Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Embedding machine-learnt sub-grid variability improves climate model biases

Daniel Giles, James Briant, Cyril J. Morcrette, Serge Guillas

TL;DR

The paper addresses persistent cloud- and precipitation-biases in coarse-resolution climate models by embedding sub-grid thermodynamic variability learned from high-resolution data into a coarse GCM. A trained Multi-Output Gaussian Process (MOGP) predicts columnar variability in temperature $T$ and specific humidity $Q$, and is coupled in-situ with the SPEEDY atmospheric GCM to perturb its thermodynamic profiles every 6 hours. Across a 10-year simulation, the hybrid SPEEDY+MOGP setup reduces precipitation RMSE globally by $18\%$ and in the tropics by $22\%$ relative to ERA5, outperforming naive stochastic baselines and providing physical insight via Hadley cell adjustments and lifted-index changes. A $+4\ \mathrm{K}$ SST experiment demonstrates the method’s potential to preserve cloud- and precipitation-change signals in a warming climate, highlighting its promise for scalable, uncertainty-aware sub-grid variability representations in climate models.

Abstract

The under-representation of cloud formation is a long-standing bias associated with climate simulations. Parameterisation schemes are required to capture cloud processes within current climate models but have known biases. We overcome these biases by embedding a Multi-Output Gaussian Process (MOGP) trained on high resolution Unified Model simulations to represent the variability of temperature and specific humidity within a climate model. A trained MOGP model is coupled in-situ with a simplified Atmospheric General Circulation Model named SPEEDY. The temperature and specific humidity profiles of SPEEDY are perturbed at fixed intervals according to the variability predicted from the MOGP. Ten-year predictions are generated for both control and ML-hybrid models. The hybrid model reduces the global precipitation bias by 18\% and over the tropics by 22\%. To further understand the drivers of these improvements, physical quantities of interest are explored, such as the distribution of lifted index values and the alteration of the Hadley cell. The control and hybrid set-ups are also run in a plus 4K sea-surface temperature experiment to explore the effects of the approach on patterns relating to cloud cover and precipitation in a warmed climate setting.

Embedding machine-learnt sub-grid variability improves climate model biases

TL;DR

The paper addresses persistent cloud- and precipitation-biases in coarse-resolution climate models by embedding sub-grid thermodynamic variability learned from high-resolution data into a coarse GCM. A trained Multi-Output Gaussian Process (MOGP) predicts columnar variability in temperature and specific humidity , and is coupled in-situ with the SPEEDY atmospheric GCM to perturb its thermodynamic profiles every 6 hours. Across a 10-year simulation, the hybrid SPEEDY+MOGP setup reduces precipitation RMSE globally by and in the tropics by relative to ERA5, outperforming naive stochastic baselines and providing physical insight via Hadley cell adjustments and lifted-index changes. A SST experiment demonstrates the method’s potential to preserve cloud- and precipitation-change signals in a warming climate, highlighting its promise for scalable, uncertainty-aware sub-grid variability representations in climate models.

Abstract

The under-representation of cloud formation is a long-standing bias associated with climate simulations. Parameterisation schemes are required to capture cloud processes within current climate models but have known biases. We overcome these biases by embedding a Multi-Output Gaussian Process (MOGP) trained on high resolution Unified Model simulations to represent the variability of temperature and specific humidity within a climate model. A trained MOGP model is coupled in-situ with a simplified Atmospheric General Circulation Model named SPEEDY. The temperature and specific humidity profiles of SPEEDY are perturbed at fixed intervals according to the variability predicted from the MOGP. Ten-year predictions are generated for both control and ML-hybrid models. The hybrid model reduces the global precipitation bias by 18\% and over the tropics by 22\%. To further understand the drivers of these improvements, physical quantities of interest are explored, such as the distribution of lifted index values and the alteration of the Hadley cell. The control and hybrid set-ups are also run in a plus 4K sea-surface temperature experiment to explore the effects of the approach on patterns relating to cloud cover and precipitation in a warmed climate setting.
Paper Structure (14 sections, 8 figures, 1 table)

This paper contains 14 sections, 8 figures, 1 table.

Table of Contents

  1. Data generation, coarse-graining and training design
  2. Global mean precipitation over 10-year simulation (1982-1992)
  3. Physical explanation for precipitation differences
  4. Hadley cell circulation
  5. Lifted index and extreme events
  6. Plus 4K sea-surface temperature simulation
  7. Discussion
  8. Online Methods
  9. MOGP
  10. Training
  11. Validation
  12. Two-way coupling of GCM and MOGP
  13. Experimental set up
  14. Naive stochastic perturbations

Figures (8)

  • Figure 1: Examples of MOGP predictions of temperature and humidity standard deviation on 00:00UTC 1 January and 12:00UTC 1 June 1987 at 925 hPa.
  • Figure 2: a) Location of the 80 limited-area models (LAMs) across the globe. Each LAM is then split into a 2x2 array, where coarse-graining is performed. An average or a standard deviation can be calculated over these patches, as shown by the shading which, as an example, represents the standard deviation of near-surface temperature [K] at 00Z on 1 Jan 2020. Examples of the 1.5 km surface temperature data are shown for b) the Arabian Sea and c) Southern India and Sri Lanka, while examples of their standard deviation are shown in d) and e).
  • Figure 3: Maps of precipitation error between the control (top left) and the hybrid experiment (top right) versus the ERA5 data for the same period. The weighted RMSE values are included in the titles. Bottom centre: the difference of the absolute errors of the control and hybrid runs.
  • Figure 4: Differences in mean precipitation, top-of-atmosphere outgoing long-wave radiation, top-of-atmosphere short-wave radiation and cloud cover between the hybrid and control SPEEDY simulations. Statistically insignificant results are masked.
  • Figure 5: Top Left: Zonal averages of the temperature over the limited area (168.75$^{\circ}$ W to 78.75$^{\circ}$ W ) in the Pacific for the control run. Bottom Left: Difference of the zonal averages of the temperature between the control and hybrid run, see top left subplot of Fig. \ref{['fig:field-diffs']} for the bounding box. Top Right: Lifted index histograms for both the control and hybrid simulation in central Africa (33.75$^{\circ}$ E, -1.856 $^{\circ}$ S). Bottom Right: Precipitation histogram with the zero counts removed.
  • ...and 3 more figures