Session R37: Predictability of the Climate System

Data Assimilation and Uncertainty Quantification in the Geosciences

Data assimilation is the name commonly given to the estimation process that
generates moments of probability density functions of time dependent processes
modeled by physics, and observations. The inherent uncertainties of model and data
are taken into account using a Bayesian framework. Data assimilation is presently
used in applications as diverse as weather forecasting and spacecraft navigation.
I will appeal to familiar statistical physics to present the general methodoloy. I will
briefly descirbe a couple of computational implementations of the method
summarize some of the key research challenges that arise in their application.
I will also describe some novel applications of the methodology, potentially
useful in tracking targets, hurricanes, and ongoing research in the application
of stochastic parametrization and machine learning for the purpose
of dimension reduction.
  

Climate Change and Climate Variability: A Unified Framework

The “death of stationarity” poses a substantial challenge to climate predictability and to the climate sciences in general. This challenge is addressed herein by formulating the problems of change in the climate’s intrinsic variability within the framework of the theory of nonautonomous and random dynamical systems (NDS and RDS) with time-dependent forcing. A key role in this theory is played by the pullback attractors (PBAs) that replace the strange attractors of the more familiar theory of autonomous dynamical systems, in which there is no explicit time dependence of either forcing or coefficients.
The concepts and methods of the NDS and RDS approach will be introduced and will be illustrated using a stochastically perturbed version of the Lorenz (1963) convection model. This illustration will be followed by applications to models of the wind-driven ocean circulation and the El Niño–Southern Oscillation (ENSO). One finds that two local PBAs, a quiescent and a chaotic one, coexist within the wind-driven ocean model's decadally modulated global PBA, whereas a critical transition between two types of chaotic behavior occurs in the seasonally forced ENSO model.
Implications for the climate sciences in the era of anthropogenic change will be discussed.
Reference
Ghil, M. and V. Lucarini: The physics of climate variability and climate change, Rev. Mod. Phys., submitted, arXiv:1910.00583.   

Quantifying uncertainty in climate predictability using perturbed physics ensembles and climate model emulation

Climate models are essential tools for understanding and predicting Earth system processes and feedbacks, but uncertainties in their future projections remain challenging to characterize. Improvements in physical process realism and the representation of human influence arguably make models more comparable to reality, but also increase the degrees of freedom in model configuration leading to parametric and structural uncertainties in projections. Perturbed physics ensembles sample the uncertainty space through different choices of parameter settings. Climate model emulators can be a computationally efficient method of producing large ensembles of climate model output, in order to study different sources of uncertainty. In this work we use a machine learning algorithm to build an emulator for the land surface component of a climate model. Using a perturbed physics ensemble of model simulations, we train the emulator to predict model output given a set of parameter values as input. We optimize parameter values by comparing emulated model output with observations across multiple relevant metrics, including global carbon and water flux benchmarks. We also account for structural and observational uncertainty through a novel Bayesian calibration approach. By sampling the resulting posterior distributions and running future climate simulations, we can then estimate the contribution of land model parameter uncertainty in future projections of climate change.   

Earth System Modeling 2.0: Toward Data-Informed Climate Models With Quantified Uncertainties

While climate change is certain, precisely how climate will change is less clear. But breakthroughs in the accuracy of climate projections and in the quantification of their uncertainties are now within reach, thanks to advances in the computational and data sciences and in the availability of Earth observations from space and from the ground. To achieve a leap in accuracy of climate projections, we are developing a new Earth system modeling platform. It will fuse an Earth system model (ESM) with global observations and targeted local high-resolution simulations of clouds and other elements of the Earth system. The ESM is being developed by the Climate Modeling Alliance (CliMA), which encompasses Caltech, MIT, and the Naval Postgraduate School. CliMA will capitalize on advances in data assimilation and machine learning to develop an ESM that automatically learns from diverse data sources, be they observations from space or data generated computationally in high-resolution simulations. It will also engineer the ESM from the outset to be performant on emerging computing architectures, including heterogeneous architectures that combine traditional CPUs with hardware accelerators such as graphical processing units (GPUs). This talk will cover key new concepts in the ESM, including turbulence, convection, and cloud parameterizations and fast and efficient algorithms for assimilating data and quantifying uncertainties.   

Bayesian Inference for Climate prediction

Bayesian Inference in the geosciences is called data assimilation. It studies how to best combine information from complex numerical models with information from observations of the system at hand, given limited computational resources. This requires knowledge of the physics, numerical modeling including computer architecture, quantification of deficiencies in the numerical models, characteristics of observation errors, and Bayesian inference for very high dimensional highly nonlinear systems.
An important characteristic of the climate system is that the different Earth system components (e.g. atmosphere, ocean ,land surface and icecaps) have vastly different internal time scales. The main work horse for weather prediction, a (variational) smoother in which observations over a time window of 6-12 hours are used to find the best starting point for predictions, is problematic because the optimal time window length is substantially different for the different components. Even after 20 years of intensive research no satisfying smoother solution has been found.
This suggests to use a filter solution without an assimilation window, but the main workhorse there, the Ensemble Kalman Filter, suffers from too small ensemble sizes to accommodate the large number of observations (even when so-called localization is applied).
Another issue is that with its many feedbacks the climate system is highly nonlinear, while the standard methods for weather predictions are only optimal for linear, and perhaps weakly nonlinear systems. Furthermore, system updates are typically too abrupt and need to be added incrementally during the prediction.
We will discuss potential solutions based on existing techniques, and alternative ideas based on so-called particle flows.The latter are fully nonlinear while combining the strong points of smoothers and filters mentioned above, and have the potential to make substantial strides forwards towards better climate prediction.