Application of Reduced-Order Modeling to Uncertainty in the Vulnerability of Permafrost Carbon to Climate Change

Abstract:

Earth System Models (ESMs) are used to estimate terrestrial feedbacks to climate change, such as the permafrost-carbon feedback. Fully characterizing the range of possible future behavior given uncertainty in model structure and parameterization is computationally expensive. Because current ESMs typically employ land models with non-interacting gridcells, there is an opportunity for computational speedup by simulating a subset of the gridcells that are representative of the full model grid. Here, we use a reduced-order modeling technique Gappy Principle Orthogonal Decomposition with Empirical Interpolation (GPOD-EIP) to demonstrate the feasibility of efficiently characterizing the range of future responses to climate. We analyze 31 variable-configuration simulations in the Community Land Model 4.5 (CLM4.5) with GPOD-EIP and create a 20-element POD basis representing the characteristic spatial patterns of soil carbon concentration along with a representative subset of 20 to 100 of the original 20975 gridcells. We train GPOD-EIP with 10 of the 31 simulations chosen adaptively, and reconstruct the time-varying soil carbon distribution in the remaining simulations with a relative error of less than 1%. Highly-variable quantities such as the latent heat flux require a much larger basis and gridcell subset for accurate reconstruction, although their global-mean can be represented more readily. We demonstrate the use of GPOD-EIP in uncertainty quantification by efficiently characterizing the dependence of future carbon feedback on the explicit depth-dependence of soil carbon decomposition, interpolating between the extreme cases investigated in a recent published analysis. We conclude that GPOD-EIP represents a promising reduced-order modeling technique for ESM land-model data and other applications where gridded spatial fields need to be efficiently emulated.