In this study we investigate a data-driven stochastic methodology to parametrize small-scale features in a prototype multiscale dynamical system, the Lorenz '96 (L96) model. We propose to model the small-scale features using a vector autoregressive process with exogenous variables (VARX), estimated from given sample data. To reduce the number of parameters of the VARX we impose a diagonal structure on its coefficient matrices. We apply the VARX to two different configurations of the 2-layer L96 model, one with common parameter choices giving unimodal invariant probability distributions for the L96 model variables, and one with nonstandard parameters giving trimodal distributions. We show through various statistical criteria that the proposed VARX performs very well for the unimodal configuration, while keeping the number of parameters linear in the number of model variables. We also show that the parametrization performs accurately for the very challenging trimodal L96 configuration by allowing for a dense (nondiagonal) VARX covariance matrix.

Constrained autoregressive models, Linear number parameters, Lorenz '96, Multiscale modeling, Stochastic parametrization
Communications in Applied Mathematics and Computational Science
Centrum Wiskunde & Informatica, Amsterdam, The Netherlands

Verheul, N, & Crommelin, D.T. (2021). Stochastic parametrization with VARX processes. Communications in Applied Mathematics and Computational Science, 16(1), 33–57. doi:10.2140/CAMCOS.2021.16.33