- University: AgroParisTech
- Place of defence: Montpellier
- Hosting institution(s): Max Planck Institute for Biogeochemistry
Abstract
The TRENDY ensemble of land surface and dynamic global vegetation models provides the basis for the Global Carbon Budget’s bottom-up estimates of the land carbon sink. While widely used, these models are computationally expensive and suffer from documented errors in leaf area index (LAI), including bias shifts in the phenological cycle. This study trains a transformer neural network to emulate the TRENDY ensemble mean across 15 variables representing carbon and water cycles and LAI. The TRENDY factorial simulations isolate the primary drivers of the land carbon sink—CO2 fertilisation, climate change, and land-use change—enabling process-informed emulation. The model is then bias corrected by fine-tuning using the satellite-based LAI product AVH15C1. The emulator shows strong predictive skill for the non-disturbance fluxes, with weaker performance for disturbance-driven fluxes such as fire and land-use change emissions. Long autoregressive rollouts from 1901–2023 reveal some instability in carbon state variables, which is substantially reduced by training with progressively longer autoregressive windows. LAI bias correction proves successful, but does not meaningfully constrain other fluxes. This work demonstrates the potential of emulators as a bridge between process- and data-driven approaches, and is a step toward a publicly available tool allowing lightweight simulation of land surface dynamics.
