No abstract

We introduce the first benchmark for emulation of key spatially resolved climate variables derived from a full complexity Earth System Model • Three baseline emulators are presented which are able to predict regional temperature and precipitation with varying skill • Evaluation metrics and areas for future research are presented to encourage further development of trustworthy data-driven climate emulators

Many different emission pathways exist that are compatible with the Paris climate agreement, and many more are possible that miss that target. While some of the most complex Earth System Models have simulated a small selection of Shared Socioeconomic Pathways, it is impractical to use these expensive models to fully explore the space of possibilities. Such explorations therefore mostly rely on one-dimensional impulse response models, or simple pattern scaling approaches to approximate the physical climate response to a given scenario. Here we present ClimateBench -a benchmarking framework based on a suite of CMIP, AerChemMIP and DAMIP simulations performed by a full complexity Earth System Model, and a set of baseline machine learning models that emulate its response to a variety of forcers. These emulators can predict annual mean global distributions of temperature, diurnal temperature range and precipitation (including extreme precipitation) given a wide range of emissions and concentrations of carbon dioxide, methane and aerosols, allowing them to efficiently probe previously unexplored scenarios. We discuss the accuracy and interpretability of these emulators and consider their robustness to physical constraints such as total energy conservation. Future opportunities incorporating such physical constraints directly in the machine learning models and using the emulators for detection and attribution studies are also discussed. This opens a wide range of opportunities to improve prediction, consistency and mathematical tractability. We hope that by laying out the principles of climate model emulation with clear examples and metrics we encourage others to tackle this important and demanding challenge.

<p>Exploration of future emissions scenarios mostly relies on one-dimensional impulse response models, or simple pattern scaling approaches to approximate the physical climate response to a given scenario. Such approaches are unable to reliably predict climate variables which respond non-linearly to emissions or forcing (such as precipitation) and must rely on heavily simplified representations of e.g., aerosol, neglecting important spatial dependencies.</p><p>Here we present ClimateBench - a benchmark dataset based on a suite of CMIP, AerChemMIP and DAMIP simulations performed by NorESM2, and a set of baseline machine learning models that emulate its response to a variety of forcers. These surrogate models can skilfully predict annual mean global distributions of temperature, diurnal temperature range and precipitation (including extreme precipitation) given a wide range of emissions and concentrations of carbon dioxide, methane and spatially resolved aerosol. We discuss the accuracy and interpretability of these emulators and consider their robustness to physical constraints such as total energy conservation. Future opportunities incorporating such physical constraints directly in the machine learning models and using the emulators for detection and attribution studies are also discussed. This opens a wide range of opportunities to improve prediction, consistency and mathematical tractability.</p><p>We hope that by defining a clear baseline with appropriate metrics and providing a variety of baseline models we can bring the power of modern machine learning techniques to bear on the important problem of efficiently and robustly sampling future climates.</p>

Mixup -a neural network regularization technique based on linear interpolation of labeled sample pairshas stood out by its capacity to improve models robustness and generalizability through a surprisingly simple formalism. However, its extension to the field of object detection remains unclear as the interpolation of bounding boxes cannot be naively defined. In this paper, we propose to leverage the inherent region mapping structure of anchors to introduce a mixup-driven training regularization for region proposal based object detectors. The proposed method is benchmarked on standard datasets with challenging detection settings. Our experiments show enhanced robustness to image alterations along with an ability to decontextualize detections, resulting in an improved generalization power.