Using Ensemble of Neural Networks to Learn Stochastic Convection Parameterizations for Climate and Numerical Weather Prediction Models from Data Simulated by a Cloud Resolving Model

Abstract

A novel approach based on the neural network (NN) ensemble technique is formulated and used for development of a NN stochastic convection parameterization for climate and numerical weather prediction (NWP) models. This fast parameterization is built based on learning from data simulated by a cloud-resolving model (CRM) initialized with and forced by the observed meteorological data available for 4-month boreal winter from November 1992 to February 1993. CRM-simulated data were averaged and processed to implicitly define a stochastic convection parameterization. This parameterization is learned from the data using an ensemble of NNs. The NN ensemble members are trained and tested. The inherent uncertainty of the stochastic convection parameterization derived following this approach is estimated. The newly developed NN convection parameterization has been tested in National Center of Atmospheric Research (NCAR) Community Atmospheric Model (CAM). It produced reasonable and promising decadal climate simulations for a large tropical Pacific region. The extent of the adaptive ability of the developed NN parameterization to the changes in the model environment is briefly discussed. This paper is devoted to a proof of concept and discusses methodology, initial results, and the major challenges of using the NN technique for developing convection parameterizations for climate and NWP models.

Prognostic validation of a neural network unified physics parameterization

“…On the other hand, n hid = 5 makes -13-more than twice as much error in the 64-step prediction than n hid = 128 does. These results confirm that just fitting the apparent source does not ensure long-term accuracy, which requires using more neurons that Krasnopolsky et al [2013] suggest.…”

“…For fixed T, increasing the number of hidden neurons hardly improves the R 2 scores for the apparent heat sources. Even networks with n hid = 5 can explain 60% and 70% of the variance of the apparent sinks of moisture and temperature, respectively, as also found by Krasnopolsky et al [2013]. On the other hand, n hid = 5 makes -13-more than twice as much error in the 64-step prediction than n hid = 128 does.…”

“…An analogous approach for machine learning parameterization is to choose the loss function as the MSE between the neural network's predicted Q 1 and Q 2 and the discrete estimate thereof that is derived from the model output. This is the approach taken by Krasnopolsky et al [2010Krasnopolsky et al [ , 2013, but they did not test their scheme in a prognostic fashion. In our own experiments, training in this fashion produces solutions that diverge to machine infinity after just a few timesteps even when the forcing is constant in time (i.e.…”

“…In an early paper, Krasnopolsky et al [2005] showed that a NN can emulate a realistic radiative transfer code but with much lower computational expense. They, then, trained a NN-based cumulus parameterization using a limited-area CRM, and showed that the scheme could accurately diagnose cloud fractions and precipitation [Krasnopolsky et al, 2010[Krasnopolsky et al, , 2013. However, a scheme which successfully predicts the instantaneous precipitation over the training data could be numerically unstable when predicting over multiple timesteps.…”

“…However, a scheme which successfully predicts the instantaneous precipitation over the training data could be numerically unstable when predicting over multiple timesteps. Krasnopolsky et al [2010Krasnopolsky et al [ , 2013 by training NN parameterizations on a near-global CRM simulation and testing the prognostic accuracy in a single column modeling framework. Single column models, which decouple the dynamics of adjacent grid cells, are commonly used to evaluate a parameterization's prognostic performance.…”