Massoud, E. C., Collier, N., Sharma, B., Kumar, J., & Hoffman, F. M. (2024). Enhancing Photosynthesis Simulation Performance in ESMs with Machine Learning-Assisted Solvers. In 2024 IEEE International Conference on Big Data (BigData) (pp. 4351–4356). IEEE. https://doi.org/10.1109/bigdata62323.2024.10825207

Improvements in the efficiency of photosynthesis solvers within Earth System Models (ESMs) are achieved by integrating machine learning (ML) techniques, offering a data-driven approach to complement traditional numerical methods while preserving model physics.

The Science

This work integrates machine learning (ML) techniques into the photosynthesis solvers of the E3SM Land Model (ELM) to improve computational efficiency. By leveraging logged simulation data, neural networks (NN) and multi-linear regression (MLR) were developed to estimate optimal initial guesses for solvers at both leaf and canopy scales. These approaches maintain the model’s inherent physics while reducing solver iterations, showcasing how data-driven methods can complement numerical approaches in ESMs.

The Impact

The NN-based initial guesses reduced total solver iterations by 26.23% at the leaf level and by 2.62% at the canopy scale, while MLR achieved a 3.12% reduction at the canopy scale. These improvements enable faster model runs, reducing the computational cost and making high-resolution, long-term climate simulations more feasible. This work demonstrates how AI-driven approaches can complement traditional numerical methods, paving the way for more scalable and efficient Earth system modeling efforts that retain accuracy and reliability.

Summary

Earth System Models (ESMs) play a critical role in understanding and predicting climate dynamics. Photosynthesis simulations, a computationally intensive component of these models, were optimized in this study by implementing ML-based initial guesses for their solvers. Using logged data from point-scale simulations at the Duke site, NN and MLR models were trained to mimic solver response surfaces and estimate better initial guesses. This approach resulted in significant reductions in solver iterations, particularly at the leaf level, demonstrating the potential for ML techniques to accelerate ESM simulations while preserving model fidelity. These advancements not only enhance computational efficiency but also open pathways for integrating AI-driven solutions into other complex components of ESMs.