HYPE-CAFE: Towards a Hybrid Model for Improved Marine Primary Production Estimates

Marine Primary Production (MPP) is a key component in understanding ocean ecosystems and their atmospheric carbon sequestration capacity. However, numerous challenges exist for obtaining MPP estimates. Algorithm variability is a significant issue, since various MPP models (chlorophyll-based or carbon-based algorithms) yield divergent results. Furthermore, the lack of observational data and periodic vertical profiles of the surface ocean hinder the ability to validate and refine such models.

This work focuses on improving MPP estimations by extending the state-of-the-art Carbon, Absorption, and Fluorescence Euphotic-resolving (CAFE) net primary production model with machine learning techniques to overcome current limitations. To improve the model's accessibility and versatility to be extended with data-driven methods, the original C code was rewritten in Python, resulting in a more user-friendly version named PyCAFE [https://github.com/jcrespinesteve/PYCAFE.git]. Using PyCAFE, simulations of MPP from 2003 to 2023 were conducted, producing a comprehensive dataset for training, validation, and testing. First, we train a random forest (RF) model using 500 random locations to emulate PyCAFE and to test global upscaling of MPP estimates. Our results show that the RF model has a strong capability for extrapolating MPP predictions with high accuracy [R²=0.96]. Second, we develop a hybrid model approach to simulate MPP: the HYPE-CAFE model (HYbrid marine Primary production Estimates based on the Carbon, Absorption, and Fluorescence Euphotic-resolving model). HYPE-CAFE combines the physical processes of the PyCAFE model with a neural network predicting the light-use efficiency (LUE), i.e., MPP is calculated as the product of absorbed photons and the predicted LUE. Preliminary results indicate that HYPE-CAFE provides an improvement over the predictions made with the CAFE model alone, especially in regions with variable environmental conditions. However, the lack of observational data limits the learning process. Therefore, in a next step we test a transfer learning approach to improve MPP predictions by HYPE-CAFE.

In conclusion, this project paves the way for the development of advanced hybrid modeling approaches, such as HYPE-CAFE, for global MPP estimation, and offers a transformative avenue for deepening our understanding of global ocean productivity, particularly in the context of climate change.