Source-linked AI summary
Using machine learning to parameterize moist convection: potential for modeling of climate, climate change and extreme events
Paul A. O'Gorman, John G. Dwyer
TL;DR
This paper examines whether a machine-learning parameterization of moist convection remains stable and useful when coupled to a general circulation model, including for extreme precipitation and climate change. Using a random-forest parameterization trained on conventional-model output, it finds accurate control-climate simulations, captures extremes without special training, and reproduces climate change when training spans both climates or uses only warm-climate data.
Problem
The behavior of machine-learning moist-convection parameterizations when fully coupled in a general circulation model, and their usefulness for climate change and extreme-event simulations, remain poorly understood.
Method
The study trains a random-forest moist-convection parameterization on output from a conventional parameterization and evaluates it in idealized general circulation model simulations.
Results
The coupled model stably and accurately simulates the control climate and precipitation extremes without special extreme-event training; climate change is captured with training from both climates or only the warm climate.
Takeaways & Limitations
Random-forest parameterizations can preserve energy conservation and non-negative surface precipitation while supporting climate simulation and diagnostics of convection–environment interactions.
Takeaways & Limitations
The study uses an idealized aquaplanet GCM and learns from a conventional parameterization rather than high-resolution simulations; the tested climate change increases global-mean surface temperature by 6.5K.
Abstract
from arXiv · showhide
The parameterization of moist convection contributes to uncertainty in climate modeling and numerical weather prediction. Machine learning (ML) can be used to learn new parameterizations directly from high-resolution model output, but it remains poorly understood how such parameterizations behave when fully coupled in a general circulation model (GCM) and whether they are useful for simulations of climate change or extreme events. Here, we focus on these issues using idealized tests in which an ML-based parameterization is trained on output from a conventional parameterization and its performance is assessed in simulations with a GCM. We use an ensemble of decision trees (random forest) as the ML algorithm, and this has the advantage that it automatically ensures conservation of energy and non-negativity of surface precipitation. The GCM with the ML convective parameterization runs stably and accurately captures important climate statistics including precipitation extremes without the need for special training on extremes. Climate change between a control climate and a warm climate is not captured if the ML parameterization is only trained on the control climate, but it is captured if the training includes samples from both climates. Remarkably, climate change is also captured when training only on the warm climate, and this is because the extratropics of the warm climate provides training samples for the tropics of the control climate. In addition to being potentially useful for the simulation of climate, we show that ML parameterizations can be interrogated to provide diagnostics of the interaction between convection and the large-scale environment.
1 Introduction
The paper examines whether machine-learning parameterizations can be safely coupled to a GCM and used for climate, climate-change, extreme-event, and process-diagnostic applications. It uses a random forest trained on RAS output in idealized GCM tests.
- GCM parameterizations represent unresolved processes but contribute substantial uncertainty and bias because relevant scales cannot be fully resolved.
- Moist convection is a promising ML target because CRM data can provide training samples and conventional convective schemes are a major source of atmospheric-model uncertainty.
- Earlier work compared ML convective tendencies diagnostically but did not test a fully coupled ML parameterization in a GCM.
- The study trains a random forest on RAS temperature and specific-humidity tendencies, implements it in an idealized GCM, and compares simulations with RAS.
- Random forests are selected partly because their predictions preserve energy conservation and non-negative surface precipitation, while remaining more robust than the tested shallow ANNs.
- The experiments assess control-climate statistics, precipitation extremes, climate-change generalization, and diagnostics of convection–environment interactions.
2 Machine learning algorithm: random forest
A random forest learns the mapping from atmospheric profiles to convective tendencies through an ensemble of decision trees. Its averaging structure supports physical constraints and robustness when coupled to a GCM.
- A random forest is an ensemble of decision trees whose predictions average the predictions from individual trees.
- Training uses supervised regression to minimize mean squared error between known and predicted continuous outputs.
- Averaging non-negative training precipitation values guarantees non-negative predicted precipitation.
- Because moist enthalpy is linear in temperature and specific humidity and RAS conserves it, RF averaging preserves column-integrated energy conservation.
- Averaging predictions over training subsets limits deviations from the training outputs and may improve stability during GCM extrapolation.
3 Convection scheme and idealized GCM simulations
The experiments use an idealized aquaplanet GCM with RAS convection, a mixed-layer ocean, and simplified radiative and surface settings. Control and warm climates provide the simulation targets.
- RAS represents shallow and deep convection with an ensemble of entraining plumes and outputs temperature and specific-humidity tendencies.
- The idealized GCM couples a spectral atmospheric dynamical core to a 0.5 m shallow thermodynamic mixed-layer ocean.
- The model has no land, ice, seasonal cycle, or diurnal cycle, and uses perpetual-equinox insolation and simplified radiation.
- RAS simulations are spun up for 700 days and then integrated for 3300 days to generate training data; RF simulations use 700-day spinups and 900-day integrations.
- The control climate has a global-mean surface air temperature of 288K, while the warm climate has 295K after increasing longwave optical thickness by 1.4.
4 Training and validation of the random forest
The RF maps atmospheric profiles and surface pressure to convective temperature and humidity tendencies, with precipitation implied by the humidity tendency. Offline tests show accurate control-climate predictions and physically constrained precipitation and energy.
- The RF uses temperature and specific-humidity profiles plus surface pressure as 43 features and predicts 42 scaled convective-tendency outputs.
- The learned mapping is y = f(x), with x = (T, q, ps) and y = (cp∂T/∂t|conv, L∂q/∂t|conv).
- Using one column-based RF for all levels preserves energy conservation and non-negative precipitation, unlike separate level-specific RFs.
- 0.82 overall test-set R2 for the RF compares with 0.86 on the training dataset, with higher tendency skill where convective variability is large.
- 0.95 R2 and 7 × 10^-5 mm day^-1 mean bias describe the RF’s control-test precipitation performance.
- 0.77 tendency R2 and 0.93 precipitation R2 were obtained when the warm-climate RF was evaluated on warm-climate test data.
- 0.2 W m^-2 RMSE measures the RF’s small error in conserving column-integrated moist enthalpy on both training and test predictions.
5 Implementation in GCM and simulation of control climate
The RF parameterization was integrated into the GCM without numerical instability and reproduced important control-climate statistics, including precipitation extremes, relative to RAS.
- Implementation in GCM: The RF parameterization replaced RAS in the GCM without creating numerical instability.The RF also runs three times faster than RAS.
- Control-climate simulation: Figure 3 compares RAS, RF, and no-convection simulations using vertical profiles and latitudinal precipitation distributions.The plotted variables include tropical equivalent potential temperature, tropical eddy kinetic energy, mean precipitation, and extreme precipitation.
- Control-climate simulation: The RF correctly captures tropical equivalent potential temperature, tropical eddy kinetic energy, mean precipitation, and the 99.9th percentile of daily precipitation.These statistics are compared with the GCM simulation using RAS.
- Control-climate simulation: The RF GCM adequately simulates means, wind variances represented by eddy kinetic energy, and precipitation extremes.The assessment uses higher-order climate statistics in addition to mean quantities.
6 Climate change and training in different climates
RF climate-change performance depends on training coverage: combined-climate or warm-climate training reproduces the response, whereas control-only training fails. Warm-climate training generalizes because warmer-climate extratropical samples cover conditions relevant to the control tropics, although the tested warming is large and extrapolation remains a concern.
- Training strategies: Control-only training produces incorrect and much-too-large tropical and subtropical precipitation changes because warm-climate conditions are absent from its training data.The control-trained RF has no skill in warm-climate convective-temperature tendencies equatorward of roughly 25° latitude.
- Climate-change response: 6.5K global-mean surface warming is the climate-change perturbation tested in these simulations.The authors note that generalization might be better for a smaller climate change.
- Training strategies: Figure 6 compares precipitation-change responses across four training designs: separate, combined, control-only, and warm-only RF training.Responses are shown as precipitation percentage change normalized by the change in zonal- and time-mean surface-air temperature.
- Training strategies: Warm-only training surprisingly gives good climate-change results because higher-latitude warm-climate samples represent control-climate tropical conditions.Extratropical warm-climate training predicts control-climate tropical tendencies well but fails for warm-climate tropical tendencies.
7 Feature importance of convection and sensitivity to perturbations
The RF parameterization can be interrogated through linear-response functions and feature importance to diagnose how temperature and humidity at different levels influence convection.
- Feature importance: Vertically integrated feature importance is 0.75 for specific humidity, 0.24 for temperature, and 0.01 for surface pressure.
- Diagnostic applications: The RF can efficiently generate linear-response diagnostics and feature-importance profiles without requiring additional perturbed convection-resolving-model simulations.
- Linear-response function: The RF generates linear-response functions by perturbing temperature, specific humidity, and surface pressure inputs and measuring resulting precipitation changes.Perturbations are applied to non-zero-precipitation test samples, with sensitivities rescaled to represent a 50 hPa-deep input perturbation.
- Linear-response function: Surface precipitation increases with moistening, especially at lower levels, while upper-level moistening above σ = 0.7 produces nearly zero sensitivity.The lower-level response is consistent with moisture increasing parcel buoyancy through initial moisture and entrainment.
- Linear-response function: Near-surface warming increases precipitation, whereas warming higher in the atmosphere decreases it more strongly.
- Feature importance: Feature importance measures the total reduction in tree mean-squared error attributable to each input feature and includes both convection occurrence and strength.Unlike linear-response functions, it does not require small perturbations and allows comparison between variables with different units.
8 Replacing both the large-scale condensation and convection schemes
The study also replaces both moist convection and large-scale condensation with one RF, using modified inputs, scaling, and sampling to represent their combined tendencies.
- Training design: The combined RF uses relative humidity as an input, applies stronger upper-tropospheric output weighting, and removes cosine-latitude sampling weights.These changes target large-scale condensation’s sensitivity to saturation and its importance at higher latitudes.
- Offline and coupled evaluation: The combined RF predicts test-dataset tendencies and precipitation accurately, with overall R2 values of 0.83 for tendencies and 0.93 for precipitation.
- Offline and coupled evaluation: Replacing both schemes with the combined RF produces accurate simulations of the control climate in the GCM.
9 Conclusions
In idealized GCM tests, RF-based convection produces stable climate simulations and captures precipitation extremes, while climate-change performance depends on training-climate coverage. The RF also provides diagnostics of convection–environment interactions, but the conclusions remain bounded by the idealized setup.
- Conclusions: The RF parameterization produces robust, accurate control-climate simulations and preserves physical constraints such as energy conservation through its decision-tree structure.It also simulates extreme precipitation events without specialized training on extremes.
- Conclusions: Climate change is accurately simulated when training includes control and warm climates, but not when training uses only the control climate.
- Conclusions: Training only on the warm climate nevertheless captures climate change because warmer-climate extratropical samples represent control-climate tropical conditions.The reported asymmetry relates to internal temperature variability, meridional temperature gradients, and the greater tropical importance of moist convection.
- Conclusions: Interrogating the RF yields linear-response functions and feature-importance measures for studying environmental controls on convection.Feature importance can also be separated into effects on convection occurrence and intensity.
- Scope and limitations: The conclusions come from an idealized aquaplanet GCM and training on a conventional parameterization rather than high-resolution simulations.Applying the approach to resolved convection introduces additional processing and interpretation challenges.