The ecLand model is the land surface component of the ECMWF Integrated Forecasting System (IFS) , building on the Carbon–Hydrology Tiled ECMWF Scheme for Surface Exchanges over Land CHTESSEL;. It simulates the exchange of energy, water, and carbon between the land surface and the atmosphere, including associated subsurface processes. The model can be run either fully coupled within the IFS or in offline (uncoupled) mode forced by prescribed atmospheric fields.

Each grid box at the land–atmosphere interface is subdivided into surface tiles, with seven distinct classifications used over land: high and low vegetation, bare ground, shaded and exposed snow, intercepted water, lake, and urban. Each tile has its own physical parameters, allowing separate computation of water, heat, and momentum fluxes, and solution of an energy balance equation for tile skin temperature. Grid-box mean fluxes are obtained by area-weighted aggregation of tile-level fluxes and are passed to the atmospheric model.

For the ML emulator, we consider a subset of ecLand variables. In aiLand, soil water content and soil temperature in three layers (0–7, 7–28, and 28–100 cm) are treated as prognostic states, as in ecLand. Snow cover fraction is also treated as prognostic in aiLand, although in ecLand it is a diagnostic variable derived from the underlying prognostic snow water equivalent and snow density. Latent and sensible heat fluxes, 2 m air temperature, 2 m dew point temperature, and skin temperature are predicted as diagnostic outputs. In this study, ecLand CY49R2 (https://github.com/ecmwf-ifs/ecland, last access: May 2026) is run in offline mode, forced by ERA5 meteorological reanalysis . A global simulation covering 1998–2024 was performed at hourly temporal resolution on the native ERA5 Tco399 reduced Gaussian grid (∼31 km). Outputs were aggregated to 6-hourly time steps for training, to be consistent with that used in the AIFS.

Surface physiographic fields are assumed to be static and describe land surface characteristics such as orography, vegetation, soil type, land–sea mask, and surface albedo, capturing spatial heterogeneity relevant for land surface parameterisations. These fields determine the tile properties in ecLand and influence the computation of turbulent fluxes and tile skin temperatures. The fields are prescribed independently of model initial conditions. Leaf area index is provided using monthly climatologies to represent seasonal variability. All datasets are remapped to the model grid before use. In this study, physiographic fields derived from ERA5 and their updated versions are used, as described in .

ERA5 is the fifth-generation atmospheric reanalysis produced by ECMWF . It is generated using IFS cycle 41R2 with 4D-Var data assimilation for the atmosphere, a land data assimilation system , and a coupled land-atmosphere modelling approach using ecLand . ERA5 provides hourly atmospheric and surface fields at approximately 31 km horizontal resolution with 137 vertical levels extending to 80 km height. Uncertainty estimates are available at ∼63 km resolution from a 10-member ensemble. In this study, hourly ERA5 surface fields at ∼31 km resolution on the cubic octahedral reduced Gaussian grid (Tco399; N320) are used as dynamic forcing inputs.

Flux-tower measurements from the FLUXNET network provide in situ observations of land–atmosphere exchanges of energy, water, and carbon based on the eddy covariance technique. These observations include turbulent fluxes such as latent heat (LE) and sensible heat (H), together with meteorological drivers (e.g., radiation, air temperature, and precipitation) typically measured at half-hourly resolution.

In this study, we use the harmonised flux tower dataset provided by FluxDataKit , which compiles and standardises observations from several openly available flux data products including PLUMBER-2 , AmeriFlux , and ICOS releases e.g., Warm. The dataset provides quality-controlled and gap-filled fluxes and meteorological variables with consistent formatting across sites, as well as site metadata and continuous multi-year records suitable for model calibration and benchmarking. We use FluxDataKit's energy-balance-closed variant of the turbulent fluxes, in which raw H and LE are scaled proportionally to satisfy H+LE=Rn−G at each timestep (where Rn is net radiation and G is ground (soil) heat flux). This follows standard FLUXNET2015 conventions and removes the well-documented systematic underestimation of turbulent fluxes by raw eddy-covariance measurements and preserves the observed Bowen ratio at each site.

The observational dataset is split into training and validation subsets (see Appendix ), ensuring spatial separation between training and evaluation periods. For the fine-tuning part of this study, we train the model over the years 2000–2019 for 199 calibration sites, retaining 42 validation sites to also be evaluated over this period. For the main evaluation, the validation period of 2020–2023 was used, at the 84 training and 17 validation sites with valid post-2020 FLUXNET data. All metrics are computed on daily means, only where both prediction and observation are non-missing.

All datasets are prepared using the anemoi-datasets framework (https://anemoi.ecmwf.int/, last access: 10 November 2026) and stored in Zarr format, chunked in time to optimise input/output (I/O) performance. Dataset generation is controlled via YAML-based recipe files to ensure reproducibility. Feature-wise statistics (mean and standard deviation) are computed over the ecLand training period as part of the dataset creation pipeline and stored for subsequent use during model training.

The final dataset used for model pretraining comprises global ecLand simulations and ERA5 reanalysis time series for 1998–2024 at 6-hourly temporal resolution, together with static surface physiographic fields, defined on the model grid (N320, 171039 land points). Of this period, 1998–2019 is used for pretraining, 2022 for validation, and 2020–2023 is withheld for out-of-sample evaluation (Sect. ); the evaluation window ends in 2023 to match the availability of FLUXNET observations. Two separate observational datasets are created for training and validation for the fine-tuning step, covering 199 and 42 sites respectively, reducing the spatial dimension compared to the full model grid. FLUXNET observations (or missing values) replace model-derived surface fluxes, while other prognostic and diagnostic variables are retained from the simulations.

The MLP emulator is a feed-forward neural network approximating ecLand’s mapping from static, dynamic, and prognostic inputs to land surface states and fluxes: 1f:X→Y. Here, the input vector x∈X consists of static, dynamic, and prognostic state variables, while the output vector y∈Y contains prognostic state variable increments and diagnostic variables at the next timestep. The true mapping f is approximated by a learned function f*(x) via stochastic gradient-based optimisation.

The network comprises a shared prognostic backbone that produces a latent representation, which is projected to prognostic targets, and a diagnostic branch that combines climatological, meteorological, and prognostic information to predict diagnostic variables. Prognostic outputs are applied in a residual formulation, where predicted increments are added to the previous state, enabling stable autoregressive rollouts. Diagnostic variables are predicted as absolute values without memory of their previous value, mirroring ecLand, where skin temperature and the turbulent fluxes are computed from the tile energy balance at each time step rather than carried over from the previous one as prognostic states. Post-processing enforces variable-specific physical bounds. Physical consistency is therefore inherited from the ecLand training target and reinforced by the residual formulation and output bounds, rather than imposed through conservation laws or physical penalties within the network (see Sect.  for details).

Each hidden layer applies a fully connected (linear) transformation, followed by layer normalisation , followed by a rectified linear unit (ReLU) activation. Layer normalisation standardises each sample's activations across the feature dimension, stabilising training and reducing sensitivity to the differing scales of the input variables; ReLU introduces the non-linearity needed to capture ecLand's non-linear input–state mapping. The base model uses six hidden layers of width 512, with approximately 1.3 M trainable parameters, and a single-block diagnostic head. Figure  illustrates the architecture. Note that this model is wider and deeper than the version presented in to account for the additional variables.

The model is trained using a smooth L1 loss (Huber loss with β=1). Loss is accumulated across the autoregressive rollout: errors at each forecast step are summed and then scaled by the inverse rollout length. Global weights balance prognostic and diagnostic contributions, with optional per-variable weighting for individual targets (here all set to 1); when per-variable weights are used, the weighted losses are averaged over the number of variables. Prognostic increments are normalised before the loss by dividing by tendency scalers – the standard deviation of 6-hourly increments over the training set – ensuring balanced contributions across variables with different characteristic evolution rates. Diagnostic variables, which are standardised like all other targets, receive no additional tendency scaling since they are predicted as absolute values rather than increments. During fine-tuning to observations, the loss is computed only over valid (non-NaN) target elements, allowing the model to learn from sparse observational datasets where only a subset of variables or grid points carry valid values at any given time. Input features (meteorological forcing, prognostic states, and static physiographic fields) are sourced from the complete ERA5 reanalysis and ecLand simulations, which do not contain missing values at valid land grid points; a residual safety guard replaces any unexpected NaN in inputs with zero in normalised space.

Pretraining uses global ecLand simulations at N320 resolution from 1998–2019 (approximately 32 000 six-hourly timesteps over 171 039 land points), with 2022 reserved for validation and checkpoint selection. The years 2020–2023 are excluded from training and used exclusively for out-of-sample evaluation (Sects.  and ). Input features are normalised using feature-wise statistics computed over the full dataset, with variable-specific schemes (e.g., standardisation, max-scaling, or identity for categorical inputs). Missing values are masked during loss computation and set to zero in inputs.

Training proceeds in two phases with increasing rollout horizon. In Phase 1, the model is trained for 80 epochs (160 000 steps) with rollout length R=4 (24 h). The Adam optimiser is used with a peak learning rate of 5×10-4, decayed to 3×10-7 using a cosine schedule with 1000-step linear warmup. In Phase 2, the best checkpoint is fine-tuned with rollout R=8 (48 h) for 8 epochs (16 000 steps) at a reduced learning rate of 3×10-5. Gradient clipping (global norm 5.0) is applied throughout, and gradient checkpointing is enabled for longer rollouts to manage memory.

Each training step processes one sample, defined as a rollout window of R consecutive 6-hourly timesteps spanning all valid land grid points (171 039 for global pretraining; 199 for FLUXNET fine-tuning). Training is performed on 4 GPUs using distributed data parallelism with mixed precision, giving an effective batch of 4 rollout windows per gradient update. The smaller effective batch during fine-tuning (constrained by FLUXNET data volume) interacts with the choice of low fine-tuning learning rates: the very low rates used (10⁻⁵ to 10⁻⁷, Table ) help mitigate noise from small batches, but rigorous separation of batch-size effects from data-distribution effects on fine-tuning quality would require a dedicated ablation study, which we defer to future work. The best checkpoint is selected based on validation loss.

For the second part of this study, we fine-tune the aiLand-base emulator using in situ flux data. The observational data are normalised using the same statistics as during pretraining for consistency. We investigate five fine-tuning strategies (Table ) spanning a methodological spectrum from minimum-intervention to full-model adaptation, designed to test how aggressively the pretrained backbone should be modified during fine-tuning. Higher backbone learning rates risk catastrophic forgetting of the pretrained dynamics, while overly low rates limit fine-tuning's ability to correct inherited LSM biases; the five strategies span this trade-off. Diagnostic-only strategies (S1, S2) freeze the prognostic backbone entirely and update only the diagnostic head, ensuring that the state dynamics learned from ecLand pretraining are preserved exactly. S2 tests whether additional learning capacity in the diagnostic head improves flux predictions, by extending it to 3 blocks (two randomly initialised) before fine-tuning. Backbone-adaptive strategies (S3, S4, S5) update the prognostic backbone at a substantially lower learning rate than the diagnostic head, allowing gradual adaptation while protecting the pretrained dynamics. S3 uses a single training phase with explicit differential learning rates (backbone 10⁻⁷, diagnostic 10⁻⁴). S4 adopts a two-phase procedure: first updating only the diagnostic head, then unfreezing the full model at a uniform low learning rate (10⁻⁶) to mitigate catastrophic forgetting. S5 updates the full model from the start with a uniform learning rate (5×10-5) across all components.

To evaluate model performance, we use common metrics such as the root-mean-square error (RMSE), the mean absolute error (MAE) and the anomaly correlation coefficient (ACC). The ACC is calculated as the Pearson correlation between predicted and reference anomalies, defined as deviations from the climatology. The climatology is computed from the ecLand dataset over a multi-year reference period (here 1998–2024) using a sliding-window approach. For each 6-hourly timestep, samples are assigned to their hour-of-year bin and to neighbouring bins within a ±3 d window (cyclically), and averaged across all years to obtain a smoothed climatological mean for each timestep, variable, and grid point. All metrics are on the N320 reduced Gaussian grid, with glacier and coastal points excluded throughout.

To test the transferability of aiLand-base, we partitioned the land surface into climate–land–surface regimes by clustering those in a feature space that combines climatological indicators with land-surface states (soil moisture, snow cover, Bowen ratio). Working in a metric feature space lets us relate transfer skill to the distance between regimes, and test whether a single global model matches regime specialists, both on regimes it has seen and on one withheld from training. Climatological features (2009–2020) were computed from ERA5, including monthly mean and seasonal amplitude of 2 m temperature and precipitation, aridity index, top-layer soil moisture, snow cover fraction, leaf area index (LAI), and Bowen ratio. The features define static spatial regions, so a decade-long window suffices; computing them over the full training period instead assigns 98.9 % of land points to the same regime, which does not provide a justifiable improvement over the shorter period. All features were standardised to zero mean and unit variance prior to analysis to ensure comparable contributions across variables with different units. Regimes were identified using k-means clustering . The number of clusters was chosen by evaluating three cluster validity metrics over k=2–12: the silhouette score , the Davies–Bouldin index and the Calinski–Harabasz index . The first two are optimised at k=7, which lies within 1.3 % of the optimum of the third at k=9; we therefore use k=7. One cluster predominantly corresponding to glaciated or ice-covered regions was masked prior to transfer experiments, leaving six effective regimes. Spatial coherence of the regime labels was enforced using a local majority filter to suppress isolated cells while preserving large-scale structure.

We interpret the resulting clusters as climate-vegetation regimes (summarised in Fig.  and Table ). The six-cluster partition captures broad climate-vegetation regimes rather than following the usual regime naming conventions: “temperate continental” includes both grasslands and forests with similar climates, and “tropical/subtropical humid” is distinguished from “tropical rainforest” by precipitation and LAI rather than by latitude. The arid/semi-arid and dry-desert clusters are separated by an order-of-magnitude difference in mean precipitation (640 mm yr⁻¹ vs. 58 mm yr⁻¹) and vegetation cover (LAI 1.2 vs. 0.4), reflecting genuinely different surface energy and water budgets. However, the regime are not internally homogeneous: hard k-means clustering of continuous climate features inevitably assigns transition-zone grid points to single clusters despite genuinely intermediate characteristics. An internal margin analysis (relative distance to nearest vs. second-nearest centroid) indicates that around 30 % of all grid points fall in transition zones (borderline assignments or grid points reassigned by the spatial-coherence filter). The dry desert cluster (R4) is the most internally consistent, with only 11 % of its grid points in transition zones, while the tropical/subtropical humid cluster (R0) is the most heterogeneous (40 % transition-zone cells). Cross-regime transfer experiments (Sect. ) should therefore be interpreted as testing transferability across overlapping climate regimes rather than cleanly separated ecosystems.

For each regime, we extract latitude/longitude coordinates to define training subsets. Pairwise Euclidean distances between cluster centroids in feature space are computed to quantify regime similarity for these transferability analyses. Each specialist and the global model is evaluated on every regime, with RMSE normalised per variable by the mean in-regime specialist RMSE, so that 1.0 corresponds to average specialist skill. The transfer penalty of a model on a regime is its normalised RMSE there relative to that of the regime's own specialist, minus one.

We evaluate aiLand-base in two complementary protocols, both forced by ERA5 atmospheric reanalysis (the same forcing used in the corresponding ecLand simulations) and both initialised from ecLand state at the start time. The first protocol consists of 138 independent 90 d integrations launched every two days from January through early October 2022 (the latest start date allowing a full 90 d window within the evaluation year), used to characterise short- to sub-seasonal error growth as aiLand-base departs from ecLand when both are initialised from the same state and driven by identical forcing (Table , Fig. a). The second consists of continuous autoregressive integrations over 2020 (1 year) and 2020–2023 (4 year), used to assess long-range stability and the representation of the seasonal cycle (Table , Fig. b).

Errors remain well controlled over the 90 d integration horizon (Table , Fig. a). Surface soil temperature (stl1) RMSE rises by 29 % and surface soil moisture (swvl1) by 19 %. Deep soil layers have very small day-1 errors (stl3: 0.03 K; swvl3: 0.002 m³ m⁻³), reflecting long memory timescales, so their apparent growth ratios are large (stl3: 21.7×) despite absolute errors remaining smaller than at the surface throughout the time integration. Unlike the surface layer, where errors are continually corrected by the forcing, deep-layer increments depend mostly on the state itself, so errors there accumulate rather than being corrected. Snow cover RMSE peaks at intermediate lead times (0.046 at 60 d) before declining slightly, consistent with transition-season timing errors rather than monotonic drift. Diagnostic variables (2 m temperature: 0.69 K; skin temperature: 1.06 K; sensible heat flux: 11.7–12.5 W m⁻²; latent heat flux: 10.1–11.1 W m⁻²) show near-constant RMSE across all lead times, confirming that they are effectively re-diagnosed from the prognostic state at each step rather than accumulating error.

This stability is confirmed over continuous integration (Table ). While stl1 RMSE gradually increases from 1.47 K (1 year) to 1.72 K (4 year), a 17 % increase, and stl2 and stl3 grow by 34 % and 37 %, respectively, correlation coefficients remain high across all soil variables (≥0.993 for temperature; ≥0.993 for moisture), and anomaly correlations stay above 0.93 for all prognostic variables except swvl3, which decreases from 0.888 to 0.855. The stl1 RMSE trace (Fig. b) shows a seasonal cycle but no upward trend, confirming stable autoregressive behaviour with no dynamical divergence. Soil moisture RMSE grows more slowly than temperature (14 %–20 % over four years), and snow cover is nearly stable (+4 %), indicating negligible drift in non-thermal prognostic variables. Diagnostic variables degrade minimally (2 m temperature: 0.61 to 0.63 K; sensible heat flux: 11.8 to 12.1 W m⁻²; latent heat flux: 9.8 to 10.4 W m⁻²), with correlations remaining near unity (∼0.998–0.999) and anomaly correlations above 0.99.

The dominant systematic feature is a +0.40 K warm bias in stl1, present at 89 % of land grid points and largely established within the first year of integration. Equivalent integrations within the training period yield comparable biases (+0.44 to +0.47 K), confirming that the offset is an intrinsic property of the learned dynamics rather than distribution shift. The bias is consistent with a conditional-mean offset induced by the Smooth L1 loss under the right-skewed soil-temperature increment distribution (training skewness +0.94), with an additional Arctic component reflecting underestimated polar warming. The skew weakens with depth (0.42 for stl2, and approximately zero for stl3 once the most extreme 1 % of grid points is excluded), consistent with the bias being largest in the surface layer, which receives the diurnal forcing directly. The bias confined to soil thermal variables; moisture biases are negligible and diagnostic temperatures show small negative offsets consistent with per-step re-diagnosis.

The global RMSE values in Sect.  mask strong geographic structures (Fig. ). At latitudes >60° N the 90 d RMSE reaches 2.42 K for stl1 and 0.018 m³ m⁻³ for swvl1 – roughly three times its tropical value. The soil moisture comparison is harder to interpret directly because hydraulic properties and dynamic ranges differ between regions, and freezing affects the modelled state in cold winters; we therefore focus the analysis on relative spatial patterns rather than absolute magnitudes. A spatial decomposition (Fig. a) shows that the high-latitude error amplification is dominated by seasonal snow cover, although freeze-thaw cycling and reduced atmospheric coupling in cold winters likely contribute as well. Grid points with seasonal snow (fractional snow-cover RMSE > 0.02; 33 % of valid land area) account for 57 %, 82 %, and 76 % of the global stl1, stl2, and stl3 mean squared error at 90-day integration time. In these zones stl1 RMSE is 1.83 K (2.1× the snow-free value), stl2 is 1.31 K (4.3×), and stl3 is 1.16 K (3.5×); swvl3 is 2.9× the snow-free value. Seasonal snow covers 95 % of all land poleward of 60° N and 62 % of the 30–60° N band but less than 1 % of the tropics, explaining the latitudinal error gradient.

Consistent with the diagnostic re-diagnosis pathway established above, diagnostic outputs are largely unaffected by snow regime (Fig. b): 2 m temperature RMSE is nearly identical across regimes (0.70 K vs. 0.69 K snow-free), and surface heat fluxes are actually lower under snow cover. This is physically expected: snow's high albedo and the frozen surface state both suppress the absolute magnitude of turbulent exchange, so absolute flux errors are smaller in snow-covered conditions. Only skin temperature shows mild amplification (1.3×), reflecting its direct coupling to the prognostic soil state.

Error seasonality at snow-affected sites differs sharply between thermal and hydrological variables, as illustrated by point-level traces from the continuous 4-year integration (Fig. ). At a representative warm arid site (Fig. a), aiLand-base tracks ecLand essentially perfectly year-round, confirming that the residual errors identified above are regime-specific rather than systematic. At the boreal site (Fig. b), soil temperature errors concentrate during sustained snow cover rather than at onset or melt transitions: stl2 RMSE under continuous snow is 1.71 K vs. 0.37 K in snow-free months (4.6×). Snow acts as an insulating layer that decouples the soil column from atmospheric forcing, weakening the learned mapping from meteorological inputs to soil-state increments; sub-surface variables are disproportionately affected because they are furthest from any remaining surface coupling. Soil moisture exhibits the opposite seasonality: swvl1–3 errors peak after the snow has gone, during the active hydrological season when freeze-thaw cycling and snowmelt-driven rewetting produce complex transient dynamics that are harder for the emulator to track. Under snow cover itself, by contrast, soil moisture is largely frozen and temporally static, so low variability yields low error.

This site-level contrast generalises to the global error field. Grid points with seasonal snow show soil moisture RMSE 1.3–2.7× higher than permanently snow-free regions, and 85 % of snow-zone grid points exceed the snow-free median despite the errors occurring primarily outside the snow-covered season. Insulation and freeze–thaw mechanisms therefore act on different variables and in different seasons, but both have a strong impact on the annual error budget of snow-affected regions. Future work could improve this regime by extending the emulator to predict surface and sub-surface runoff, allowing it to better account for snowmelt routing and the post-melt soil-moisture dynamics.

A key question for ML-based land surface emulators is whether learned input-output mappings remain valid across climate-land-surface regimes, including extreme climates not well sampled during training. The snow-coupling analysis above predicts that cold-climate regimes should be among the hardest targets, because specialists trained on narrow climate regimes will not encounter enough snow-insulated soil dynamics to learn the distinct forcing-state relationship that applies under a snowpack. To test this and to evaluate broader regime sensitivity, we trained six regime-specialist models on data-driven clusters (see Table  and Fig.  for full characterisation). Each specialist was evaluated within its home regime and across all others (Fig. a), alongside the globally trained model evaluated under the same protocol.

Regimes are not uniformly learnable (Fig. a, left panel). R4 and R5 are the most easily fitted by their respective specialists, reflecting their relatively clean cluster definitions (R4: only 11 % transition-zone cells; R5: 24 %; see Sect. ). R2 and R3 show above-average in-regime errors reflecting a combination of greater internal process heterogeneity and substantial transition-zone content (∼26 %–28 %). Against this in-regime baseline, transfer penalty scales with dissimilarity in the training feature space (r=0.68; Fig. a, right panel). Transfers between similar regimes remain close to parity, while extreme shifts degrade strongly: R5 to R4 transfer reaches 4.9× the in-regime baseline. R4 emerges as the most challenging target overall (mean 3.1×), consistent with its outlier climatology (Table ): on the precipitation axis it overlaps only with R1, so transfers into it are extrapolations for every other specialist, whereas R1, spanning 120–2330 mm, overlaps both R4 and R0. The R4 penalty is partly amplified by near-zero in-regime variability for several variables.

The globally trained model achieves a mean normalised RMSE of 1.16; only 16 % above the specialist in-regime baseline. Its largest penalties occur in dry regimes (1.47 for R1; 1.60 for R4), reflecting different mechanisms in each case. R1 makes up 21% of land grid points (not under-sampled) but has substantial transition-zone content (∼31 % borderline or reassigned cells; Sect. ), so its representation is diluted into neighbouring clusters during global training. R4 is internally consistent (only 11 % transition-zone cells) but combines outlier climatology (58 mm yr⁻¹ precipitation vs. 550–2700 in other regimes, LAI 0.4 vs. 1.2–6.1) with a small sample share (12 % of land points), so the global training distribution provides limited explicit coverage of its distinctive surface energy and water budget. Notably, the global model outperforms cross-regime specialist transfers for the temperate (R2) and rainforest (R5) regimes. These are regimes where the specialists themselves face challenging training distributions: R2 spans grasslands and forests with similar climates but distinct surface physics, and R5 cells appear partly in R0 due to the continuous precipitation-LAI gradient (Sect. ). The global model benefits from having seen the full climate continuum without enforced cluster boundaries.

Because the global model has seen every regime in training, its advantage over cross-regime specialist transfer could reflect exposure rather than generalisation. To separate the two, six leave-one-regime-out models were trained on all regimes but one and evaluated on the held-out regime (Table ). Removing a regime degrades performance on it only slightly: four of six regimes are within 5 % of the global model, and the largest degradation is 16 % for R4, against cross-regime specialist penalties reaching 4.9×. The cost of omitting a regime does not scale with its share of the training data: removing R0, the largest regime in Table , has no effect, whereas removing R4, one of the smallest, is the most damaging. What matters is whether the remaining regimes contain the same physical conditions; R1 and R4 are the only water-limited regimes in the partition, so removing either removes most of the exposure to soil-moisture-limited evaporation.

The prognostic-diagnostic asymmetry identified in Sect.  persists across regimes (Fig. b). Prognostic variables incur moderate transfer penalties (mean ratios 0.92–0.71), with soil temperature layers transferring almost perfectly (stl1: 0.92; stl2: 1.01) and soil moisture and snow cover showing larger penalties (1.34–1.71). Diagnostic variables are substantially more sensitive (median 2.0× vs. 1.1× for prognostics), with latent heat flux (3.4×) and skin temperature (2.9×) the most affected. The stacked decomposition in Fig. b shows that global training recovers the majority of the diagnostic cross-regime penalty, indicating that training over diverse climate samples largely compensates for regime-specific surface-atmosphere coupling.

Overall, a single global model provides the most robust performance across regimes, and this is not simply because it has seen every regime in training: removing a regime degrades performance on it by at most 16 %, against transfer penalties of up to 4.9× when a specialist is applied outside its training regime. What the global model gains is exposure to conditions (e.g., water limitation, snow-insulated soil) that no single-regime training set contains. Specialists offer an advantage primarily for diagnostic variables within climatically well-defined regimes. Beyond cross-regime transfer, we also tested whether the same prognostic-diagnostic asymmetry holds across spatial resolutions by training a second global aiLand-base instance on an O96 (∼125 km) ecLand simulation. Prognostic state predictions transfer cleanly between the two grids in both directions, while diagnostic outputs, particularly 2 m temperature and surface fluxes, remain sensitive to the training grid's spatial characteristics (Appendix ). For operational applications targeting screen-level scores, this sensitivity could be mitigated by training directly at the target resolution, by multi-resolution training to encourage resolution-invariant features, or by diagnostic-only post-hoc bias correction. We plan to investigate these options systematically in future work.

Across Sect. –, aiLand-base reproduces ecLand's prognostic dynamics with high fidelity, remains stable over multi-year integrations, and generalises cleanly across regimes and spatial resolutions (Appendix ). Diagnostic variables are consistently more sensitive to regime changes than prognostic ones, reflecting their absolute-value prediction pathway vs. the residual formulation used for prognostic states.

Two physically distinct error sources stand out: snow-soil coupled dynamics, where the insulating snowpack decouples the soil column from atmospheric forcing and weakens the learned increment mapping, and extreme aridity, where the desert regime's surface energy balance differs sufficiently from vegetated training environments to degrade cross-regime transfer by up to ∼5×. Both align with known limitations of ecLand and land surface models more broadly. In cold regions, uncertainties in snow insulation, freeze-thaw processes, and soil thermal coupling lead to systematic errors in sub-surface temperature and moisture. In arid environments, surface fluxes are highly sensitive to soil moisture thresholds and sparse vegetation, making models calibrated on humid conditions difficult to transfer to water-limited regimes. The spatial and regime-dependent error patterns exposed by aiLand-base provide a structured diagnostic of where ecLand's input-output relationships are hardest to learn, and therefore where process representations may benefit most from targeted development. In this sense the emulator serves not only as a surrogate for ecLand but as an interpretive tool for identifying priority areas for physical model improvement.

All five strategies substantially reduce flux RMSE relative to aiLand-base at both training and validation sites (Table ). LE improvements exceed those for H: validation LE RMSE decreases by 17 %–32 %, while H decreases by 8 %–17 %. This asymmetry reflects the structure of the inherited bias. ecLand carries a large systematic LE bias (+8.8 W m⁻² at validation sites) that fine-tuning readily corrects, whereas H bias is significantly smaller (+4.6 W m⁻²), leaving a less correctable signal. LE also exhibits a strong seasonal cycle linked to soil moisture and vegetation – inputs well represented in the model – while H depends more strongly on boundary-layer processes and the surface-air temperature gradient, making it sensitive to atmospheric conditions not provided as inputs. Fine-tuning largely removes the inherited LE bias: at training sites, |bias|<1 W m⁻² across all strategies; at validation sites, residual biases range from -1.7 to -4.5 W m⁻², with aiLand-S4 the smallest.

The strategies differ most in how well gains transfer to unseen sites and unseen years. Spatially, the training-validation RMSE gap ranges from 2.9 W m⁻² (aiLand-S5) to 9.8 W m⁻² (aiLand-S2) for LE, and from -0.5 to 5.3 W m⁻² for H. aiLand-S2 achieves the lowest training RMSE but the largest gap, indicating that its randomly initialised diagnostic blocks overfit training-site patterns. aiLand-S4 combines a small spatial gap with the best validation LE correlation (r=0.786, up from 0.722 for aiLand) and validation H correlation (r=0.719, up from 0.657).

Temporal generalisation reinforces this pattern. Comparing validation-site LE improvements between the training period (2000–2019) and the out-of-sample evaluation period (2020–2023), aiLand-S4 is essentially stable (30.2 % vs. 29.6 %) and aiLand-S1 shows only mild attenuation (29.7 % vs. 28.2 %), whereas aiLand-S2 drops markedly (24.2 % vs. 16.5 %). aiLand-S5 is the one case where validation-site improvement strengthens in the evaluation period (27.9 % to 31.9 % for LE; 19.7 % to 15.8 % for H), suggesting that full-model fine-tuning with a low learning rate learns a structural correction rather than memorising site-specific patterns, though at some cost to training-site skill.

To illustrate site-level behaviour, we focus on aiLand-S4 as a representative fine-tuning strategy and delivers the best flux RMSE and correlations across validation sites (Table ). At the site level, aiLand-S4 improves LE at 14 of 17 validation sites and H at 14 of 17; 11 sites improve in both fluxes simultaneously, while no site degrades in both (Fig. ). The largest LE gains occur at water-limited sites (IT-Lsn, -60.6 %; US-xGR, -56.4 %) and extend across temperate wetland (US-ALQ, -57.2 %) and Mediterranean (IT-Tor, -53.9 %) areas. Three sites show LE degradation: US-NC4 (coastal wetland, +45.2 %), CH-Dav (+13.8 %), and US-xCP (+0.3 %), though all three improve in H. Conversely, three sites with strong LE improvement degrade in H – IT-Tor (+21.9 %), US-Me1 (+36.5 %), and CH-Cha (+44.2 %) – predominantly high-altitude or complex-terrain locations where the ∼60 km grid cannot resolve topographic variability. Figure  illustrates these improvements at few validation sites of varying climates and vegetation. We see that ecLand systematically overestimates LE during the growing season. Fine-tuning pulls the model toward the observed flux magnitude while preserving the seasonal phase and inter-annual variability.

Reductions in LE and H RMSE can, in principle, be achieved either by improving the emulator's energy partitioning or by damping flux magnitudes toward zero. To distinguish these mechanisms, we evaluate two physical-consistency metrics at the FLUXNET sites: the Bowen ratio (BR=H/LE), which measures how surface available energy is partitioned between sensible and latent fluxes, and the total turbulent energy balance (H+LE error relative to observations), which tests whether any gains come at the cost of closure. ecLand's median Bowen ratio at validation sites is 0.44 against an observed median of 0.53 (Table ), reflecting a systematic tendency to partition too much available energy into latent heat relative to sensible heat – consistent with the LE overestimation identified in Sect.  and with analogous biases reported for other LSMs in PLUMBER2 . All fine-tuned strategies correct this partitioning bias: per-site Bowen ratio MedAE drops from 0.26 (aiLand-base) and 0.22 (ecLand) to between 0.15 and 0.20 across strategies, with aiLand-S4 achieving the lowest value (0.15, a 42 % reduction relative to aiLand-base). Notably, aiLand-S5 recovers the observed median BR exactly but has a higher MedAE (0.20) than aiLand-S4, indicating that a correct bulk ratio does not guarantee correct site-level partitioning. This confirms that the flux RMSE improvements in Sect.  reflect genuine energy-partitioning corrections rather than magnitude damping.

Energy balance closure provides a stricter test. ecLand overestimates both fluxes at validation sites (+8.84 W m⁻² for LE, +4.54 W m⁻² for H; time-averaged per site), producing a compounding closure offset of +13.38 W m⁻². Fine-tuning reverses this bias: all strategies slightly underestimate the turbulent flux sum, with aiLand-S5 achieving near-perfect closure (-0.10 W m⁻²) and aiLand-S4 second-best (-2.24 W m⁻²). aiLand-S2 retains the largest offset among fine-tuned models (-3.42 W m⁻²), consistent with its tendency to overcorrect LE (-4.47 W m⁻²). Because the observational target used here is an energy-balance-closed flux budget, these results indicate that fine-tuning recovers a physically consistent surface energy balance at the tower scale.

Fine-tuning also produces physically consistent changes in the screen-level diagnostic variables (2 m temperature, 2 m dew point, skin temperature; Fig. ). Across strategies (excluding aiLand-S2), reduced evaporative cooling at the surface produces a modest warming of skin temperature (+0.02 to +0.05 K) and 2 m temperature (+0.03 to +0.05 K) relative to aiLand-base, while 2 m dew point decreases slightly (-0.01 to -0.04 K), consistent with reduced moisture flux into the boundary layer. aiLand-S4 is notable for near-zero annual temperature change (Δ2t=0.00 K, Δskt=-0.01 K), reflecting its more conservative two-phase training schedule. The seasonal partitioning of this response is informative: skin temperature warming peaks in MAM (+0.06 to +0.12 K all strategies excluding aiLand-S2) and remains positive in DJF and SON for most strategies (+0.02 to +0.06 K), but is near-neutral in JJA (-0.02 to -0.00 K, excluding aiLand-S4). aiLand-S4 shows slight cooling in all seasons outside MAM (-0.02 to -0.07 K), reflecting its two-phase schedule. Despite the largest absolute LE correction occurring in summer (∼17 W m⁻² in JJA, where aiLand-base overestimates the total upward turbulent flux), fine-tuning corrects this primarily through LE reduction while H changes remain an order of magnitude smaller (<2 W m⁻²), leaving surface temperature near its baseline value. aiLand-S2 is an exception – its deep diagnostic head produces a cooler skin temperature (-0.18 K annually) and 2 m temperature (-0.08 K) despite similar LE reductions, indicating that an over-parameterised diagnostic-only architecture can decouple the temperature prediction from the surface energy balance.

Strategies aiLand-S1 and aiLand-S2 freeze the prognostic backbone, so their state trajectories are identical to aiLand-base by construction. The backbone-adaptive strategies (aiLand-S3, aiLand-S4, aiLand-S5) modify the backbone and could, in principle, improve flux predictions at the expense of degraded soil moisture or temperature. To assess this, we compare the free-running prognostic trajectories of each fine-tuned model against aiLand-base over the full 24-year period (2000–2023), with focus on the out-of-sample evaluation period (Table ).

Across the backbone-adaptive strategies, soil moisture is slightly wetter than aiLand-base by 0.3–1.8×10-3 m³ m⁻³ and soil temperatures are marginally cooler by 0.02–0.16 K. These offsets are consistent between training and validation sites, indicating a genuine equilibrium shift rather than overfitting. Within-evaluation-period drift is small: soil moisture trends remain below 1×10-3 m³ m⁻³ over four years at training sites, rising to ∼3×10-3 m³ m⁻³ at validation sites – an order of magnitude smaller than the soil moisture range across regimes, but indicating somewhat more equilibrium shift at unseen locations. Soil temperature trends stay within 0.15 K across both site sets. Snow cover shows a comparable pattern, with validation-site drift of ∼0.014 in fractional cover over four years – small relative to the ∼0.06 snow-cover RMSE already present in the pretrained model (Table ) and unlikely to materially interact with the snow-coupled error regime identified in Sect. .

The sign of these changes is physically consistent with the flux corrections: the fine-tuned models reduce LE magnitude relative to aiLand-base (correcting a systematic overestimation of evapotranspiration), which is qualitatively consistent with the slight wetting of the soil profile, while the associated reduction in evaporative cooling is consistent with the marginal temperature decrease.

Comparing strategies on their combined flux skill and state integrity, aiLand-S3 achieves the smallest prognostic biases and near-zero drift in all variables, while aiLand-S4 and aiLand-S5 show comparable biases but slightly larger variability at validation sites – particularly in soil moisture drift (∼3×10⁻³ m³ m⁻³ over 4 years) and snow cover. Constraining the backbone learning rate (aiLand-S3) or freezing it entirely (aiLand-S1, aiLand-S2) therefore provides the most robust guard against prognostic instability, while aiLand-S4 and aiLand-S5 trade marginally more state uncertainty for comparable or better flux skill, with downstream effects on screen-level diagnostics consistent with this flux–state tradeoff.

Bringing together the different metrics, aiLand-S4 (two-phase) and aiLand-S5 (full low LR) emerge as the top-performing strategies, with the choice between them depending on application priorities. aiLand-S4 achieves the best flux RMSE, correlations, and per-site Bowen ratio skill (MedAE 0.15) at validation sites, making it preferable when flux prediction accuracy is the primary target. aiLand-S5 achieves near-perfect energy balance closure at validation sites (-0.10 W m⁻²), recovers the observed median Bowen ratio exactly (0.53), and shows the smallest training–validation generalisation gap, at some cost to training-site flux skill and per-site Bowen ratio accuracy. It is therefore preferable for coupled applications where bulk physical consistency of the surface energy budget matters most. Between the two, aiLand-S5 is also easier to implement since it requires only one phase of fine-tuning.

The same architecture trained from scratch on the fine-tuning dataset improves latent heat flux at the tower sites, but with opposing LE and H biases and a soil-moisture departure from ecLand three to six times that of aiLand-base, suggesting that the prognostic stability of aiLand-base comes from pretraining on the global grid.

Relating these results to the two error sources identified in Sect. : fine-tuning on flux observations reduces the diagnostic-variable errors that underlie the regime sensitivity of latent heat flux, with the largest validation improvements at water-limited sites (IT-Lsn -60.6 %, US-xGR -56.4 %) and consistent gains at boreal, temperate and Mediterranean sites. Snow-coupled errors are not addressed by diagnostic-head fine-tuning, as both originate in the prognostic increment function. Correcting these prognostic-level errors would require updating the state-evolution network itself, for example through fine-tuning on snow-specific observational data or online bias correction, which we leave to future work.

A consideration in interpreting these results is the scale mismatch between FLUXNET tower footprints (∼100 m) and the N320 reduced Gaussian grid (∼31 km at the equator, ∼60 km at mid-latitudes). The model represents grid-cell-averaged fluxes, whereas tower observations capture local conditions that may not be representative of the surrounding landscape – particularly at sites with heterogeneous land cover, complex terrain, or coastal boundaries. This mismatch contributes to irreducible representation error in the metrics and likely explains some of the site-level degradations observed at high-altitude (CH-Cha, US-Me1) and coastal (US-NC4) locations, where sub-grid variability is largest.

Similarly, while we use energy-balance-closed FLUXNET observations, some caveats remain. The closure correction distributes the residual energy proportionally between H and LE, preserving the observed Bowen ratio but making absolute flux magnitudes conditional on this convention rather than uniquely determined. Net radiation and ground heat flux measurements used in the correction have their own ∼10 %–15 % uncertainty , which propagates into the closed fluxes. Bowen ratio as a metric is also disproportionately sensitive to low-LE conditions where the ratio diverges.