Search

The path toward realizing the potential of seasonal forecasting and its socioeconomic benefits relies on improving general circulation model (GCM) based dynamical forecast systems. To improve dynamical seasonal forecasts, it is crucial to set up forecast benchmarks, and clarify forecast limitations posed by model initialization errors, formulation deficiencies, and internal climate variability. With huge costs in generating large forecast ensembles, and limited observations for forecast verification, the seasonal forecast benchmarking and diagnosing task proves challenging. Here, we develop a probabilistic deep learning‐based statistical forecast methodology, drawing on a wealth of climate simulations to enhance seasonal forecast capability and forecast diagnosis. By explicitly modeling the internal climate variability and GCM formulation differences, the proposed Conditional Generative Forecasting (CGF) methodology enables bypassing crucial barriers in dynamical forecast, and offers a top‐down viewpoint to examine how complicated GCMs encode the seasonal predictability information. We apply the CGF methodology for global seasonal forecast of precipitation and 2 m air temperature, based on a unique data set consisting 52,201 years of climate simulation. Results show that the CGF methodology can faithfully represent the seasonal predictability information encoded in GCMs. We successfully apply this learned relationship in real‐world seasonal forecast, achieving competitive performance compared to dynamical forecasts. Using this CGF as benchmark, we reveal the impact of insufficient forecast spread sampling that limits the skill of the considered dynamical forecast system. Finally, we introduce different strategies for composing ensembles using the CGF methodology, highlighting the potential for leveraging the strengths of multiple GCMs to achieve advantgeous seasonal forecast. Plain Language Summary: Seasonal forecast benefits a broad range of societal sectors. However, current dynamical seasonal forecast systems are considerably hindered by observation, model, and computation limitations. We develop a machine learning probabilistic forecast model that learns from climate simulations to infer possible climate patterns a season ahead. The model achieves competitive performance for global precipitation and temperature forecast, compared to the costly dynamical forecasts. More importantly, it offers crucial implications for understanding the limitations and improving the current dynamical forecast systems. Key Points: We develop the Conditional Generative Forecasting (CGF) methodology that learns from climate simulation for probabilistic seasonal forecastWe create a unique data set consisting 52,201 years of climate simulation to support data‐driven seasonal forecastsWe demonstrate the competitive skill of the CGF methodology for global seasonal forecast of precipitation and 2 m air temperature [ABSTRACT FROM AUTHOR]

The fidelity of climate projections is often undermined by biases in climate models due to their simplification or misrepresentation of unresolved climate processes. While various bias correction methods have been developed to post‐process model outputs to match observations, existing approaches usually focus on limited, low‐order statistics, or break either the spatiotemporal consistency of the target variable, or its dependency upon model resolved dynamics. We develop a Regularized Adversarial Domain Adaptation (RADA) methodology to overcome these deficiencies, and enhance efficient identification and correction of climate model biases. Instead of pre‐assuming the spatiotemporal characteristics of model biases, we apply discriminative neural networks to distinguish historical climate simulation samples and observation samples. The evidences based on which the discriminative neural networks make distinctions are applied to train the domain adaptation neural networks to bias correct climate simulations. We regularize the domain adaptation neural networks using cycle‐consistent statistical and dynamical constraints. An application to daily precipitation projection over the contiguous United States shows that our methodology can correct all the considered moments of daily precipitation at approximately 1° resolution, ensures spatiotemporal consistency and inter‐field correlations, and can discriminate between different dynamical conditions. Our methodology offers a powerful tool for disentangling model parameterization biases from their interactions with the chaotic evolution of climate dynamics, opening a novel avenue toward big‐data enhanced climate predictions. Plain Language Summary: Accurate climate prediction is crucial for understanding climate change and implementing effective climate adaptation strategies. However, climate models that are used to generate climate predictions have multifaceted biases that often need to be corrected before predictions can be considered usable. We develop a data‐driven methodology that detects and corrects climate model biases using a game theory inspired machine learning technique. By applying physical and statistical constraints, our predictions are not only more accurate, but also more trustworthy as judged by our physical understandings. Key Points: Identify and correct climate projection biases using unpaired climate simulation and observation dataRegularize the data‐driven bias corrector using statistical and dynamical constraintsSignificant improvement of daily precipitation estimation regarding a broad range of spatiotemporal statistics [ABSTRACT FROM AUTHOR]

Statistical models of high-resolution climate models are useful for many purposes, including sensitivity and uncertainty analyses, but building them can be computationally prohibitive. We generated a unique multiresolution perturbed parameter ensemble of a global climate model. We use a novel application of a machine learning technique known as random forests to train a statistical model on the ensemble to make high-resolution model predictions of two important quantities: global mean top-of-atmosphere energy flux and precipitation. The random forests leverage cheaper low-resolution simulations, greatly reducing the number of high-resolution simulations required to train the statistical model. We demonstrate that high-resolution predictions of these quantities can be obtained by training on an ensemble that includes only a small number of high-resolution simulations. We also find that global annually averaged precipitation is more sensitive to resolution changes than to any of the model parameters considered. [ABSTRACT FROM AUTHOR]

We perform the first uncertainty quantification analysis of the turn‐of‐the‐century drought in the western United States using a large perturbed‐parameter ensemble of the Community Atmosphere Model version 4.0 (CAM4). We develop several metrics to characterize the aridity bias, spatial extent, and tropical forcing of the drought and use statistical models to ascertain that the modeled drought was mainly sensitive to CAM4 parameters related to deep convection and clouds. Deep convection parameters account for over half the variance across the drought metrics. We employ observed estimates of these drought metrics to infer probability distributions of the model parameters that lead to a better agreement with observations, thereby providing guidance on how to improve the simulation of drought in CAM4. We find that in some cases, the suggested parameter values that improve the simulation of one drought characteristic would degrade the simulation of another, suggesting that there is a complex relationship between the model parameters and drought in the western United States. We also demonstrate reductions in the uncertainty of the drought metrics of up to 30% by constraining with observations of all metrics. Furthermore, we demonstrate for the first time that improvements to the simulation of the spatial extent of the turn‐of‐the‐century drought also lead to improvements in the spatial extent of the Australian Millennium drought, suggesting that physics of drought as encoded by the model parameters may be generalizable. Key Points: Deep convection parameters account for more than half the CAM4 ensemble variance in metrics used to quantify droughtWe found optimal parameter values that may lead to a better reproduction of drought in CAM4We reduce the uncertainty of the CAM4 simulation of a metric that quantifies the spatial extent of drought by 40% [ABSTRACT FROM AUTHOR]