Process-Constrained Bias Correction of NOAA UFS Subseasonal-to-Seasonal (S2S) Precipitation Forecasts through Machine Learning

Subseasonal-to-Seasonal (S2S) scale predictions are critically important to societal goals like the protection of life and property, public health and safety, economic prosperity, and quality of life. However, because of the chaotic nature of the atmospheric circulation, predictability within the Earth system is limited after about two weeks. Beyond this time, some predictive skill can still be tapped from low frequency oscillations such as tropical convectively coupled waves and their downstream teleconnections, but the uncertainty associated with these processes is very large. Accurately modeling such convectively coupled waves is a longstanding challenge in numerical weather and climate prediction since they inherently depend on processes that are only partially resolved in global prediction systems. Due to the role these waves can play in modulating precipitation, a model’s failure to capture them accurately can decrease the skill of its precipitation forecasts. Despite this, many bias-correction and post-processing techniques focus only on the variable of interest, and ignore the representation of such underlying physical processes. Lack of consideration for precipitation-modulating processes like convectively coupled waves can allow bias correction and forecasting techniques to generate process-inconsistent precipitation forecasts which display unrealistic behavior, despite any amount of skill. Simply adding precipitation-modulating processes as predictor variables in machine learning models may not be sufficient to prevent process-inconsistent forecasts, because doing so does not explicitly penalize process-inconsistency. Constraining the bias-correction of a model’s precipitation field by explicitly penalizing unrealistic representations of the physical processes which modulate precipitation may enforce the process-consistency of the precipitation forecasts and therefore increase their skill. The criticality of such process-consistency to accurate S2S scale predictions has cultivated a broad interest in the development of process-informed bias correction and machine learning methods. Here, we explore developing a process-constrained bias-correction method based on a single-layer multiple-input, multiple-output (MIMO) neural network. While a MIMO neural network fits a separate output neuron for each target variable, the collective error from all target variables is propagated through the shared input and hidden layer neurons. If precipitation-modulating processes like convectively coupled waves are included as target variables during training, the parts of the network associated with predictions of precipitation will therefore be penalized both for poor predictions of precipitation, and for failing to capture the associated modulating processes accurately. In this study, we evaluate whether this configuration will serve to constrain the network to produce process-consistent forecasts, and compare the efficacy of such a machine learning-based process-constrained bias correction method to that of simpler process-informed bias correction methods and standard methods. We also approach this challenge in terms of “forecasts of opportunity” by using NOAA’s UFS pre-operational GFSv17/GEFSv13 prototypes to evaluate the extent to which this process-constrained bias correction method can improve representations of the precipitation-modulating physical processes during notable events.