betaFeedback

WorksComputer Science

Tabular Foundation Models as Emulators for Forecasting Terrestrial Ecosystem Dynamics

Abstract

Process-based terrestrial biosphere models such as the Ecosystem Demography (ED) model underpin global carbon budgets, but their cost limits global, high-resolution and scenario-rich simulation. Deep emulators such as DeepED reduce that cost but require task-specific architectures and training. Tabular foundation models offer a different route: TabPFN predicts new rows by in-context learning, without being trained on the task. We compare TabPFN-3.5 with two DeepED baselines, DeepEDLSTMDeepED_{LSTM} and DeepEDTransformerDeepED_{Transformer}, on CarbonGlobe's overall evaluation, which trains on a uniform 1/16 sample of sites and tests on the other 50,779, predicting the 40-year state of all seven ED outputs. We then test every model, trained on present-day data only, on ED runs forced by SSP1-2.6 and SSP5-8.5 for 2021–2060 in six continental forest regions. With no training, TabPFN-3.5 has the lowest present-day height error of the three models and is within 4% of DeepEDLSTMDeepED_{LSTM} for soil carbon, but is less accurate for AGB, LAI and the carbon fluxes. In 2060, errors are several times larger than at the same sites under present-day forcing, and every model underestimates the biomass that ED gains. Under SSP5-8.5, TabPFN-3.5 is the most accurate model for height, AGB, soil carbon and LAI, with the lowest AGB error in all six regions, although, unlike ED, it predicts no larger biomass gain than under SSP1-2.6. Under SSP1-2.6, no model is best overall.

Introduction

Forests store a large share of terrestrial carbon, and how much they keep absorbing under a changing climate is a central uncertainty of the global carbon cycle [11]. Process-based models answer such questions by simulating vegetation demography and biogeochemistry explicitly. The Ecosystem Demography model (ED; Moorcroft et al. [23], Hurtt et al. [14], Ma et al. [22]) represents the size- and age-structure of forest stands and their carbon fluxes; it contributes to the Global Carbon Budget and supports NASA’s Carbon Monitoring System [32]. Its cost, however, grows with spatial resolution, the number of scenarios and the length of the simulation, which limits exactly the global, scenario-rich runs that policy needs.

Machine-learning emulators trained on ED outputs can replace most of this computation. DeepED [31] emulates ED’s 40-year carbon trajectories with recurrent networks and age-aware inputs, and the CarbonGlobe benchmark [32] made such emulation a standard machine-learning task: global coverage at 0.5∘0.5^\circ, 54,152 sites, 136 monthly drivers, 15 initial forest ages and seven carbon and vegetation outputs over 40 years. These emulators are effective, but each is a task-specific architecture that must be designed, trained and tuned.

Foundation models suggest another route. Pretrained once at scale, they are applied to new tasks with little or no task-specific training [5]; in Earth science, foundation models now

Preprint. forecast weather and climate [24, 4], and time-series foundation models forecast unseen series zero-shot [1, 7]. For tabular data, TabPFN [12, 13] is a transformer pretrained on synthetic tasks that predicts the targets of new rows by in-context learning: the training rows are its context, and no weights are updated. Its newest version, TabPFN-3.5 [15], takes contexts far larger than the ten thousand rows that TabPFN v2 was built for. Whether such a model can emulate a process-based ecosystem model, and how it compares with purpose-built emulators in the present and in a future climate, has not been tested.

We explore this question on CarbonGlobe. We cast ED emulation as tabular regression, with one row per forest stand (a site started from one of its initial states) and one in-context fit per output variable, and compare TabPFN-3.5 with two DeepED baselines: DeepEDLSTM, built on a recurrent network, and DeepEDTransformer_{\mathrm{Transformer}}, built on a Transformer. Our contributions are:

  • A tabular foundation model as an ED emulator. Without any training, TabPFN-3.5 predicts the 40-year state of all seven ED outputs from in-context fits on CarbonGlobe’s training sites (§ 4).

  • A present-day comparison with trained emulators. On CarbonGlobe’s overall evaluation, 50,779 test sites, TabPFN-3.5 has the lowest height error and is close to the best baseline for soil carbon, but is less accurate for AGB, LAI and the carbon fluxes (§ 5.1).

  • A future-climate test in six continental forest regions. All models, trained on present-day data only, are evaluated on ED runs under SSP1-2.6 and SSP5-8.5 for 2021–2060 at 1,858 forest sites in six regions, one per continent, and compared with the same sites under present-day forcing (Figure 1; § 5.2). Errors grow several-fold, and every model underestimates the biomass ED gains by 2060. Under SSP5-8.5, TabPFN-3.5 is the most accurate model for height, AGB, soil carbon and LAI.

Study area of the future-climate evaluation

Figure 1. Study area of the future-climate evaluation. Shading: aboveground biomass (AGB) of the oldest initial state (500-year stand) of each CarbonGlobe cell where that stand is at least 5 m tall; other land in gray. Boxes: the six regions, with the number of sites used in each (§ 3, Table 1).

Related work

Process-based models and their emulators. Vegetation demographic models such as ED represent the size and age structure of forests [23, 10]. Machine learning increasingly emulates or complements process models in Earth science [28], often guided by domain knowledge [17, 16]. Emulators of land models support parameter estimation [6], and deep emulators now approximate numerical weather prediction at global scale [18, 19, 3], with benchmarks such as WeatherBench 2 [27]. For forest carbon, DeepED [31] introduced long-horizon emulation of ED with age-aware recurrent networks, and CarbonGlobe [32] released a global ML-ready dataset with recurrent, transformer-based [30, 35] and linear baselines, evaluated with the RMSE, the MAE and two problem-driven metrics, the delta and cumulative errors. EcoDiffusion [20] makes such emulation uncertainty-aware, generating long sequences from a single initial step with a conditional diffusion model.

Emulators under climate change. What-if simulations, such as future climate scenarios, are a central use of emulators, and an Earth system world model has recently been built for them, with terrestrial ecosystems as a case study [33]. An emulator trained on the present climate, however, has to extrapolate in a future scenario: temperature, precipitation and CO2_2 leave the range it was trained on. Beucler et al. [2024] showed that machine-learning parameterizations trained in one climate generalize poorly to a warmer one unless their inputs are made climate-invariant. We test this setting for ED emulators, with scenario runs at sites never used for training.

Foundation models for science and time series. Pretrained models for the atmosphere and Earth system [24, 4] and for time series [1, 7, 34] transfer to new tasks with little training. Time-series models forecast from past observations of the series itself; ED emulation instead starts from an initial ecosystem state and a long sequence of forcing, a setting closer to conditional regression.

Tabular foundation models. TabPFN [12] showed that a transformer pretrained on synthetic data can solve small tabular tasks in context; TabPFN v2 [13] extended this to regression and to datasets of up to about ten thousand rows. Later models scale the context further [26, 21, 15] or retrieve a local context for each test row [29]. To our knowledge, tabular foundation models have not been evaluated as emulators of process-based ecosystem models, nor under a climate shift.

Problem definition

Data. CarbonGlobe [32] contains ED v3.0 simulations at 54,152 land sites on a 0.5∘0.5^\circ global grid. At each site, ED is started from 15 initial forest states, taken at stand ages a∈{1,10,20,30,50,70,100,150,200,250,300,350,400,450,500}a \in\{1, 10, 20, 30, 50, 70, 100, 150, 200, 250, 300, 350, 400, 450, 500\} years, and run for the 40 years from 1981 under observation-based forcing (MERRA-2 meteorology and NOAA CarbonTracker CO2_2). We call a site started from one of its initial states a stand. The forcing X∈R40×12×136X \in\mathbb{R}^{40 \times12 \times136} holds 136 monthly drivers: air temperature, precipitation and soil temperature at six depths (8); disturbance (1); diurnal cycles, as 24 hourly values, of atmospheric CO2_2, specific humidity, air temperature, shortwave radiation and wind speed (120); and seven soil properties. The state yt∈R7y_t \in\mathbb{R}^7 is annual: vegetation height (m), aboveground biomass (AGB), soil carbon, leaf area index (LAI, m2^2 m−2^{-2}), gross and net primary production (GPP, NPP) and heterotrophic respiration (Rh); carbon stocks and fluxes are in kg C m−2^{-2} (fluxes as annual totals).

Task. Given a stand’s initial state y0y_0 at age aa and the 40-year forcing XX of its site, predict the state of all seven outputs after 40 years, y40y_{40}. The deep emulators produce the whole trajectory y1,…,y40y_1, \ldots, y_{40} on the way; TabPFN-3.5 predicts y40y_{40} directly.

Present-day evaluation. We follow CarbonGlobe’s overall evaluation [32]. Models learn from a uniform 1/16 sample of the sites, every fourth cell in latitude and longitude (train4: 3,373 sites on a 2∘2^\circ grid, of which 337 are held out for early stopping), and are tested on all other 50,779 sites (761,685 stands). Uniform sampling mirrors how an emulator would be used: ED runs on a coarse grid train it, and it then fills in the fine grid.

Future-climate runs. To test generalization beyond the present-day climate, we use ED runs forced by the GFDL-ESM4 Earth system model [8] under SSP1-2.6 and SSP5-8.5, a low- and a high-emission scenario [25]. They start in 2020 from 15 initial states per site, at the same stand ages as above except that the youngest is 0 years, and we use their first 40 years, 2021–2060. These runs differ from the present-day runs in two ways: the climate changes, and the forcing comes from a climate model instead of observations.

Six forest regions. We evaluate on six forest regions, one per continent, that span tropical moist, subtropical, temperate and eucalypt forest (Figure 1, Table 1). Every site we use is a 0.5∘0.5^\circ cell inside a region that (i) is not in CarbonGlobe’s training or validation split and (ii) is forest: its oldest initial state, a 500-year stand, is at least 5 m tall in 2020 under both scenarios, the tree-height threshold of the FAO forest definition [9]. In all, we use 1,858 sites, each with 15 initial states. They learn from present-day train4 data only, so they have never seen these locations. Over 2021–2060 the regions are 0.9–1.2 K (SSP1-2.6) and 1.2–1.8 K (SSP5-8.5) warmer than in their present-day forcing. Atmospheric CO2_2, 411–442 ppm in the last present-day year at these sites, reaches 474 ppm (SSP1-2.6) and 643 ppm (SSP5-8.5) by 2060.

The six future-climate regions. Sites: 0.5∘0.5^\circ cells used, each with 15 initial states. AGB: median AGB of the 500-year stand in 2020 (kg C m−2\mathrm{kg\,C\,m^{-2}}). TT: mean annual air temperature of the present-day forcing (∘C^\circ\mathrm{C}). ΔT\Delta T: mean temperature of the GFDL-ESM4 forcing over 2021–2060 minus TT (K); it includes the difference between the climate model and the observation-based present-day forcing.

RegionBoundsForestSitesAGBTTSSP1-2.6SSP5-8.5
E. North America35–47°N, 90–70°Wtemperate broadleaf31612.29.1+1.2+1.8
Amazon10°S–2°N, 72–50°Wtropical moist43622.225.7+1.1+1.5
C. and E. Europe45–58°N, 5–30°Etemperate mixed44210.28.2+1.1+1.5
Congo Basin5°S–5°N, 12–30°Etropical moist31520.824.3+0.9+1.3
E. China22–32°N, 108–120°Esubtropical broadleaf18618.117.3+1.1+1.2
E. Australia44–25°S, 145–154°Eeucalypt1636.317.3+0.9+1.2
All six1,858

Table 1. The six future-climate regions. Sites: 0.5∘0.5^\circ cells used, each with 15 initial states. AGB: median AGB of the 500-year stand in 2020 (kg C m−2\mathrm{kg\,C\,m^{-2}}). TT: mean annual air temperature of the present-day forcing (∘C^\circ\mathrm{C}). ΔT\Delta T: mean temperature of the GFDL-ESM4 forcing over 2021–2060 minus TT (K); it includes the difference between the climate model and the observation-based present-day forcing.

Metric. For each output variable we report the root-mean-square error at year 40 over all stands, [1N∑i(y^i,40−yi,40)2]1/2\left[\frac{1}{N}\sum_i(\hat{y}_{i,40}-y_{i,40})^2\right]^{1/2}, in the variable’s physical units. This is CarbonGlobe’s cumulative error [32] in root-mean-square form: the error left at the end of the 40-year forecast, after it has accumulated over all years.

Methods

TabPFN-3.5 as an in-context emulator

Each stand is one row. Its features are the 40-year mean of each of the 136 drivers, the initial state y0y_0 (7), the stand age divided by 500 (1), and two age references from DeepED [31]: the mean initial state of all sites at that age and its slope across ages (14), 158 features in all. Drivers and states are standardized with CarbonGlobe’s statistics. One regressor predicts the year-40 value of each output variable, seven in all. A fit only places the context rows, the training stands with that target, in the model’s key-value cache: no weights are updated and nothing is tuned. We use TabPFN-3.5 [15] with two estimators, one per target transform of the checkpoint, take the mean of its predictive distribution, and use the 3,036 training sites of train4 as context (45,540 rows). The weights are used under Prior Labs’ TabPFN-3.5 Non-Commercial License (v1.0).

DeepED baselines

DeepED [31] emulates ED with a sequence model that steps from the initial state through the 40 years of forcing, with the stand age and age-dependent reference states as additional inputs. We use two versions as baselines.

DeepEDLSTM_{\mathrm{LSTM}}. We use DeepED’s LSTM with an improved design. Each year, an LSTM with 160 units reads the year’s 12 monthly drivers together with the previous state, the stand age and the reference states, and a small multilayer perceptron turns the year’s summary into the change of all seven states. The network is trained through the whole 40-year rollout on its own predictions, so training sees the same error accumulation as forecasting (AdamW with learning rate 10−310^{-3} halved on plateaus, batches of 1,024 stands, gradient clipping at 1, early stopping on the held-out sites). It has 0.29 M parameters.

DeepEDTransformer. An Informer encoder–decoder [35] replaces the recurrent network and predicts all 40 years in one pass. Each token is one month. The encoder reads 12 tokens, the first year’s drivers with the initial state; the decoder reads these 12 tokens followed by the 480 forecast months with the state set to zero, and its output at the last month of each forecast year is that year’s state. The tokens also carry the stand age and the reference states. It has 11.8 M parameters and is trained on the squared error of the states plus half the squared error of their year-to-year changes (Adam with learning rate 10−410^{-4} halved every epoch, batches of 16 stands, early stopping on the held-out sites).

Protocol

Each deep baseline is trained with three seeds, and we report the mean prediction of the three. Early stopping uses only the 337 held-out train4 sites, and the six regions were fixed before any model was scored on them. Future inputs are standardized with the same present-day statistics as the training data. All experiments run on NVIDIA A100 GPUs (40 GB).

Experiments

Present day

Table 2 gives the year-40 RMSE of each output on CarbonGlobe’s 50,779 test sites.

ModelHeightAGBSoil CLAIGPPNPPRh
TabPFN-3.51.6170.5850.5510.3770.2560.1270.143
DeepEDLSTM1.6670.4730.5330.3060.1720.0850.102
DeepEDTransformer1.8560.6310.6600.3530.2040.1010.125

Table 2. Present day, CarbonGlobe’s overall evaluation: year-40 RMSE over the 50,779 test sites (761,685 stands); every model learns from train4 only. Height in m, LAI in m2^2 m−2^{-2}, the other variables in kg C m−2^{-2} (fluxes per year). DeepEDLSTM and DeepEDTransformer: mean prediction of three training seeds. Best per column in bold.

Without training, TabPFN-3.5 is competitive for height and soil carbon. DeepEDLSTM is the most accurate of the three models for six of the seven outputs (Table 2). TabPFN-3.5, which is never trained, has the lowest height error of the three (1.62 m, against 1.67 m for DeepEDLSTM and 1.86 m for DeepEDTransformer), and its soil-carbon error is within 4% of DeepEDLSTM’s. Its errors are larger for AGB (+24%) and LAI (+23%), and largest for the three fluxes (+40 to +49%). A likely reason is that annual fluxes follow the weather of the final year, which the 40-year mean drivers seen by TabPFN-3.5 do not resolve. DeepEDTransformer is less accurate than DeepEDLSTM for every output, and less accurate than TabPFN-3.5 for height, AGB and soil carbon. Averaging three training seeds helps the deep baselines: single DeepEDLSTM seeds have AGB errors of 0.515–0.529 kg C m−2^{-2}, their mean prediction 0.473.

Future climate in six regions

Table 3 compares each model’s year-40 error at the 1,858 region sites under present-day forcing with its error in 2060 under the two scenarios, and Table 4 breaks the AGB error down by region.

Table 3: Six future regions, 1,858 sites (27,870 stands): year-40 RMSE under present-day forcing (the same sites inside CarbonGlobe’s test set) and in 2060 under SSP1-2.6 and SSP5-8.5. All models learn from present-day train4 data only. Units as in Table 2; best per column and block in bold.

ModelHeightAGBSoil CLAIGPPNPPRh
Present day, the same sites
TabPFN-3.51.0610.6880.5170.4540.2730.1340.176
DeepEDLSTM_{\mathrm{LSTM}}1.2330.5280.4620.3110.1800.0880.159
DeepEDTransformer_{\mathrm{Transformer}}1.4120.9090.6830.4130.2300.1140.181
SSP1-2.6, 2060
TabPFN-3.52.8092.0661.4821.4611.1460.5820.457
DeepEDLSTM_{\mathrm{LSTM}}3.7494.2251.4320.8360.6840.3510.247
DeepEDTransformer_{\mathrm{Transformer}}2.5442.4071.6211.3210.8840.4670.402
SSP5-8.5, 2060
TabPFN-3.53.5652.7461.9661.2431.4640.7370.513
DeepEDLSTM_{\mathrm{LSTM}}6.1546.5102.1701.7801.0060.5030.462
DeepEDTransformer_{\mathrm{Transformer}}6.5396.2123.6611.4361.6720.8570.493

Table 3.

Table 4: Year-40 AGB RMSE (kg C m−2^{-2}) in each region (Table 1): present-day forcing at the same sites, and 2060 under SSP1-2.6 and SSP5-8.5. Best per column and block in bold.

ModelE. North AmericaAmazonC. and E. EuropeCongo BasinE. ChinaE. Australia
Present day, the same sites
TabPFN-3.50.201.170.330.610.290.73
DeepEDLSTM_{\mathrm{LSTM}}0.210.820.350.480.350.63
DeepEDTransformer_{\mathrm{Transformer}}0.291.540.430.770.471.00
SSP1-2.6, 2060
TabPFN-3.51.262.361.042.962.491.93
DeepEDLSTM_{\mathrm{LSTM}}0.608.311.201.792.201.50
DeepEDTransformer_{\mathrm{Transformer}}1.442.751.533.512.602.00
SSP5-8.5, 2060
TabPFN-3.52.032.411.614.124.042.03
DeepEDLSTM_{\mathrm{LSTM}}3.7710.274.015.616.913.81
DeepEDTransformer_{\mathrm{Transformer}}4.256.843.649.137.663.95

Table 4.

Errors grow several-fold, and every model underestimates ED’s biomass gain. Under present-day forcing, the errors at the region sites are of the same order as on the whole test set (Table 3). In 2060 the same models are much less accurate: height and AGB errors are 2–8 times larger under SSP1-2.6 and 3–12 times larger under SSP5-8.5. Averaged over the region stands, ED’s AGB grows between 2020 and 2060 by 1.4 kg C m−2^{-2} under SSP1-2.6 and by 2.3 under SSP5-8.5, to 11.1 and 12.1. TabPFN-3.5 predicts smaller gains, 0.6 and 0.4, and both DeepED baselines predict losses: 0.9 and 3.2 for DeepEDLSTM, 0.4 and 2.7 for DeepEDTransformer. No model reproduces ED’s larger gain under the high-emission scenario: TabPFN-3.5 predicts a slightly smaller gain under SSP5-8.5, and the deep baselines larger losses.

Under SSP5-8.5, TabPFN-3.5 is the most accurate model for height, AGB, soil carbon and LAI. It also has the lowest AGB error in all six regions (Table 4); over all regions its AGB error, 2.7 kg C m−2^{-2}, is less than half that of DeepEDLSTM_{\mathrm{LSTM}} (6.5) or DeepEDTransformer_{\mathrm{Transformer}} (6.2). DeepEDLSTM_{\mathrm{LSTM}} remains the most accurate model for the three fluxes.

Under SSP1-2.6, no model is best overall. DeepEDLSTM_{\mathrm{LSTM}} has the lowest errors for soil carbon, LAI and the three fluxes, and the lowest AGB error in four of the six regions, but its AGB error in the Amazon, 8.3 kg C m−2^{-2} against 2.4 for TabPFN-3.5, makes it the least accurate model for AGB over all regions. TabPFN-3.5 has the lowest AGB error over all regions, and DeepEDTransformer_{\mathrm{Transformer}} the lowest height error.

The deep baselines fail most in tropical forest, and depend on the training seed. Their largest AGB errors are in the two tropical regions: the Amazon for DeepEDLSTM_{\mathrm{LSTM}} and the Congo Basin for DeepEDTransformer_{\mathrm{Transformer}} (Table 4). In the Amazon, the warmest region, precipitation also falls by 9% (SSP1-2.6) and 14% (SSP5-8.5); there DeepEDLSTM_{\mathrm{LSTM}} predicts AGB to drop to 5.6 kg C m−2^{-2} by 2060 under SSP5-8.5, while ED keeps it near 15. Their future errors also vary far more with the training seed than in the present day: under SSP5-8.5, single DeepEDLSTM_{\mathrm{LSTM}} seeds have AGB errors from 3.7 to 9.6 kg C m−2^{-2} and single DeepEDTransformer_{\mathrm{Transformer}} seeds from 5.8 to 6.9, against 0.515–0.529 and 0.637–0.653 in the present day.

Discussion and limitations

What the tabular route offers. Given only CarbonGlobe’s 1/16 training sample as context, with no training and no tuning, a pretrained tabular model predicts ED’s 40-year height and soil carbon about as accurately as emulators designed and trained for the task, though not AGB, LAI and the fluxes. Under the larger climate shift of SSP5-8.5 it is the most accurate of the three models for height, AGB, soil carbon and LAI, while the trained emulators drift far from ED in some regions and vary with the training seed.

What no model captures. TabPFN-3.5’s advantage in the future appears to come from responding little to the change in forcing, not from following ED: its mean 2060 biomass is about the same under both scenarios, whereas ED’s is higher under SSP5-8.5. No model trained on present-day data reproduces the biomass that ED gains by 2060. Emulators meant for scenario studies therefore need either training data from future climates or knowledge of the processes that make biomass respond to them.

Limitations. (i) We compare the models only at year 40. TabPFN-3.5 can predict any forecast year with its own in-context fit, which would allow a comparison along the trajectory with CarbonGlobe’s delta error. (ii) TabPFN-3.5’s climate features are 40-year means of each driver, so it sees neither the order of events within the 40 years nor the weather of the final year. (iii) All models learn from train4 only, and TabPFN-3.5 is not fine-tuned; whether more present-day data or fine-tuning helps in a future climate is not tested. (iv) The future test uses one Earth system model and two scenarios, and it changes the source of the forcing as well as the climate.

Conclusion

We explored a tabular foundation model as an emulator of the ED model on CarbonGlobe. With no training, TabPFN-3.5 predicts the 40-year height more accurately than both DeepED baselines and soil carbon within 4% of the better one, DeepEDLSTM_{\mathrm{LSTM}}, but is less accurate for AGB, LAI and the carbon fluxes. In six continental forest regions under SSP1-2.6 and SSP5-8.5, the errors of every model trained on present-day data grow several-fold, and none reproduces the biomass that ED gains by 2060. Under SSP5-8.5, TabPFN-3.5 is the most accurate model for height, AGB, soil carbon and LAI, while the deep baselines’ errors are largest in tropical forest and vary strongly with the training seed.

References

  1. [1]Abdul Fatir Ansari, Lorenzo Stella, Caner Turkmen, et al. Chronos: Learning the language of time series. Transactions on Machine Learning Research, 2024.arxiv.org/abs/2403.07815
  2. [3]Kaifeng Bi, Lingxi Xie, Hengheng Zhang, Xin Chen, Xiaotao Gu, and Qi Tian. Accurate medium-range global weather forecasting with 3D neural networks. Nature, 619:533–538, 2023.DOI
  3. [4]Cristian Bodnar, Wessel P. Bruinsma, Ana Lucic, Megan Stanley, et al. A foundation model for the Earth system. Nature, 641(8065):1180–1187, 2025.DOI
  4. [5]Rishi Bommasani, Drew A. Hudson, Ehsan Adeli, et al. On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258, 2021.
  5. [6]Katherine Dagon, Benjamin M. Sanderson, Rosie A. Fisher, and David M. Lawrence. A machine learning approach to emulation and biophysical parameter estimation with the Community Land Model, version 5. Advances in Statistical Climatology, Meteorology and Oceanography, 6:223–244, 2020.DOI
  6. [7]Abhimanyu Das, Weihao Kong, Rajat Sen, and Yichen Zhou. A decoder-only foundation model for time-series forecasting. In International Conference on Machine Learning (ICML), 2024.arxiv.org/abs/2310.10688
  7. [8]John P. Dunne, Larry W. Horowitz, Alistair J. Adcroft, et al. The GFDL earth system model version 4.1 (GFDL-ESM 4.1): Overall coupled model description and simulation characteristics. Journal of Advances in Modeling Earth Systems, 12(11):e2019MS002015, 2020.DOI
  8. [9]FAO. Global forest resources assessment 2020: Terms and definitions. Forest Resources Assessment Working Paper 188, Food and Agriculture Organization of the United Nations, Rome, 2018.
  9. [10]Rosie A. Fisher, Charles D. Koven, William R. L. Anderegg, Bradley O. Christoffersen, Michael C. Dietze, Caroline E. Farrior, Jennifer A. Holm, George C. Hurtt, Ryan G. Knox, Peter J. Lawrence, et al. Vegetation demographics in earth system models: A review of progress and priorities. Global Change Biology, 24(1):35–54, 2018.DOI
  10. [11]Pierre Friedlingstein, Michael O’Sullivan, Matthew W. Jones, Robbie M. Andrew, et al. Global carbon budget 2023. Earth System Science Data, 15(12):5301–5369, 2023.
  11. [12]Noah Hollmann, Samuel Müller, Katharina Eggensperger, and Frank Hutter. TabPFN: A transformer that solves small tabular classification problems in a second. In International Conference on Learning Representations (ICLR), 2023.arxiv.org/abs/2207.01848
  12. [13]Noah Hollmann, Samuel Müller, Lennart Purucker, Arjun Krishnakumar, Max Körfer, Shi Bin Hoo, Robin Tibor Schirrmeister, and Frank Hutter. Accurate predictions on small data with a tabular foundation model. Nature, 637:319–326, 2025.DOI
  13. [14]George C. Hurtt, Paul R. Moorcroft, Stephen W. Pacala, and Simon A. Levin. Terrestrial models and global change: challenges for the future. Global Change Biology, 4(5):581–590, 1998.DOI
  14. [15]Benjamin Jäger, Nick Erickson, Léo Grinsztajn, Felix Birkel, Klemens Flöge, et al. TabPFN-3.5: Technical report. arXiv preprint arXiv:2609.17895, 2026.arxiv.org/abs/2609.17895
  15. [16]Xiaowei Jia, Jared Willard, Anuj Karpatne, Jordan S. Read, Jacob A. Zwart, Michael Steinbach, and Vipin Kumar. Physics-guided machine learning for scientific discovery: An application in simulating lake temperature profiles. ACM/IMS Transactions on Data Science, 2(3):1–26, 2021.arxiv.org/abs/2001.11086
  16. [17]Anuj Karpatne, Gowtham Atluri, James H. Faghmous, Michael Steinbach, Arindam Banerjee, Au-roop Ganguly, Shashi Shekhar, Nagiza Samatova, and Vipin Kumar. Theory-guided data science: A new paradigm for scientific discovery from data. IEEE Transactions on Knowledge and Data Engineering, 29(10):2318–2331, 2017.
  17. [18]Thorsten Kurth, Shashank Subramanian, Peter Harrington, Jaideep Pathak, Morteza Mardani, David Hall, Andrea Miele, Karthik Kashinath, and Anima Anandkumar. FourCastNet: Accelerating global high-resolution weather forecasting using adaptive Fourier neural operators. In Proceedings of the Platform for Advanced Scientific Computing Conference, pages 1–11, 2023.arxiv.org/abs/2208.05419
  18. [19]Remi Lam, Alvaro Sanchez-Gonzalez, Matthew Willson, Peter Wirnsberger, Meire Fortunato, Ferran Alet, Suman Ravuri, Timo Ewalds, Zach Eaton-Rosen, Weihua Hu, et al. Learning skillful medium-range global weather forecasting. Science, 382(6677):1416–1421, 2023.DOI
  19. [20]Ruohan Li, Zhihao Wang, Xiaowei Jia, Gengchen Mai, Lei Ma, George C Hurtt, Quan Shen, Zhili Li, and Yiqun Xie. Ecodiffusion: Uncertainty-aware emulation of ecosystem processes with conditional diffusion for long sequences with single-step initialization. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 40, pages 38880–38888, 2026.DOI
  20. [21]Junwei Ma, Valentin Thomas, Rasa Hosseinzadeh, Hamidreza Kamkari, Alex Labach, Jesse C. Cresswell, Keyvan Golestan, Guangwei Yu, Maksims Volkovs, and Anthony L. Caterini. TabDPT: Scaling tabular foundation models. arXiv preprint arXiv:2410.18164, 2024.arxiv.org/abs/2410.18164
  21. [22]Lei Ma, George Hurtt, Lesley Ott, Ritvik Sahajpal, Justin Fisk, Rachel Lamb, Hao Tang, Steve Flanagan, Louise Chini, Abhishek Chatterjee, et al. Global evaluation of the ecosystem demography model (ED v3.0). Geoscientific Model Development, 15(5):1971–1994, 2022.DOI
  22. [23]Paul R. Moorcroft, George C. Hurtt, and Stephen W. Pacala. A method for scaling vegetation dynamics: the ecosystem demography model (ED). Ecological Monographs, 71(4):557–586, 2001.
  23. [24]Tung Nguyen, Johannes Brandstetter, Ashish Kapoor, Jayesh K. Gupta, and Aditya Grover. ClimaX: A foundation model for weather and climate. In International Conference on Machine Learning (ICML), 2023.arxiv.org/abs/2301.10343
  24. [25]Brian C. O’Neill, Claudia Tebaldi, Detlef P. van Vuuren, Veronika Eyring, et al. The scenario model intercomparison project (ScenarioMIP) for CMIP6. Geoscientific Model Development, 9: 3461–3482, 2016.DOI
  25. [26]Jingang Qu, David Holzmüller, Gaël Varoquaux, and Marine Le Morvan. TabICL: A tabular foundation model for in-context learning on large data. In International Conference on Machine Learning (ICML), volume 267 of Proceedings of Machine Learning Research, pages 50817–50847, 2025.arxiv.org/abs/2502.05564
  26. [27]Stephan Rasp, Stephan Hoyer, Alexander Merose, Ian Langmore, Peter Battaglia, et al. WeatherBench 2: A benchmark for the next generation of data-driven global weather models. Journal of Advances in Modeling Earth Systems, 16(6):e2023MS004019, 2024.arxiv.org/abs/2308.15560
  27. [28]Markus Reichstein, Gustau Camps-Valls, Bjorn Stevens, Martin Jung, Joachim Denzler, Nuno Carvalhais, and Prabhat. Deep learning and process understanding for data-driven Earth system science. Nature, 566:195–204, 2019.
  28. [29]Valentin Thomas, Junwei Ma, Rasa Hosseinzadeh, Keyvan Golestan, Guangwei Yu, Maksims Volkovs, and Anthony L. Caterini. Retrieval & fine-tuning for in-context tabular models. In Advances in Neural Information Processing Systems (NeurIPS), volume 37, 2024.arxiv.org/abs/2406.05207
  29. [30]Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems (NeurIPS), 2017.arxiv.org/abs/1706.03762
  30. [31]Zihao Wang, Yiqun Xie, Xiaowei Jia, Lei Ma, and George Hurtt. High-fidelity deep approximation of ecosystem simulation over long-term at large scale. In Proceedings of the 31st ACM International Conference on Advances in Geographic Information Systems, pages 1–10, 2023.DOI
  31. [32]Zihao Wang, Lei Ma, George Hurtt, Xiaowei Jia, Yanhua Li, Ruohan Li, Zhili Li, Shuo Xu, and Yiqun Xie. CarbonGlobe: A global-scale, multi-decade dataset and benchmark for carbon forecasting in forest ecosystems. In Advances in Neural Information Processing Systems (NeurIPS), Datasets and Benchmarks Track, 2025.DOI
  32. [33]Zihao Wang, Ruichen Wang, Ruohan Li, Lei Ma, George Hurtt, Xiaowei Jia, Gengchen Mai, Shaowen Wang, and Yiqun Xie. Earth system world model for what-if simulations: A case study for terrestrial ecosystems. arXiv preprint arXiv:2609.08855, 2026.arxiv.org/abs/2609.08855
  33. [34]Gerald Woo, Chenghao Liu, Akshat Kumar, Caiming Xiong, Silvio Savarese, and Doyen Sahoo. Unified training of universal time series forecasting transformers. In International Conference on Machine Learning (ICML), 2024.arxiv.org/abs/2402.02592
  34. [35]Haoyi Zhou, Shanghang Zhang, Jieqi Peng, Shuai Zhang, Jianxin Li, Hui Xiong, and Wancai Zhang. Informer: Beyond efficient transformer for long sequence time-series forecasting. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 11106–11115, 2021.arxiv.org/abs/2012.07436

Paper details

vlnc:W_xnsfx4zb
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 10

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

Not recorded for this paper

View statement
Looking for related papers…
Contents