Climate projections increasingly use neural emulators, which are fast networks trained to imitate expensive land-surface simulators. Practitioners also prune these emulators, removing weights so that very large ensembles run quickly on GPUs. This note argues that the two practices interact in a way nobody measures. Land models used to ignore nitrogen and phosphorus. That omission made forests absorb far more CO₂ than they can when soil nutrients run short. Adding nutrient cycles cut the projected land carbon-concentration feedback by about 10% in JSBACH (Carbon–nitrogen interactions in idealized simulations with JSBACH (version 3.10)). It cut tropical gross primary productivity by 36% in ORCHIDEE-CNP (Convergence in phosphorus constraints to photosynthesis in forests around the world). The nutrient brake is a damping of the response to CO₂. It matters most in particular regions, such as old weathered tropical soils, and it barely shows up in average error over all grid cells and time steps. A pruned emulator could therefore keep a low overall error while losing exactly this brake. It would drift back toward the old carbon-only behaviour, recreating the omission error that the field spent a decade fixing.
The proposal makes that failure measurable and then tests a fix. The measurement compares the pruned emulator's CO₂ sensitivity against two paired runs of the same simulator, one with nutrients and one without. The fix is a pruning criterion that protects the nutrient contrast directly. The experiments ask how this "re-omission" depends on pruning granularity. Granularity means how weights are removed: individually and freely, in hardware-friendly 2-of-every-4 patterns, or as whole channels.
Core Claim
Removing capacity from an emulator of a nutrient-enabled land model preferentially erases the nutrient-limitation pathway before it degrades average accuracy. Coarser pruning should erase it faster at the same average error. The quantity to track is an omission fraction for a region $r$:
$$\Omega_r = \frac{S_r^{\text{pruned}} - S_r^{\text{CNP}}}{S_r^{\text{C}} - S_r^{\text{CNP}}}$$
Each $S_r$ is the change in cumulative productivity per unit CO₂ increase over region $r$, so it measures how strongly growth responds to rising CO₂:
- $S_r^{\text{CNP}}$ is the sensitivity of the nutrient-enabled simulator, which the emulator should reproduce.
- $S_r^{\text{C}}$ is the sensitivity of the same simulator with nutrient limitation switched off.
- $S_r^{\text{pruned}}$ is the sensitivity of the compressed emulator.
$\Omega_r = 0$ means the emulator is faithful. $\Omega_r = 1$ means it has become a carbon-only model.
The claim is falsifiable in two directions. Re-omission may not occur: $\Omega_r$ could stay near zero at every granularity whenever average error stays low. In that case compression is safe for this use, which is a useful negative result. If re-omission does occur, a criterion that protects the nutrient contrast should lower $\Omega_r$ at equal sparsity, equal speedup and equal average error.
Research Program
Training data. The first step builds factorial training data that existing emulator work lacks. Wesselkamp et al. trained LSTM, gradient-boosting and MLP emulators of ECMWF's ECLand scheme, but only on hydrological states with no nutrient cycles (Advances in Land Surface Model-based Forecasting). We will run ORCHIDEE-CNP under historical forcing as a 2×2 design: ambient versus elevated CO₂, crossed with nutrients limited versus unlimited. This mirrors the limited-versus-unlimited phosphorus scenarios that produced the roughly 70 Pg C tropical difference reported by Ellsworth et al. The emulator sees meteorology, CO₂, soil properties and the model's nutrient-state variables, and it predicts productivity and carbon fluxes. It is trained only on the nutrient-enabled runs. The carbon-only runs serve purely as the $\Omega$ reference.
Emulator. Following Wesselkamp et al., we use an LSTM applied separately to each grid cell, the architecture that gave them the best accuracy. Grid cells and ensemble members are batched, so the matrix multiplications are large enough for sparse hardware to matter.
Pruning sweep. The dense emulator is then pruned at matched sparsities under four granularities:
- unstructured magnitude pruning;
- 2:4 semi-structured pruning, which keeps two of every four weights, trained with the sparse-refined straight-through estimator from Learning N:M Fine-grained Structured Sparse Neural Networks From Scratch and corrected with ARMOR's block-diagonal factors (ARMOR);
- a hierarchical dense / 2:4 / fully-pruned region allocation with channel permutation, following FlexHiNM-GP;
- structured channel removal with dependency tracking and low-rank recovery, following LLM-Pruner.
These papers supply the methods. All of them were judged on image accuracy or language perplexity, never on whether a physical sensitivity survives compression.
Contrast-preserving criterion. The proposed fix borrows the form of Calibration-Preserving Pruning, which mixes a standard importance score with the gradient of a conformal score so that classifier uncertainty sets survive pruning (Calibration-Preserving Pruning). We replace the conformal term with a nutrient-contrast term, $I = (1-\lambda) I_{\text{base}} + \lambda I_{\text{contrast}}$. Here $I_{\text{base}}$ is the standard importance score of a weight or group. $I_{\text{contrast}}$ is the first-order change, caused by removing that weight or group, in the emulator's elevated-minus-ambient productivity difference evaluated on nutrient-poor calibration inputs. The weight $\lambda$ sets the balance between the two. The criterion plugs into each granularity's own masking rule. The fixed reference curve is the simulator's own CO₂ response with nutrients on, so no carbon-only run is needed during pruning.
What the emulator is judged against. ORCHIDEE-CNP has known biases. It underestimates the Northern Hemisphere sink, imposes overly strict phosphorus limits near the poles, and degrades the tropical seasonal cycle (Global evaluation of ORCHIDEE-CNP v1.2). A pruned emulator that weakens the polar phosphorus brake might therefore look better against observations while being less faithful to its simulator. For this reason $\Omega_r$ is always scored against the simulator. Comparisons with observations are reported separately and are never used to select a pruning method.
Baselines
Dense emulator. The naive reference is the unpruned LSTM trained on the nutrient-enabled runs. It sets the floor for $\Omega_r$ that comes from emulation alone, before any compression. Wesselkamp et al. report RMSE generally below 1.58 and MAE below 0.84 for an LSTM on the European subset of ECLand states. That result is on hydrological variables, not productivity. No reported result was found on Lacuna for any emulator of a nutrient-enabled land model.
Strongest hardware-friendly pruning method. This is 2:4 sparsity trained with the sparse-refined straight-through estimator, with ARMOR correction. Its 2:4 pattern is what Ampere-class sparse tensor cores accelerate. SR-STE kept ImageNet ResNet-50 at 77.0% top-1 against a 77.3% dense baseline. ARMOR reached a 1.57× layer speedup on Qwen 2.5-14B, against 1.86× for naive 2:4. No reported result was found on Lacuna for any pruning method on a land-surface or climate emulator.
Hierarchical pruning. FlexHiNM-GP is the strongest method at high sparsity. It reports 79.46% versus 73.52% top-1 for DeiT-Base at 80% sparsity.
Fairness rules. Every method gets the same training runs, the same calibration grid cells, the same GPU budget for mask search and recovery fine-tuning, and the same hyperparameter-search budget. Methods are compared at matched measured latency, not only matched sparsity.
Ablations. The ablation that isolates the key ingredient sets $\lambda = 0$ with everything else held fixed. A second control spends the same calibration budget on nutrient-poor cells but uses only $I_{\text{base}}$. This separates "protect the contrast" from "look at the right data."
What counts as winning. The contrast criterion wins if it halves the tropical $\Omega_r$ relative to the strongest baseline at equal latency, while average-error degradation stays within the baseline's.
Why not just…?
Why not just train a small dense emulator instead of pruning a large one? This is the most serious threat. If a small dense network keeps $\Omega_r$ near zero at the same latency, pruning granularity is beside the point. For that reason small dense networks, with and without the contrast term added to their training loss, are part of the main sweep. They are plotted on the same latency-versus-$\Omega$ curve. If they dominate, the contribution narrows to two things: the omission diagnostic, and the finding that any loss of capacity, however it happens, re-omits nutrients unless the contrast is protected.
Why not just fine-tune after pruning, the way LLM-Pruner recovers with low-rank adapters? Recovery fine-tuning minimises average error. The hypothesis is that average error is exactly what hides re-omission, so fine-tuning may restore error while leaving $\Omega_r$ high. Every pruned model is recovered under the same budget before $\Omega_r$ is measured. If recovery alone brings $\Omega_r$ to zero, the criterion is unnecessary, and the experiments will show that directly.
Why not just multiply a carbon-only emulator's output by a nutrient correction factor? Nutrient limits do not scale linearly. In Ellsworth et al., a five-fold rise in leaf nitrogen raised carboxylation capacity only 1.8-fold in low-phosphorus species, against 3.3-fold in moderate-phosphorus species. In ORCHIDEE the limiting nutrient switches from nitrogen on a 300-year-old Hawaiian soil to phosphorus on a 4.1-million-year-old one (A representation of the phosphorus cycle for ORCHIDEE). We will still fit a per-biome scalar correction as a baseline, and expect it to fail the chronosequence check described below.
Why not just use gradient boosting? Boosted trees are cheap and need no pruning. However, Wesselkamp et al. found the LSTM reduced MAE by up to 37% relative to the other emulators. Trees also cannot be differentiated through for coupled runs. Boosted trees are included as a non-neural point on the speed-versus-$\Omega$ curve.
Decisive Checks
Re-omission versus granularity. The headline experiment plots tropical and boreal $\Omega_r$ against sparsity and measured latency for each granularity at matched average error. The hypothesis fails if all granularities give $\Omega_r$ indistinguishable from the dense emulator.
Hydrology-only control. The same pruning sweep is repeated on an ECLand hydrology emulator, which has no nutrient pathway. Here we measure loss of a comparably regime-specific sensitivity, the soil-moisture response to precipitation in arid cells. This tests whether the effect is specific to nutrient-style damping or general to rare, region-specific behaviour.
Chronosequence switch. At the Hawaiian sites, the pruned emulator must still show the simulator's nitrogen response on young soil and its phosphorus response on old soil.
Nitrogen slope under low phosphorus. The emulator's productivity response to leaf nitrogen under low versus moderate phosphorus inputs must reproduce the simulator's own slope ratio, which encodes the Ellsworth constraint.
Sanity check against FACE. The elevated-CO₂ productivity gain is reported against the FACE value of 23% cited by Goll et al. This is context only and is not used for model selection.
References
- Carbon–nitrogen interactions in idealized simulations with JSBACH (version 3.10)
- Convergence in phosphorus constraints to photosynthesis in forests around the world
- Advances in Land Surface Model-based Forecasting: A comparative study of LSTM, Gradient Boosting, and Feedforward Neural Network Models as prognostic state emulators
- Learning N: M Fine-grained Structured Sparse Neural Networks From Scratch
- ARMOR: High-Performance Semi-Structured Pruning via Adaptive Matrix Factorization
- FlexHiNM-GP: Flexible Hierarchical Pruning via Region Allocation and Channel Permutation
- LLM-Pruner: On the Structural Pruning of Large Language Models
- Calibration-Preserving Pruning: Compression as a Reliability Contract
- Global evaluation of the nutrient-enabled version of the land surface model ORCHIDEE-CNP v1.2 (r5986)
- A representation of the phosphorus cycle for ORCHIDEE (revision 4520)