Estimates of how much carbon humans release by clearing, logging, and regrowing land disagree wildly. Over 2000–2023 they range from net emissions of 1.9 ± 0.6 PgC per year in process-based vegetation models to net removals of about −1.0 PgC per year in satellite-based products (Differences and uncertainties in land-use CO2 flux estimates). Much of that gap does not come from measurement error. It comes from a counterfactual question that observations cannot answer directly: what would this land have done if people had not touched it? In observations, deforestation, fire-prone degradation, and drought tend to occur together. In the Amazon, for example, human-influenced land accounted for 60% of gross carbon losses in El Niño years (Human influence on Amazon's aboveground carbon dynamics intensified over the last decade). A data-driven model trained on such data can credit a flux to either driver and fit the data equally well. Its "what if no land-use change" answer is therefore whatever its arbitrary internal arrangement happens to produce.
Computational neuroscience has a precise result about this kind of arbitrariness. Neural networks have continuous symmetries: rotations and rescalings of hidden units leave the output unchanged, so the internal arrangement is not pinned down. Two biological constraints break these symmetries. Requiring firing rates to be non-negative removes the rotations, and charging an energy cost for activity removes the rescalings. Together they force each hidden unit to track one underlying cause (Disentanglement with Biological Constraints: A Theory of Functional Cell Types). We propose to test whether the same two constraints, built into an emulator of land carbon fluxes, give more accurate counterfactual land-use emissions when the model is trained only on the "real world" and must infer the world without land-use change. We can score this exactly, because process-model ensembles provide both worlds.
Core Claim
Train a recurrent emulator on the factual simulation alone. This is the vegetation-model run forced by changing climate, CO2, and land use, called S3 in the TRENDY protocol. Then switch off its land-use inputs. Its prediction should then match the counterfactual run S2 (no land-use change) more closely when its hidden units are non-negative and pay an activity cost than when they are unconstrained. The advantage should grow with how strongly land-use and climate drivers are correlated in the training data.
A second, more specific prediction concerns the "loss of additional sink capacity" (LASC). LASC is the carbon that cleared forest would have absorbed from rising CO2 had it stayed standing, and it raises process-model land-use emissions by 42–77% (Differences and uncertainties in land-use CO2 flux estimates). LASC is an interaction between land use and environment, not a separable term. We predict that a constrained emulator will develop a distinct group of units that respond only when clearing and CO2 rise co-occur. If so, the main accounting conventions differ only in which unit groups they sum. Bookkeeping models exclude the interaction group. Process models include it. National inventories add the natural sink on managed land. The bookkeeping-plus-managed-land-sink reconciliation reported by the same synthesis would then appear as an identity over learned components.
What would be new is not "a neural net for carbon fluxes." It is the use of symmetry-breaking constraints as a tool for identifying counterfactuals, scored against exact factorial ground truth.
Research Program
The emulator is a recurrent network run per grid cell at annual resolution. Its inputs are gridded land-use transitions, including gross transitions and wood harvest (LUH2), climate forcing (CRU-NCEP v4.0), and atmospheric CO2. Its output is the net land–atmosphere flux. Hidden units come in two kinds. Fast units carry a firing-rate cost. Slow units are stored the way short-term synaptic plasticity stores memory: variables with decay and facilitation dynamics that cost almost nothing to hold. This follows the activity-silent working-memory networks of Circuit mechanisms for the maintenance and manipulation of information in working memory, which held information in synaptic states during simple delays and fired persistently only when information had to be transformed.
The ecological analogue is legacy carbon. Slash, dead wood, and soil pools decay quietly for decades after a disturbance. That is far beyond the roughly two-week memory that standard LSTMs exploit for photosynthesis (Understanding and Modeling Climate Impacts on Photosynthetic Dynamics with FLUXNET Data and Neural Networks).
The output is written as gross gain minus gross loss. Each is a non-negative readout of non-negative units, matching how satellite studies report gross fluxes. The training objective is
$$\mathcal{L} = \sum_{t}\left(F_t - \hat F_t\right)^2 + \lambda \sum_{t,i} r_{t,i}^2,$$
where $F_t$ is the simulated net flux in year $t$, $\hat F_t$ is the emulator's prediction, $r_{t,i} \ge 0$ is the activity of fast unit $i$, and $\lambda$ sets the energy price.
One known risk shapes the design. An activity cost penalizes carrying large gain and large loss at the same time, so it may push the model to represent only net flux and underestimate gross flux. Gross flux matters: omitting gross transitions alone biases emissions by 11–38%. We therefore sweep $\lambda$ and report gross-flux error alongside counterfactual error. This mirrors Effects of noise and metabolic cost on cortical task representations, where only a medium-cost regime reproduced primate prefrontal dynamics and both extremes failed. We expect a similar interior optimum here.
The theory carries a real caveat. The disentanglement proofs assume independent causes (Disentanglement with Biological Constraints), and correlated drivers are exactly our regime. That same paper reports "warped" but still modular geometries when factors are correlated. Whether this partial robustness carries over to biogeochemical drivers is the empirical question at the center of the proposal, not an assumption.
We also borrow one negative result as a design constraint. Non-negativity alone did not produce sparse codes in a rectified Gaussian autoencoder, which stayed near 50% sparsity at every regularization strength. A Poisson-coupled cost raised sparsity from 7% to 96% (Metabolic cost of information processing in Poisson variational autoencoders). The "non-negativity only" ablation is therefore expected to fail, and we include a Poisson-rate variant of the cost as a second treatment arm.
The program has two stages. The first is a controlled test on TRENDY v13 gridded outputs, where S2 and S3 give exact counterfactuals for each model. The second is a transfer to the Amazon. There, aboveground carbon from disaggregated SMOS L-VOD (L-band vegetation optical depth, a satellite microwave measure of biomass), combined with TMF forest-transition maps and ERA5-Land/CHIRPS climate data, provides only the factual world. In the Amazon the emulator's attribution between drought and degradation can be compared with the published decomposition but not verified against ground truth.
Baselines
The naive reference is the unconstrained version of the same recurrent emulator: same inputs, architecture, and data, with signed hidden units and no activity cost. Its counterfactual comes from the same input ablation. We use an LSTM because, among feedforward, LSTM, and Transformer models, LSTMs were the strongest on FLUXNET photosynthesis (Understanding and Modeling Climate Impacts on Photosynthetic Dynamics). It is the competitor any reviewer would reach for first. An LSTM with standard weight decay and dropout is a second version of this baseline, so that "any regularization" is separated from "symmetry-breaking regularization."
The strongest structured competitor is an additive model with explicit pairwise interaction terms between land-use, climate, and CO2 drivers. Such a model can represent LASC by construction. We also include gradient-boosted trees (XGBoost), which Comparative Analysis of Two Machine Learning Algorithms in Predicting Site-Level Net Ecosystem Exchange found fast and accurate within the training range but unable to extrapolate. We expect it to fail on zero-land-use counterfactuals in heavily cleared cells. Finally, as an accounting reference, a bookkeeping-style estimate under fixed environment shows what ignoring the interaction costs.
The ablations isolate the key ingredient. They are non-negativity only, activity cost only, both, the Poisson-rate cost, and slow synaptic memory on versus off.
Every arm gets the same training cells and years, the same input set, and an equal hyperparameter search budget, using an identical number of trials selected on held-out factual fit only, never on counterfactual error. Each arm is trained across five seeds per configuration.
Beating the baseline means two things at once. First, lower counterfactual land-use flux error than both the unconstrained LSTM and the interaction-additive model. Second, factual fit no worse than 5% above the LSTM's, across a majority of TRENDY models. For context, the ensemble-mean land-use flux to be recovered is 1.9 ± 0.6 PgC per year over 2000–2023 (Differences and uncertainties). We found no reported result on Lacuna for counterfactual attribution error of a data-driven emulator trained on factual simulations, so all baseline numbers will be produced in this study.
Why not just…?
Why not just ablate the inputs of any well-fit emulator? That is our naive baseline, and it is a real threat. With weakly correlated drivers it may do as well as the constrained model. The proposal predicts that its counterfactual error grows faster with driver correlation than the constrained model's does. The correlation sweep below tests exactly that. If the two curves coincide, the core claim is false.
Why not just fit an additive model with interaction terms? Such a model needs the analyst to choose the interaction structure. It also represents legacy effects only through hand-designed lag features. If explicit pairwise interactions with lagged land-use inputs match the constrained emulator's counterfactual error, the honest conclusion is that the neuroscience constraints add interpretability but not accuracy. We will report it that way. The emulator's possible edge lies in learning the decades-long memory and the interaction groups without specifying them.
Why not just use non-negative matrix factorization or Shapley attributions? Non-negative factorization is non-negativity without an activity cost. The P-VAE result (Metabolic cost of information processing) suggests non-negativity alone leaves the scale symmetry unbroken, and our non-negativity-only ablation measures this directly. Shapley values split interaction credit by a fixed convention. For LASC, that convention is precisely the disputed accounting choice, so Shapley values restate the disagreement rather than resolve it.
Proposed Experiments
The decisive experiment uses each TRENDY v13 model separately. We train every arm on S3 only, then predict the world without land-use change and score it against S2. Metrics are cell-level root-mean-square error and the global land-use flux error in PgC per year. We control how correlated the drivers are by building training sets from regions or periods where land-use change and drought anomalies co-vary to different degrees. The headline result is error plotted against driver correlation for each arm. The leave-one-model-out version trains on one vegetation model and tests whether the learned structure, not the specific fit, recovers that model's own LASC.
To quantify LASC, we need a land-use counterfactual under a fixed environment. We will use participating models' fixed-environment land-use runs where they exist, or run such a simulation with an open-source vegetation model where they do not.
Module analysis measures how purely each unit group responds to land-use, climate, CO2, or their interaction, using a mutual-information-gap score with driver labels. It then tests whether summing groups reproduces the bookkeeping, process-model, and inventory conventions.
The Amazon transfer reports the constrained emulator's split of gross losses between degradation and drought in El Niño years. It compares this with the published totals: deforestation −0.55 ± 0.04 PgC, degradation −0.42 ± 0.08 PgC, and 60% of El Niño losses on human-influenced land. These comparisons serve as a plausibility check, not as proof.
A negative outcome is informative in either direction. A flat correlation curve would mean symmetry breaking does not help identify counterfactuals. Failure of the interaction group to emerge would mean LASC must stay an explicit modeling choice.
References
- Differences and uncertainties in land-use CO2 flux estimates
- Human influence on Amazon’s aboveground carbon dynamics intensified over the last decade
- Disentanglement with Biological Constraints: A Theory of Functional Cell Types
- Circuit mechanisms for the maintenance and manipulation of information in working memory
- Understanding and Modeling Climate Impacts on Photosynthetic Dynamics with FLUXNET Data and Neural Networks
- Effects of noise and metabolic cost on cortical task representations
- Metabolic cost of information processing in Poisson variational autoencoders
- Comparative Analysis of Two Machine Learning Algorithms in Predicting Site-Level Net Ecosystem Exchange in Major Biomes