Clouds persist as the most significant source of uncertainty in climate model predictions of climate change. In the most general terms, my research seeks to understand the dynamics of clouds and the mechanisms by which they change with climate. I build massively parallel, high resolution, observationally verified models that explicitly simulate the turbulent dynamics of clouds. These models serve as the numerical laboratory in which I conduct my research. Below are a few highlights of my work.

To get a qualitative feeling for the high resolution simulations we do, please see Visualization.

Using large eddy simulation driven by climate models as a numerical laboratory for studying cloud feedbacks

Collaborators: Tapio Schneider and Zhihong Tan

This is a project that uses large eddy simulations and climate models to understand cloud feedbacks in simulations that resolve the turbulent structure of clouds, feel the effects of changes in the large-scale circulation, and satisfy physically realizable surface and top of atmosphere energy balances.


Much of the uncertainty related to clouds comes about because their dynamics are not resolved at the resolution of current climate models. Thus cloud processes in climate models must be represented semi-empirically.

The approach we take is to simulate cloud feedbacks using large eddy simulations ( LES). The advantage to using LES, which are run at orders of magnitude higher resolution than climate models, is that the dynamics of clouds are directly represented in terms of well understood physical laws, like Newton's laws of motion and the laws of thermodynamics, along with physically motivated model closures. Therefore, LES avoid much of the empiricism inherent in climate models. However, due to their large computational expense, LES are necessarily limited area models. This means that LES do not naturally feel the changes in large-scale atmospheric dynamics that accompany climate change nor do they naturally satisfy physically realizable energy and mass balances. This makes direct application of LES to the problem of cloud feedbacks a challenge.

Large Eddy Simulations Driven by a Climate Model

To overcome these challenges we use a climate model to provide dynamically consistent forcing to the LES and require the LES to satisfy physically realizable surface and top of the atmosphere energy balances. Providing dynamically consistent large-scale forcing to the LES is accomplished by envisioning the LES domain as a single grid column of the climate model. The budget equations for the climate model grid column can be partitioned into source terms related to large-scale processes (those that are resolved in the climate model) and small-scale processes (those that are parameterized in the climate model) . The large-scale source terms are provided to the LES from the climate model and the small-scale terms are neglected as the LES will resolve these directly. Physically realizable surface and top of atmosphere energy balances are achieved by coupling the LES to a slab ocean model at the surface and using interactive radiative transfer within the LES that is consistent with the GCM. If care is taken to ensure that all large-scale budget terms are closed in a consistent manner, the LES evolve freely and reach physically realizable statistical steady states that do not require any nudging to mean states. Using LES to simulate how clouds change requires only imposing climate change on the climate model.

Statistically Steady-State LES

As a first step in a hierarchical approach to understanding climate feedbacks, we consider LES in idealized climates, namely aqua-planets, and with simplified cloud-radiation interactions. The forcing framework in no way precludes less idealized climates or cloud-radiation interactions, however these assumptions allow for a cleaner interpretation of the mechanisms of cloud changes.

In the aqua-planet climate model simulations we impose ocean heat fluxes, as shown in Figure 1, that induce a Walker Circulation. By imposing a Walker Circulation we create a region with suppressed deep convection that is conducive to the formation of boundary layer clouds. The black dots in Figure 1 show a transect of GCM grid points for which LES are performed. This transect is selected because it samples a transition from shallow boundary layer convection to deep convection. The forcing terms provided to the LES are computed based on the 100 day means of 6 hourly instantaneous fields for a single column of the GCM.

Results from equilibrium statistically steady-state LES driven by the climate model are shown in Figure 2 and Figure 3. In Figure 2, the time series of sea surface temperature (SST) show that the LES reach a statistically steady state with relatively modest drift from the initial condition, which corresponds to the 100 day mean from the GCM. This drift in SST is especially modest given that the LES is freely evolving, in the sense that the LES is not being nudged back to the GCM solution in any way, and given the large differences in the way that three-dimensional turbulence is represented in the LES and in the GCM.

In Figure 3, profiles of cloud liquid water averaged over the last ten days of the simulations are shown. Here the transition from shallow to deeper convection with increasing sea surface temperature and corresponding reductions in large scale subsidence is clearly evident and is consistent with what we see in real world climates.

Here we show that freely evolving LES driven by forcing from a GCM can yield reasonable, physically realizable, steady states for a single climate. To look at cloud feedbacks, it is only necessary to simulate climate change with the GCM and impose the resulting forcing on the LES.

The forcing framework can be used to:

  • Uncover the physical mechanisms of cloud feedbacks.
  • Investigate how deep convection changes with climate.
  • Study extreme precipitation events and how they change with climate in simulations that resolve convective processes directly.
  • Investigate rectification of time-varying forcing.
  • Provide numerical experiments to inform the development of novel GCM parameterizations.
  • Perform statistically steady-state LES sensitivity studies in a wide variety of atmospheric conditions.

Figure 1: Ocean heat fluxes added to the climate model to induce a Walker Circulation. Negative values indicate fluxes which cool the ocean. The black circles show locations along an east-west transect for which steady state LES are performed.

Figure 2: Time evolution of SST in GCM-forced LES show that simulations reach an equilibrium statistically steady state. Each line corresponds to a point in the transect shown in Figure 1, with the warmest simulations being the farthest east in the transect. The lines are labeled by the mean of the SST over the last 10 days of the simulations.

Figure 3: Vertical profiles of cloud liquid water computed from statistically steady-state LES with forcing from an idealized climate model. Each profile corresponds to a different point in the transect depicted in Figure 1 and is colored according to the equilibrium sea surface temperature.

PyCLES: A novel large eddy simulation code for the atmosphere

Collaborators: Colleen M. Kaul, Tapio Schneider, Zhihong Tan, and Sid Mishra

PyCLES is a project I initiated during my postdoctoral work at Caltech and ETH Z├╝rich and has become a tool being used by a growing community of researchers.

To get a feel for what PyCLES can do, please see Visualization.


Large eddy simulation has fast become a tool of choice for addressing research problems that involve three-dimensional turbulent flows. This is certainly the case for atmospheric flows, and in particular for studies of cloud processes, where LES will play a central role in reducing the primary sources of uncertainties in climate models. (e.g. Schneider et al, 2017). This growing use of LES to address fundamental questions in climate sciences led us to develop an Open Source, next-generation, atmospheric LES code from scratch called Python Cloud Large Eddy Simulation, or PyCLES for short.

A Tailored Design

By developing a code from scratch we were able to rethink at every level the way LES codes are designed and tailor a design for the growing community of researchers applying LES to climate problems. We prioritized four design objectives:

  1. A rigorous, thermodynamically consistent treatment of moist thermodynamics that permits simulations of deep convection.
  2. Incorporation of recent advances in numerical methods, particularly for the discretization of advective terms.
  3. Ease of use and maximization of extensibility for subject-matter, rather than software, experts.
  4. Performance, portability, and scalability across a range of hardwares. Scientifically informative simulations can be run on laptops to super-computers.

With these objectives, we have developed a massively parallel LES code that solves an energetically consistent form of the anelastic equations of motion, features closed budgets of specific entropy and total water specific humidity, offers a diverse set of numerical methods, and is written predominantly in Python and Cython. The formulation of PyCLES is discussed in more detail in Pressel et al., (2017).

A Python/Cython Workflow

The conventional workflow for doing scientific computing has been to perform large scale simulations using codes written in languages such as Fortran or C/C++, and then to do data analysis and plotting in a scripting language such as IDL, Matlab, or Python. This leads to significant startup overhead for people beginning to work with models, particularly for students with relatively little computing experience. We have written PyCLES with this issue in mind, and have sought to change this paradigm by designing the code from the outset to fit neatly within a Python workflow. This design paradigm has already shown great promise with undergraduate students and early graduate students rapidly becoming productive users and modifiers of the code.


We are always open to growing the user base of PyCLES, if you are interested in using PyCLES or have questions about it please email me.

I have blogged about this work at:

Robust large eddy simulations of boundary clouds

Collaborators: Sid Mishra, Colleen M. Kaul, Tapio Schneider, Zhihong Tan


While LES can resolve the scales of motion that are important in the dynamics of clouds in terms of well understood physical principles such as Newton's laws of motion, they are not without their own sources of modeling error. LES intercomparison studies (e.g. Stevens et al., 2005) have shown that these errors can lead to widely diverging solutions among various well constructed LES codes. Having confidence in the numerical solutions provided by LES is key to using them as a numerical laboratory for understanding climate change.

High fidelity large eddy simulation of stratocumulus clouds

Stratocumulus clouds are the most prevalent boundary layer cloud type (e.g. Wood, 2012). Thus, in order to use LES to study cloud feedbacks , they must be able to robustly simulate stratocumulus. However, stratocumulus clouds have been a long term challenge for LES (e.g. Stevens et al., 2005). In a recent study, Pressel et al. (2017), we us PyCLES to show that most of the spread in Stevens et al. (2005) is the result of anomalous turbulent fluxes that arise due to interaction of numerical error generated by the advection schemes and the sub-grid scale models used to represent the effects of turbulence that is unresolved by the LES.

We also show that using weighted essentially non-oscillatory (WENO) schemes, which have numerical error that is dissipative, for scalar and momentum transport and no explicit sub-grid scale model yields solutions that compare best with observations (Figure 1). This implies that the dissipative numerical error in the WENO schemes serves as an implicit sub-grid scale model that outperforms standard approaches. This approach is often referred to as implicit large eddy simulation (ILES) . The success of ILES is somewhat surprising given that, unlike ILES, conventional sub-grid scale models explicitly account for the effects of stratification. We also offer a simple analytical model that supports these conclusions.

Figure 1: Profile of the LES resolved vertical velocity variance simulated by PyCLES using various advection schemes and when possible without SGS models. Black circles show observations reported in Stevens et al. (2005). The simulations with WENO schemes and no explicit sub-grid scale model give the best match to observations and are shown in the bottom right panel labeled Paired-NSGS. This figure is reproduced from Pressel et al. (2017).

Robust LES of other cloud types

While it has been shown that the ILES approach allows robust simulation of stratocumulus clouds, more work remains before this conclusion can be extended to other cloud types. We are also working towards validating LES against space-based observations for a broad range of cloud conditions and over long time scales.