Machine learning for accelerating process-based computation of land biogeochemical cycles.

Area of Science:

• Global change ecology
• Computational modeling
• Machine learning applications in Earth science

Background:

• Modern global change ecology relies on large datasets (big-data) and complex mathematical models (big-model).
• High computational demands limit the development and application of terrestrial biosphere models (TBMs) for long-term simulations.
• The spin-up phase, equilibrating biogeochemical cycles, is a major computational bottleneck, consuming up to 98% of total simulation time.

Purpose of the Study:

• To introduce a machine-learning acceleration (MLA) tool to overcome computational limitations in ecological modeling.
• To significantly reduce the computational time required for the spin-up process in terrestrial biosphere models.
• To assess the impact of MLA-driven spin-up on the accuracy of carbon balance predictions.

Main Methods:

• Developed and applied a machine-learning acceleration (MLA) tool to interpolate equilibrated biogeochemical variables.
• Tested the MLA tool on three versions of the ORCHIDEE TBM within the IPSL Earth System Model.
• Focused on TBMs with complex carbon, nitrogen, and phosphorus interactions lacking analytical spin-up solutions.

Main Results:

• An unoptimized MLA tool reduced computational demand by 77%-80% for global studies.
• MLA-derived equilibrium introduced minor biases, with negligible impact on predicted regional carbon balance.
• Optimization of MLA algorithms and training strategies are expected to yield a tenfold reduction in computation demand.

Conclusions:

• Machine learning acceleration offers a viable solution to the computational bottleneck in ecological modeling.
• The MLA tool is versatile, compatible with existing methods, and benefits complex, non-linear models.
• This approach enables more extensive simulations of ecosystem carbon and nutrient dynamics under global change.