the project from late 2021 to late 2024. Two additional scholars joined in late 2023 and early 2024 to assist with the development of the AI scheme and the creation of the training dataset by embedding the Atmospheric Cloud (AC) model in every grid box of a global climate model, following our in-house “super-parameterization” (SP) approach.

Year 1: Offline AI development and testing

During the first year, the postdoctoral researcher created the initial neural-network architecture (“AI scheme”) and applied it offline to simulate the MC3E case — an observed convective storm — for comparison with a detailed cloud-model simulation.

The three-dimensional cloud simulation (see illustrative diagram) provided the training data for the neural network. Offline tests showed that the AI could reproduce cloud evolution, including heating and moistening tendencies, with reasonable fidelity to the full cloud model.

Year 2: Model refinement and integration into the global framework

In the second year, the AI scheme was refined through several key improvements:

• predicting logarithms or increments of selected variables to stabilize learning;
• rescaling outputs to conserve energy and mass;
• implementing dropout and normalization layers to improve generalization.

Preliminary training of the AI scheme with a dataset, consisting of 1 week of a global simulation with the public version of SP-CAM, was done. This was intended to enable a present-day climate simulation.

The figure below shows the offline tests of the AI scheme with a time series of global model output not used in the training. Inputs to the AI scheme were constrained to match the global model output while the prediction for the current time-step was compared with the time series value. The correlation coefficient was about 80 to 100% and the root mean square (RMS) fractional error was about 0.3 for the thermodynamic variables.

Correlation coefficient and root mean square error for all variables in the simple TCN model.

Work began on integrating the AI scheme into the Community Atmospheric Model (CAM) from NCAR (USA). This required resolving significant technical challenges, such as performing mixed-language compilation (FORTRAN–Python) and ensuring compatibility between data formats. Training with the global dataset was initiated to enable a present-day climate simulation.

Two theoretical papers by Phillips (2022, 2024), although not AI-based, informed the selection of physical input variables for the neural network by establishing relationships between warm-rain and ice processes in clouds.

Stakeholder meetings were held twice annually (e.g., May 2023 and November 2023) with representatives from SMHI and renewable-energy industries. These meetings provided external feedback on the scientific progress and potential applications of AI-accelerated climate modeling.

Relevant publications acknowledging Vinnova support include Waman et al. (2023), Zhao et al. (2023), and Gupta et al. (2023).

Year 3: Global model implementation and stability challenges

In the third year, the neural network was implemented within the global model, but numerical instability arose after approximately one week of simulation. The instability originated from super-adiabatic lapse rates forming locally and spreading horizontally, eventually causing the model to crash.

Analysis indicated that when the AI received inputs outside its training domain (e.g., unrealistically steep lapse rates from diurnal heating), it produced zero convective forcing. This prevented convection from removing the excessive heating, leading to runaway warming near the surface.

Several attempted remedies — such as artificial removal of super-adiabatic profiles and additional training — failed to resolve the problem. A promising solution was identified: to generate a “rough” training dataset using SP-CAM simulations that include artificial warm bubbles, so that the AI learns to recognize and suppress unrealistic lapse rates naturally. However, the postdoctoral contract ended before this could be implemented.

In parallel, coding for the full SP embedding of the cloud model in the global model was completed, but excessive memory requirements and machine compatibility issues prevented operational simulations.

As a result, the project’s final goal — a stable AI-driven global climate simulation — was not achieved within the project period. Nevertheless, the work produced a functioning AI prototype, extensive test-bed simulations, and critical insights into the physical and numerical behavior of AI parameterizations in climate models.

Summary of outcomes

While the main AI-accelerated climate simulations could not be completed, the project:

• Demonstrated the feasibility of offline neural-network prediction of cloud processes.
• Identified key stability issues when coupling AI modules with physical models.
• Produced transferable AI frameworks and datasets for future climate modeling.
• Built strong competence and international collaboration in AI-based atmospheric science.

Limitations, Challenges, and Future Directions

While the project made significant methodological advances, several challenges limited the achievement of the original goals. Numerical instabilities arose when integrating the AI scheme into the global climate model, preventing stable long-term simulations. Memory requirements and machine compatibility issues further constrained the operational implementation of the full AI-accelerated model. Consequently, direct evaluation of cloud-radiation feedbacks in global simulations and assessment of implications for renewable energy were not completed within the project period.

Despite these limitations, the work produced a functioning AI prototype, extensive test-bed datasets, and critical insights into the physical and numerical behavior of AI parameterizations in climate models. Future research can build on these foundations by refining training datasets, improving numerical stability, and scaling AI schemes for fully coupled global simulations. The project also established strong collaborations and built capacity in AI-based atmospheric science, providing a valuable platform for continued exploration of clouds and climate interactions.

Publications Acknowledging Vinnova Support

During the project period, ten peer-reviewed publications acknowledged Vinnova support. While the full AI-based global model was not coupled in long climate simulations, several papers report results directly related to the project’s intermediate steps, including single-column model (SCM) simulations and cloud test-bed cases. These publications document the development of the prototype AI framework, training datasets, and detailed cloud simulations that form the foundation for future AI-integrated climate modeling.

Additionally, the postdoctoral scholar supported for all years of the project, S. G. Patade, provided a technical report. It describes the offline development and evaluation of the neural network representing clouds.

Technical Report by the Senior Postdoctoral Scholar in the Project:

Patade, S. G., & Phillips, V. (2025). Final Report on AI Methods for Cloud Modeling (Vinnova Project 2020-03406). Zenodo. https://doi.org/10.5281/zenodo.17367602

SCM/MC3E Test-Bed Relevant Papers:

1. D. Waman, A. Deshmukh, A. Jadav, S. Patade, M. Gautam, and V. T. J. Phillips, “Mechanisms for indirect effects from solid aerosol particles on continental clouds and radiation.” J. Atmos. Sci., in press (2025). [MC3E test-bed relevant]
2. A. Jadav, D. Waman, C. S. Pant, S. Patade, M. Gautam, V. T. J. Phillips, A. Bansemer, D. Barahona, and T. Storelvmo, “An improved convection parameterization with detailed aerosol-cloud microphysics for a global model.” J. Atmos. Sci., 82, 197–231 (2025). [SCM simulation of MC3E]
3. D. Waman, A. Deshmukh, A. Jadav, S. Patade, M. Gautam, V. T. J. Phillips, A. Bansemer, and J. Jakobsson, “Effects from time dependence of ice nucleus activity for contrasting cloud types.” J. Atmos. Sci., 80, 2013–2039 (2023). [MC3E test-bed relevant]
4. S. Patade, D. Waman, A. Deshmukh, A. K. Gupta, A. Jadav, V. T. J. Phillips, A. Bansemer, J. Carlin, and A. Ryzhkov, “The influence of multiple groups of biological ice-nucleating particles on microphysical properties of mixed-phase clouds observed during MC3E.” Atmos. Chem. Phys., 22, 12055–12075 (2022). [MC3E test-bed relevant]
5. A. K. Gupta, D. Waman, A. Deshmukh, A. Jadav, S. Patade, V. T. J. Phillips, A. Bansemer, J. A. Martins, and F. Goncalves, “The microphysics of the warm-rain and ice-crystal processes of precipitation in simulated continental convective storms.” Commun. Earth Environ. 4, 226 (2023). [MC3E test-bed relevant]

Full List of Publications:

1. D. Waman, A. Deshmukh, A. Jadav, S. Patade, M. Gautam, and V. T. J. Phillips, “Mechanisms for indirect effects from solid aerosol particles on continental clouds and radiation.” J. Atmos. Sci., in press (2025). [MC3E test-bed relevant]
2. A. Jadav, D. Waman, C. S. Pant, S. Patade, M. Gautam, V. T. J. Phillips, A. Bansemer, D. Barahona, and T. Storelvmo, “An improved convection parameterization with detailed aerosol-cloud microphysics for a global model.” J. Atmos. Sci., 82, 197–231 (2025). [SCM simulation of MC3E]
3. V. T. J. Phillips, “A theory for the balance between warm-rain and ice-crystal processes of precipitation in mixed-phase clouds.” J. Atmos. Sci., 81, 317–339 (2024).
4. G. Sotiropoulou, A. Lewinschal, P. Georgakaki, V. T. J. Phillips, S. Patade, A. Ekman, and A. Nenes, “Sensitivity of Arctic clouds to ice microphysical processes in the NorESM2 climate model.” J. Clim., 37(16), 4275–4290 (2024).
5. M. Gautam, D. Waman, S. Patade, A. Deshmukh, V. T. J. Phillips, M. Jackowicz-Korczynski, P. Smith, and A. Bansemer, “Fragmentation in graupel–snow collisions: new formulation from field observations.” J. Atmos. Sci., 81, 2149–2164 (2024).
6. A. K. Gupta, D. Waman, A. Deshmukh, A. Jadav, S. Patade, V. T. J. Phillips, A. Bansemer, J. A. Martins, and F. Goncalves, “The microphysics of the warm-rain and ice-crystal processes of precipitation in simulated continental convective storms.” Commun. Earth Environ. 4, 226 (2023).
7. D. Waman, A. Deshmukh, A. Jadav, S. Patade, M. Gautam, V. T. J. Phillips, A. Bansemer, and J. Jakobsson, “Effects from time dependence of ice nucleus activity for contrasting cloud types.” J. Atmos. Sci., 80, 2013–2039 (2023).
8. X. Zhao, X. Liu, G. McFarquhar, S. Patade, V. T. J. Phillips, Y. Shi, and M. Zhang, “Important ice processes are missed by climate models in Southern Ocean mixed-phase clouds: bridging SOCRATES observations to model developments.” J. Geophys. Res. Atmos., 128(4), e2022JD037513 (2023).
9. S. Patade, D. Waman, A. Deshmukh, A. K. Gupta, A. Jadav, V. T. J. Phillips, A. Bansemer, J. Carlin, and A. Ryzhkov, “The influence of multiple groups of biological ice-nucleating particles on microphysical properties of mixed-phase clouds observed during MC3E.” Atmos. Chem. Phys., 22, 12055–12075 (2022).
10. V. T. J. Phillips, “Theory of in-cloud activation of aerosols and microphysical quasi-equilibrium in a deep updraft.” J. Atmos. Sci., 79, 1865–1886 (2022).

Bibliography

Bai, S., Kolter, J. Z., and V. Koltun, 2018: An Empirical Evaluation of Generic Convolutional and Recurrent Networks for Sequence Modeling, http://arxiv.org/abs/1803.01271

Gupta, A. K., D. Waman, A. Deshmukh, A. Jadav, S. Patade, V. T. J. Phillips, A. Bansemer, J. A. Martins and F. Goncalves, 2023: The microphysics of the warm-rain and ice crystal processes of precipitation in simulated continental convective storms. Commun Earth Environ. (Nature portfolio), 4, 226

Patade, S., Waman, D., Deshmukh, A., Gupta, A. K., Jadav, A., Phillips, V. T. J., Bansemer, A., Carlin, J., and A. Ryzhkov, 2022: The influence of multiple groups of biological ice-nucleating particles on microphysical properties of mixed-phase clouds observed during MC3E. Atmos. Chem. Phys., 22, 12055–12075

Phillips, V. T. J., 2022: Theory of in-cloud activation of aerosols and microphysical quasi-equilibrium in a deep updraft. J. Atmos. Sci., 79, 1865–1886

Phillips, V. T. J., 2024: A theory for the balance between warm rain and ice crystal processes of precipitation in mixed-phase clouds. J. Atmos. Sci., 81, 317–339

Waman, D., A. Deshmukh, A. Jadav, S. Patade, M. Gautam, V. T. J. Phillips, A. Bansemer, and J. Jakobsson, 2023: Effects from time dependence of ice nucleus activity for contrasting cloud types. J. Atmos. Sci., 80, 20132039

Zhao, X., X. Liu, G. McFarquhar, S. Patade, V. T. J. Phillips, Y. Shi, and M. Zhang, 2023: Important ice processes are missed by climate models in Southern Ocean mixed-phase clouds: bridging SOCRATES observations to model developments. J. Geophys. Res. Atmos., 128(4), e2022JD037513