Rising atmospheric CO2 is thought to be the largest single reason land plants now absorb more carbon than they did twenty years ago. Measuring that effect directly at real sites is hard. Flux towers record how much carbon an ecosystem takes up every half hour, but over a tower's lifetime CO2 rises almost in a straight line. Anything else that also drifts over those years gets mixed up with it: canopies grow, nitrogen accumulates, instruments age. A model that adds "CO2" as one more input and credits it with the upward trend is really crediting the calendar.
We propose a different source of evidence. Plant physiology predicts that higher CO2 does more than raise uptake. It changes how the ecosystem responds to fast weather. At the leaf scale, elevated CO2 moves the main brake on photosynthesis from the stomata (the pores that admit CO2 and leak water) to the leaf's internal biochemistry. Uptake should therefore become less sensitive to dry air and should recover faster after light changes. We borrow input-driven nonlinear latent dynamical models from systems neuroscience, which learn a low-dimensional internal state whose movement is driven by measured inputs. We use them to learn the fast weather-to-uptake response at many towers, with slowly varying CO2 allowed to reshape that response only through a channel with a physiological meaning. The claim is that the CO2 effect is identifiable from tower data as a change in response geometry, even where it is not identifiable as a trend.
Core Claim
Neuroscience models of this kind treat a slow experimental variable, such as stimulus strength, as a context that bends a fast nonlinear vector field (the rule giving the direction and speed in which the internal state moves from each point). Interpretable Nonlinear Dynamic Modeling of Neural Trajectories showed that such a field can be recovered across stimulus regimes when its basis functions are paired with a "global leak", a damping term that keeps trajectories bounded. On a decision-making benchmark with a regime change, it reached a prediction error of 0.002, against 0.244 for a locally linear model. We apply the same idea with CO2 as the slow context.
The specific claim is this. Constrain CO2 to act through a stomatal-limitation channel: it may change the sensitivity of uptake to vapor pressure deficit and the ratio of carbon gained to water lost. Let all other slow drift act through a separate structural channel that scales uptake and evaporation together, as canopy growth does. A stitched multi-site model built this way recovers the CO2 effect in simulations where the truth is known. It passes placebo tests that trend-attribution methods fail. Finally, it makes a falsifiable cross-site prediction: larger relative CO2 effects at water-limited sites.
How This Differs from Existing Work
The closest ecosystem work is Understanding and Modeling Climate Impacts on Photosynthetic Dynamics with FLUXNET Data and Neural Networks. It fits feedforward networks, a long short-term memory network (LSTM, a recurrent network) and Transformers to 30-day weather histories at four FLUXNET sites, and finds roughly two weeks of photosynthetic memory. That paper motivates our need for internal state, but it omits CO2 entirely and has no latent geometry to inspect.
Physics Informed Modeling of Ecosystem Respiration via Dynamic Mode Decomposition with Control Input fits a linear state-space model to respiration, separating the ecosystem's own evolution from temperature and soil-moisture inputs. It finds four to six dominant modes. We reuse that finding to size our latent state. Our model differs in being nonlinear, in modeling uptake rather than respiration, and in using multi-year records rather than single growing seasons.
Two static approaches quantify CO2 effects or predict uptake without any dynamics. Improved Global Gross Primary Productivity Estimation by Considering Canopy Nitrogen Concentrations and Multiple Environmental Factors is a light-use-efficiency model that attributes 30.77 Pg C of the 2001–2018 global uptake increase to CO2. Assessment of Six Machine Learning Methods for Predicting Gross Primary Productivity in Grassland benchmarks tree ensembles on 8-day satellite data. Both are memoryless. The first's CO2 attribution comes from an assumed functional form rather than from dynamics learned from data.
From neuroscience we take three methods. From BRAID: Input-driven Nonlinear Dynamical Modeling of Neural-Behavioral Data we take the separation of internal dynamics from measured inputs and its multi-step forecasting loss. We also take its warning: drivers that are not measured get absorbed into the internal state. That warning is exactly the confound we must control. From Inferring single-trial neural population dynamics using sequential auto-encoders we take "stitching": session-specific input and output layers around one shared dynamics model, which improved decoding by 0.18 R² across sessions. Here each tower plays the role of a session.
Research Program
The model keeps a latent ecosystem state of four to six dimensions. The state evolves as
$$\dot{x} = f_0(x, u) + \sum_{k} g_k(c_{\mathrm{CO_2}})\, h_k(x,u) + \sum_{j} s_j(c_{\mathrm{str}})\, r_j(x,u) - \alpha x,$$
where each symbol means the following:
- $x$ is the latent state and $u$ is the fast weather input (shortwave radiation, vapor pressure deficit, air temperature, soil water).
- $f_0$ is a shared basis-function vector field.
- $h_k$ are terms allowed to change stomatal sensitivity and water-use efficiency, scaled by smooth gains $g_k$ of the CO2 concentration $c_{\mathrm{CO_2}}$.
- $r_j$ are terms that scale carbon and water output together, scaled by gains $s_j$ of slow structural covariates $c_{\mathrm{str}}$, namely satellite leaf area and stand age where available.
- $\alpha x$ is the global leak.
Readouts predict uptake (GPP) and evapotranspiration jointly. The joint readout is what lets the two channels be told apart: canopy growth raises both outputs, while CO2 should raise carbon gained per unit water.
Each site gets its own affine input and output layers around the shared field, and the whole model is trained end to end. End-to-end training removes the two-step pipeline that Interpretable Nonlinear Dynamic Modeling of Neural Trajectories lists as a limitation. Basis-function count is held tractable by the small latent dimension the respiration study supports.
The channel constraint is an assumption, not a discovery, and we treat it as one. An unconstrained variant lets the CO2 gains enter every term. If the data favor the constrained direction, the unconstrained variant should converge toward it. If it does not, that is reported as a negative result.
Validation runs at three scales.
Leaf scale. We use the controlled induction measurements from Elevated CO2 Shifts Photosynthetic Constraint from Stomatal to Biochemical Limitations During Induction in Populus tomentosa and Eucalyptus robusta. Light steps were measured at 400 versus 800 µmol mol⁻¹. Stomatal limitation was 37.75% and 47.15% at ambient CO2, and biochemical limitation was 61.08% and 62.70% at elevated CO2, with shorter induction times. Here CO2 was manipulated by design, so a correct context model must infer faster relaxation and a weaker stomatal channel at 800.
Simulator scale. We generate synthetic tower records from a Farquhar-type photosynthesis and stomatal model driven by real FLUXNET2015 weather. We plant a known CO2 response alongside confounding slow drifts: leaf-area growth, nitrogen accumulation, and a slow sensor bias. Truth is then available for every run.
Tower scale. We use long-record FLUXNET2015 sites, including the four used in the LSTM study and the forest and grassland sites used in the respiration study.
Baselines
The naive reference point is a site-wise linear trend in annual uptake regressed on CO2. This is what "CO2 explains the trend" amounts to. On synthetic data with planted confounders it should absorb all of the drift.
The strongest predictive competitor is a gradient-boosted tree ensemble given weather, CO2 and the structural covariates, with CO2 effects read from partial dependence. It is the strongest because Assessment of Six Machine Learning Methods for Predicting Gross Primary Productivity in Grassland reports LightGBM at R² = 0.971 and RMSE = 3.149 on 8-day Mongolian grassland uptake. That setting is coarser than ours, so it is not directly comparable.
The strongest dynamical competitor is an LSTM with a 30-day window plus CO2 input, following Understanding and Modeling Climate Impacts on Photosynthetic Dynamics with FLUXNET Data and Neural Networks. That paper reports the LSTM beating its other architectures, but the file gives no numerical value, so no reported result was found on Lacuna for its error.
We also run a bilinear extension of dynamic mode decomposition with control, letting CO2 multiply the input matrix. No reported result was found on Lacuna for numerical uptake error from that method; its paper reports respiration only, described as comparable to standard partitioning. The light-use-efficiency model provides a process-style reference, with reported daily validation of R² = 0.63 against FLUXNET.
Every learned model gets the same sites, the same splits, the same input variables, and an equal hyperparameter search budget measured in trials. The splits train on early years and test on the final years, which have the highest CO2.
Two ablations isolate the proposal's ingredients. One removes the channel constraint. The other removes the joint evaporation readout. Beating the baselines means two things. First, matching or improving held-out-year forecast error. Second, and decisively, recovering the planted CO2 effect in simulation with materially lower error than the best baseline while leaving confounding drift attributed to the structural channel.
Why not just…?
Why not just add CO2 as a feature to LightGBM or an LSTM and read off its effect? This is the most serious alternative, because both models can fit tower data well. The problem is identification, not fit. Over a 15–20 year record, CO2 and a year index are nearly collinear, so any unconstrained learner can assign the drift to either one. Trees additionally cannot extrapolate beyond the CO2 range they were trained on. Our placebo experiment settles this. We replace CO2 with a year index, or give the baselines an extra fake monotonic covariate. If their attributed "CO2 effect" barely changes, they were measuring time. Our model is held to the same placebo. It passes only if its effect appears as reduced dryness sensitivity and higher water-use efficiency, and not merely as a level shift.
Why not just compute the inherent water-use-efficiency trend from tower data? This standard ecological diagnostic is a genuine threat, since it targets the same physiological signature. It is static, though: it ignores the two-week memory and cannot separate a canopy that is greener from one that is more water-efficient once soil-moisture history varies. We include it as a reference. If it recovers the planted simulator effect as well as our model does, the dynamical machinery is not justified, and we will say so.
Why not just use linear dynamic mode decomposition with CO2 as a control input? It is interpretable and already proven on tower data. But a linear operator cannot express CO2 changing the sensitivity to other inputs. The leaf data show exactly such a regime shift, and Interpretable Nonlinear Dynamic Modeling of Neural Trajectories documents how locally linear models fail across regime changes. The bilinear baseline gives it the interaction term anyway. If it matches our model in simulation, the nonlinear field is unnecessary.
Decisive Experiments
The first test is simulator recovery. Across planted CO2 sensitivities and confound strengths, we compare the error in the recovered CO2 response curve with the error from each baseline, under the condition that confounding drift is routed to the structural channel.
The second test is the leaf-scale check. The context gains fitted to the induction data must reproduce the reported direction of change: shorter induction time and a weaker stomatal share at 800 µmol mol⁻¹. We also look for the species difference in quenching as a secondary check.
The third test is the placebo and extrapolation suite on FLUXNET2015. We measure forecast error on held-out late years. We test whether the inferred CO2 effect survives substituting a year index. If it survives unchanged, the effect is time and not CO2.
The last test is the cross-site prediction that water-limited sites show a larger relative CO2-driven water-use-efficiency gain than wet sites. Because CO2 is nearly uniform across sites at any given time, this prediction draws on variation the trend cannot supply. A null or reversed result would count against the core claim.
References
- Interpretable Nonlinear Dynamic Modeling of Neural Trajectories
- Understanding and Modeling Climate Impacts on Photosynthetic Dynamics with FLUXNET Data and Neural Networks
- Physics Informed Modeling of Ecosystem Respiration via Dynamic Mode Decomposition with Control Input
- Improved Global Gross Primary Productivity Estimation by Considering Canopy Nitrogen Concentrations and Multiple Environmental Factors
- Assessment of Six Machine Learning Methods for Predicting Gross Primary Productivity in Grassland
- BRAID: Input-driven Nonlinear Dynamical Modeling of Neural-Behavioral Data
- Inferring single-trial neural population dynamics using sequential auto-encoders
- Elevated CO2 Shifts Photosynthetic Constraint from Stomatal to Biochemical Limitations During Induction in Populus tomentosa and Eucalyptus robusta