Raoult et al. (2026) aiLand v1: physics-based land surface emulator with observational fine-tuning
Identification
- Journal: Geoscientific model development
- Year: 2026
- Date: 2026-11-12
- Authors: Nina Raoult, Ewan Pinnington, Mario Santa Cruz, Florian Pinault, Baudouin Raoult, Natalie Zelenka, Gabriele Arduini, Gianpaolo Balsamo, Souhail Boussetta, Matthew Chantry, Patricia de Rosnay, Peter Dominik Dueben, Christoph Rüdiger
- DOI: 10.5194/gmd-19-10763-2026
Research Groups
European Centre for Medium-Range Weather Forecasts (ECMWF), Reading, United Kingdom and Bonn, Germany.
Short Summary
This study introduces aiLand v1, a physics-based land surface emulator of ECMWF's ecLand scheme, developed using a multi-layer perceptron architecture. It employs a two-stage learning framework, pretraining on extensive ecLand simulations and then fine-tuning with real-world FLUXNET observations, demonstrating improved accuracy and physical consistency for surface energy fluxes while maintaining prognostic state integrity.
Objective
- To develop an accurate, stable, and differentiable physics-based land surface emulator (aiLand) that leverages both synthetic model simulations and real-world observations to capture core land surface dynamics and correct systematic biases.
Study Configuration
- Spatial Scale: Global, on ERA5 Tco399 reduced Gaussian grid (approximately 31 km horizontal resolution), with some experiments at O96 (approximately 125 km resolution). FLUXNET sites represent local (approximately 100 m) scale observations.
- Temporal Scale: ecLand simulations from 1998–2024 (hourly, aggregated to 6-hourly for training). Training period for pretraining: 1998–2019. Validation period: 2022. Out-of-sample evaluation period: 2020–2023. Fine-tuning on FLUXNET observations for 2000–2019, with validation on 2020–2023. Continuous autoregressive integrations for 1 year (2020) and 4 years (2020–2023).
Methodology and Data
- Models used:
- Emulator: aiLand v1, a Multi-Layer Perceptron (MLP) feed-forward neural network.
- Physical Model: ecLand (ECMWF's land surface model), component of the Integrated Forecasting System (IFS), based on CHTESSEL.
- Data sources:
- Synthetic Simulations: ecLand offline simulations (CY49R2) forced by ERA5 meteorological reanalysis (1998–2024, hourly, aggregated to 6-hourly).
- Meteorological Forcing: ERA5 reanalysis data (hourly, approximately 31 km resolution) for ecLand simulations and as dynamic inputs for aiLand.
- Static Surface Fields: Physiographic fields derived from ERA5 and updated versions (e.g., orography, vegetation, soil type, land-sea mask, surface albedo).
- Observations for Fine-tuning: FLUXNET network eddy-covariance flux measurements (latent heat, sensible heat, meteorological drivers) from FluxDataKit (harmonized dataset, 339 sites, filtered to 241 unique sites, split into 199 training and 42 validation sites). Energy-balance-closed variant of turbulent fluxes used.
Main Results
- The pretrained aiLand emulator reproduces ecLand's prognostic soil state with a 90-day root-mean-square error (RMSE) of 1.19 K for surface soil temperature (stl1) and 0.014 m³ m⁻³ for surface soil moisture (swvl1).
- It remains stable over continuous 4-year autoregressive integrations, with stl1 RMSE increasing by 17% (from 1.47 K to 1.72 K) and soil moisture RMSE growing more slowly (14%–20%).
- A single globally trained model outperforms regime-specialist baselines in cross-regime transfer, with residual errors concentrating in snow-insulated cold regimes and extreme arid regions.
- Fine-tuning on FLUXNET observations reduces latent heat flux (LE) RMSE by 30% and sensible heat flux (H) RMSE by 20% at validation sites (using aiLand-S4 strategy).
- Fine-tuning improves the physical consistency of the surface energy budget: per-site Bowen ratio error (MedAE) drops by 42% (from 0.26 to 0.15 for aiLand-S4), and energy balance closure residuals fall from 13.4 W m⁻² (pretrained) to within 0.1 W m⁻² (aiLand-S5) at validation sites.
- Prognostic state integrity is preserved during fine-tuning, with only slight wetting (0.3–1.8 × 10⁻³ m³ m⁻³) and marginal cooling (0.02–0.16 K) of soil profiles.
Contributions
- Development of aiLand v1, a novel physics-based land surface emulator combining the strengths of extensive physical model simulations and real-world observations through a two-stage learning framework.
- Demonstration of the emulator's high fidelity in reproducing ecLand's dynamics, long-term stability, and robust generalization across diverse climate-land-surface regimes and spatial resolutions.
- Identification of specific error regimes (snow-insulated cold regions, extreme aridity) in ecLand's input-output relationships, highlighting areas for targeted physical model improvement.
- Proof-of-concept for using observational fine-tuning to correct systematic biases in surface energy fluxes inherited from the physical model, while maintaining prognostic state integrity and improving physical consistency (Bowen ratio, energy balance closure).
- Laying groundwork for future applications in data assimilation, parameter estimation, high-resolution land reanalyses, and integration into a fully data-driven Earth System Model.
Funding
- European Union's Digital Europe Programme through Destination Earth (DestinE).
- European Union's Horizon Europe research and innovation programme through CERISE (Copernicus Climate Change Service Evolution; grant no. 101082139).
Citation
@article{Raoult2026aiLand,
author = {Raoult, Nina and Pinnington, Ewan and Cruz, Mario Santa and Pinault, Florian and Raoult, Baudouin and Zelenka, Natalie and Arduini, Gabriele and Balsamo, Gianpaolo and Boussetta, Souhail and Chantry, Matthew and Rosnay, Patricia de and Dueben, Peter Dominik and Rüdiger, Christoph},
title = {aiLand v1: physics-based land surface emulator with observational fine-tuning},
journal = {Geoscientific model development},
year = {2026},
doi = {10.5194/gmd-19-10763-2026},
url = {https://doi.org/10.5194/gmd-19-10763-2026}
}
Original Source: https://doi.org/10.5194/gmd-19-10763-2026