Parameterizations in Earth System Models (ESMs) are subject to biases and uncertainties arising from subjective empirical assumptions and incomplete understanding of the underlying physical processes. Recently, the growing representational capability of machine learning (ML) in solving complex problems has spawned immense interests in climate science applications. Specifically, ML-based parameterizations have been developed to represent convection, radiation and microphysics processes in ESMs by learning from observations or high-resolution simulations, which have the potential to improve the accuracies and alleviate the uncertainties. Previous works have developed some surrogate models for these processes using ML. These surrogate models need to be coupled with the dynamical core of ESMs to investigate the effectiveness and their performance in a coupled system. In this study, we present a novel Fortran-Python interface designed to seamlessly integrate ML parameterizations into ESMs. This interface showcases high versatility by supporting popular ML frameworks like PyTorch, TensorFlow, and Scikit-learn. We demonstrate the interface’s modularity and reusability through two cases: a ML trigger function for convection parameterization and a ML wildfire model. We conduct a comprehensive evaluation of memory usage and computational overhead resulting from the integration of Python codes into the Fortran ESMs. By leveraging this flexible interface, ML parameterizations can be effectively developed, tested, and integrated into ESMs.

Recent studies reveal an anti-correlation between global cloud feedback (CF) and effective radiative forcing due to aerosol-cloud interaction (ERFaci) in climate models, but its physical plausibility remains uncertain. Here we investigate whether different turbulence representations, specifically through perturbing turbulence parameters, contribute to this relationship over the global ocean using an E3SMv2 perturbed parameter ensemble. The anti-correlation appears only in the tropical ascent regime. In the Northern Hemisphere midlatitude and high latitude regimes, there is no significant correlation, and in the tropical marine low cloud and Southern Ocean regimes, the correlation is positive. These opposite correlations are primarily driven by opposing CF responses to perturbed parameters. We find that the mean-state turbulent mixing strength affects both CF and ERFaci, enabling strong correlations in certain regimes. This study highlights the complex linkages between CF and ERFaci through turbulent processes across diverse cloud regimes.

In recent years, machine learning (ML) models have been used for improving physical parameterizations of general circulation models (GCMs). A significant challenge of integrating ML models into GCMs is the online instability when they are coupled for long-term simulation. In this study, we present a new strategy that demonstrates robust online stability when the entire physical parameterization package of a GCM is replaced by a deep ML algorithm. The method uses a multistep training scheme of the machine learning model with experience replay in which the memory of physical tendencies from the training dataset and the ML algorithm’s own output at the previous time step are used in the training. The physics memory improves the accuracy of the machine learning model, while the experience replay constrains the amplification of cumulative errors in the online coupling. The method is used to train the whole physical parameterization package for the Community Atmosphere Model version 5 (CAM5) with data from its Multi-scale Modeling Framework (MMF) high resolution simulations. Three 6-year online simulations of the CAM5 with the ML physics package at operational spatial resolution with real-world geography are presented. The simulated spatial distributions of precipitation, surface temperature and zonally averaged atmospheric fields demonstrate overall better accuracy than that of the standard CAM5 and benchmark model even without the use of additional physical constraints or tuning. This work is the first to demonstrate a solution to address the online instability problem in climate modeling with ML physics by using experience replay.

Particle-resolved direct numerical simulations (PR-DNS) play an increasing role in investigating aerosol-cloud-turbulence interactions at the most fundamental level of processes. However, the high computational cost associated with high resolution simulations poses considerable challenges for large domain or long duration simulation using PR-DNS. To address these issues, here we present a digital twin model of the complex physics-based PR-DNS developed by use of the data-driven Fourier Neural Operator (FNO) method. The results demonstrate high accuracy at various resolutions and the digital twin model is two orders of magnitude cheaper in terms of computational demand compared to the physics-based PR-DNS model. Furthermore, the FNO digital-twin model exhibits strong generalization capabilities for different initial conditions and ultra-high-resolution without the need to retrain models. These findings highlight the potential of the FNO method as a promising tool to simulate complex fluid dynamics problems with high accuracy, computational efficiency, and generalization capabilities, enhancing our understanding of the aerosol-cloud-precipitation system.

Particle-resolved direct numerical simulations (PR-DNS) are crucial for unraveling the intricate interplay of aerosol-cloud-turbulence processes. However, such models are challenged by the huge computational cost due to the extremely high resolution. Our prior work showcased that leveraging machine learning emulators could slash computational expenses by two orders of magnitude while maintaining remarkable precision for dynamic fields, and exhibited generalizability across diverse initial conditions and at super-resolution scales without retraining the emulators. Building upon this foundation, this work extends the emulator’s application to thermodynamic in the two spatial dimensions and droplet fields in the three spatial dimensions. Furthermore, to enhance the robust generalizability of the emulator for different initial values and super resolution, we introduce a novel multi-initial learning approach for the neural operator method. For the droplet fields, we introduce a novel loss function tailored to assess distribution differences using the Mallows distance, focusing particularly on droplet size distributions. Our findings indicate that the machine learning emulators hold promising potential to effectively mimic numerical PR-DNS simulations, thereby significantly advancing our understanding of the complex interactions within aerosol-cloud-turbulence processes.