There is growing interest in using IPCC-class climate model outputs in ecological research. These models provide realistic, global representations of the climate system, projections for hundreds of variables (including Essential Climate Variables), and combine observations from an array of satellite, airborne, and in-situ sensors. Unfortunately, direct use of this important class of data has been limited due to the large size and complexity of model output collections, internal file complexity, and limited means for dynamically creating derived products of interest. To address these limitations, we have developed an AI-based stochastic convergence technology, called MERRA/Max, that combines high-performance computing and Princeton’s Maximum Entropy (MaxEnt) software to rapidly subset and identify potential drivers of change among the hundreds of variables in a climate model output collection. MERRA/Max reduces dimensionality by iteratively drawing on MaxEnt’s capacity for feature selection to winnow randomly selected climate variables until a stable set of predictors is found. Preliminary work focuses on the MERRA reanalysis, a product of NASA’s GEOS-5 modeling framework. At 1PB in size, MERRA comprises over 700 climate variables and spans 1970 to the present at high temporal resolution. We evaluated MERRA/Max by modeling the bioclimatic envelope of Cassin’s Sparrow using MERRA and BioClim variables.
Results/Conclusions
We created a bioclimatic envelope model for Cassin’s Sparrow (Peucaea cassinii) using US Fish and Wildlife Service Breeding Bird Survey observations as the dependent variable and BioClim variables 01-19 as independent variables in a baseline MaxEnt run. A second baseline run used MERRA-derived BioClim surrogates as independent predictors. Both runs produced probability maps corresponding to the published distribution maps for P. cassinii, confirming the general viability of the approach. The MERRA/Max stochastic convergence method was then used to automatically down-select ten climate variables from each of the baseline collections to function as predictors in subsequent runs. In multiple trials, MERRA/Max converged on a stable “Top Ten” predictors at around 1000 iterations and produced models and maps similar to the baseline runs. Parallel processing of stochastic convergence runs reduced total run-time from eight hours on a single processor to around 15 minutes on 32 processors. These preliminary results suggest that the MERRA/Max approach may provide a practical means of harnessing climate model outputs for use in ecosystem research. In addition, near-real-time access to the full richness contained in climate model products could promote valuable experimentation. Future work will be directed at further evaluating MERRA/Max’s effectiveness and refining the technology as needed.