climate-weather/ai produced the result/Global Change Biology 2026 · v2
Neural network splits forest carbon measurements into uptake and release
Researchers trained a neural network on tower measurements of carbon dioxide at sites across the United States to separate the net flux into plant uptake and ecosystem respiration, with physical rules built into the training.
spectrum · one line per step, placed by what the step does · bright lines used AI
Constrained Carbon Partitioning: A Self‐Trained Physics‐Informed Machine Learning Model Refines GPP Estimates From Eddy Covariance Measurements
Global Change Biology, 2026
doi:10.1111/gcb.70886 · record aix-00280 v2 · checked 2026-11-05
- AI was for
- Property prediction
- Model family
- Transformer, Multilayer perceptron
- Checked by
- None stated
- Code
- available
The finding the paper is about came from the AI.
What this research was about
Towers above forests, grasslands and wetlands measure how much carbon dioxide moves between the land and the air. The technique, called eddy covariance, records rapid fluctuations in wind and gas concentration and turns them into a net figure: how much carbon the ecosystem gained or lost over half an hour. That net figure is the sum of two opposing processes. Plants absorb carbon dioxide through photosynthesis, known as gross primary productivity or GPP. At the same time plants, soil microbes and roots release carbon dioxide through respiration. The tower sees only the difference. Separating the two is unavoidable guesswork, because neither is measured on its own.
The conventional approach uses night-time readings, when photosynthesis stops, to estimate respiration, then extrapolates that into daylight hours using temperature relationships. A related method works from daytime data instead. Both rest on assumptions that may not hold. The researchers set out to do the separation differently, by training a model to produce uptake and respiration figures that add up to the measured net flux while also obeying known relationships between a plant's carbon gain and its water loss.
Where AI came in
The AI is the instrument that produces the study's central numbers. A neural network with a shared core and six output branches was trained from scratch on tower data from the National Ecological Observatory Network, covering 2019 to 2023. The branches predict quantities such as photosynthesis, respiration from different sources, the openness of leaf pores and water-use efficiency. These outputs are then fed through physical equations to reconstruct the carbon and water fluxes the tower actually measured.
Training was self-supervised: there were no reference values for uptake and respiration to copy, so the model was scored on whether its reconstructed net flux matched the measurement and on whether it respected imposed rules, including that fluxes cannot be negative, that photosynthesis is suppressed at night, and that leaf pore behaviour follows established theory. The network therefore stands in for the extrapolation assumptions of conventional partitioning, and for a measurement nobody can make directly. The paper notes the resulting differences cannot be fully checked without independent reference data.
Written by AIxSci from the checked record below, to give context for readers outside the field. It is not part of the record.
The work
Technical · from the record
A self-supervised neural network was trained on NEON eddy covariance data to split measured net ecosystem exchange into gross primary productivity and ecosystem respiration, using hard constraints for mass balance and flux non-negativity together with soft loss terms encoding stomatal optimality, water-use efficiency and carbon–water coupling. Model-derived NEE matched eddy covariance measurements with R = 0.99, and predicted GPP related to canopy transpiration with R = 0.58. At half-hourly resolution the model's GPP averaged 0.51 μmol m−2 s−1 below nighttime-partitioned GPP and 0.31 μmol m−2 s−1 below daytime-partitioned GPP, with the largest negative differences at low light. Annual GPP differences versus nighttime partitioning varied by plant functional type, from about −15% to −18% for evergreen needleleaf forests and −8% to −12% for deciduous broadleaf forests, while wetlands showed about +3%.
How AI was used
Half-hourly eddy covariance fluxes and meteorological drivers from NEON towers were combined with canopy-versus-surface CO2 and water flux components obtained by conditional eddy accumulation, then filtered and temporally aligned. These served as inputs to a neural network with a shared encoder of three stacked dense layers feeding six task-specific heads that predict stomatal conductance, GPP, above- and below-ground autotrophic respiration, heterotrophic respiration, water-use efficiency and calibration coefficients; head outputs are propagated through physically motivated forward equations to produce carbon and water fluxes. Training was self-supervised against measured NEE and the partitioned flux components, with Softplus activations enforcing non-negativity, NEE computed as a mass-balance identity, and a weighted composite loss adding soft terms for water-use efficiency, Medlyn stomatal optimality, stomatal conductance–transpiration coupling, nighttime GPP suppression, canopy and surface partitioning agreement, and a nocturnal RECO lower bound. Predictions were generated across all sites with Monte Carlo dropout at inference for uncertainty intervals, compared against nighttime and daytime partitioning computed with the hesseflux library, and examined with SHAP feature attribution, gradient-norm attribution of each physical loss, and retraining-based ablation of individual constraints.
The shape of the work
Structural · the record, drawn
Assemble NEON eddy covariance observations
Obtaining raw data, whether by measurement, download or retrieval.
We used eddy covariance (EC) observations from the National Ecological Observatory Network (NEON)where the paper describes this · verbatim
Partition high-frequency fluxes and preprocess
Cleaning, filtering, normalising or labelling data already obtained.
were obtained using conditional eddy accumulation (cea, Zahn et al. ) applied across the NEON networkwhere the paper describes this · verbatim
Train constrained multi-head network
Fitting model parameters, including fine-tuning an existing model. The AI stood in for statistical model.
The network is trained using a composite loss function combining physics‐based constraints to ensure predictions remain consistent with ecophysiological theory.where the paper describes this · verbatim
Infer GPP, RECO and stomatal variables
Running a trained model over new data to predict, classify or score. The AI stood in for unresolved measurement.
Uncertainty in GPP and RECO is estimated through Monte Carlo dropout applied at inference time.where the paper describes this · verbatim
Derive conventional partitioning baselines
Running a trained model over new data to predict, classify or score.
we used the Python library hesseflux (Cuntz ) to derive conventional nighttime‐partitioned GPP and RECOwhere the paper describes this · verbatim
Evaluate consistency and compare with baselines
Testing outputs against ground truth.
we compared KGML‐predicted NEE (calculated as RECO—GPP) against eddy covariance measurements across all sites and time periodswhere the paper describes this · verbatim
Attribute constraints and features
Extracting understanding from model behaviour.
we assessed feature importance using SHapley Additive exPlanations (SHAP) analysis to quantify the contribution of individual input variableswhere the paper describes this · verbatim
What the record says
Technical · every part carries its own basis
+ in the paper~ our reading− not reported
How to read the quotations. A quotation shows where the paper describes something. It does not quote every value beside it: one passage locates a part of the work, and values without their own quotation are our reading of that passage.
The GPP and RECO estimates the paper reports are the outputs of the trained KGML model; no non-AI route to the paper's central numbers is presented
We developed a self‐supervised knowledge‐guided machine learning (KGML) framework to partition NEE into GPP and RECO using NEON tower datawhere the paper describes this · verbatim
The supporting code and data for the KGML model for reading, processing, and exporting the partitioned datasets is availablewhere the paper describes this · verbatim
What this paper did not report
Technical · absence is published deliberately
- Trained model weightsWhether the trained model is available is not stated.
- ComputeThe hardware or time used is not stated.
- How many were testedThe paper gives no count of what was tested.
- ValidationNo validation of the AI is described.
- Version of KGML (knowledge-guided machine learning partitioning model)Which version of the model was used is not stated.
- Version of hesseflux nighttime partitioning (Reichstein et al.)Which version of the model was used is not stated.
- Version of hesseflux daytime partitioning (Lasslop et al.)Which version of the model was used is not stated.
- What step 7 replacedThe paper gives no basis for what the AI stood in for.
About this article
Record aix-00280, version 2, checked by a person on 2026-11-05. The record describes the paper; it does not assess whether the paper's findings are right. The paper is published under CC-BY; quotations are at most 25 words. How we work · Report an error
Cite this record: AIxSci (2026). Neural network splits forest carbon measurements into uptake and release. Record aix-00280, version 2, checked 2026-11-05. https://aixsci.org/articles/aix-00280/