Machine-learning emulation of a forest biogeochemistry model for efficient biosphere optimization

Management controls the spatial configuration of a number of landscapes globally, from forests to rangelands. The majority of landcover change and all land-use change is the result of human decision-making. As human populations and global temperatures continue to increase, an engineering approach is needed to ensure the persistence of biological diversity and natural capital critical to human well-being. Such an approach may be based on manipulating ecosystems to achieve desired future states, informed by the latest simulation models. Models of the land surface are now being used to inform policy in the form of planning and management practices. This often involves the application of models that include spatial dynamics and operate at a landscape scale. The strong correspondence between the resolution and extent of modeling and management activities at this scale, and ability to efficiently simulate the decadal-to-centennial time-scales of interest, provide managers with a credible scientific tool for anticipating future land states under different scenarios. The importance of such tools to managers has grown dramatically with the challenges posed by anthropogenic climate change. As ecosystem simulation models continually improve in precision, accuracy, and robustness, we posit that models may be mathematically optimized as a basis for optimizing the management of real-world systems. Since current ecosystem simulation models are coarse approximations of highly complex and dynamic real-world systems, such optimizations should ideally account for uncertainty and physical or biochemical constraints, thereby improving the tractability of the optimization problem. In this work, we demonstrate the emulation and optimization of a forest biogeochemistry model from the SORTIE-PPA family of models. In doing so, we provide the first demonstration of the concept of biosphere optimization (Erickson 2015), which may one day be extended to include computational genetic manipulation experiments. To perform this work, we utilize the open-source Earth-systems Research and Development Environment (ERDE) library, which contains built-in functions for performing these and other analyses with land models, with a particular focus on forests.