2026 Theses Doctoral

Ecosystem Responses to Rising Atmospheric Carbon Dioxide: Insights from Long-Term Observations and Hybrid Physics–Machine Learning Modeling

Zhan, Weiwei

The land absorbs roughly 25% of anthropogenic CO₂ emissions annually through photosynthesis; this photosynthetic carbon uptake is tightly coupled to water loss through plant transpiration, accounting for approximately 60% of global evapotranspiration. Despite this central role in the Earth's carbon and water cycles, how terrestrial ecosystems will respond to environmental change, especially continued CO₂ rise, remains highly uncertain, with the spread across process-based models comparable in magnitude to the mean estimate itself. This uncertainty stems from three interconnected challenges: (1) quantifying terrestrial carbon and water fluxes from available observations, particularly at large spatial scales, (2) resolving the mechanisms by which these fluxes respond to environmental drivers, and (3) translating continually improving process understanding into process-based models. This dissertation tackles these three challenges through four studies that jointly develop hybrid physics–machine learning approaches and combine them with long-term eddy covariance observations, satellite data, and Earth system model simulations.

The first study aims to improve ecosystem carbon flux quantification. A neural network framework informed by solar-induced chlorophyll fluorescence is developed to partition ecosystem CO₂ fluxes while learning the relationship between solar-induced chlorophyll fluorescence and gross primary productivity. This hybrid physics–machine learning framework improves the retrieval of ecosystem respiration and gross primary productivity under high temperature, water stress, and land management events such as harvest. It also provides the first direct, non-empirical characterization of the ecosystem-scale relationship between gross primary productivity and solar-induced chlorophyll fluorescence.

To advance process understanding, the second and third studies revisit how terrestrial ecosystems respond to rising CO₂. In the second study, the widely used space-for-time and variability-for-time empirical frameworks are shown to systematically fail to predict long-term dryland productivity trends, as they omit the CO₂ physiological effect that sustains vegetation greening even under atmospheric drying. To disentangle this CO₂ effect at the ecosystem scale, the third study develops a machine learning framework that circumvents prior assumptions derived from leaf-level experiments. This framework is applied to 63 global eddy covariance sites, revealing that the widespread CO₂-induced enhancement of ecosystem-scale intrinsic water-use efficiency is driven predominantly by reduced canopy conductance rather than productivity stimulation—a departure from leaf-level expectations. Three distinct response patterns of intrinsic water-use efficiency further emerge at the ecosystem scale, governed by nutrient supply, water availability, and biome types.

To translate these process-based insights into models, the fourth study develops a differentiable hybrid modeling framework that embeds mechanistic photosynthesis processes within a machine learning structure to infer how physiological acclimation regulates the ecosystem-scale CO₂ fertilization effect. Applied to observations from 80 global forest eddy covariance sites, the framework reveals widespread but spatially variable acclimation of two key plant traits: photosynthetic capacity and stomatal sensitivity. Acclimation of photosynthetic capacity in particular substantially dampens the CO₂ fertilization effect, reducing the cross-site median stimulation of productivity by 44% relative to a no-acclimation baseline. Together, these studies demonstrate that hybrid physics–machine learning modeling, grounded in ecosystem-scale observations, offers a flexible and principled strategy for reducing uncertainty in terrestrial carbon–water cycle projections under rising CO₂.