Skip to main content
Back to timeline
arXivSource publication:

No training needed to forecast forest carbon: TabPFN-3.5 posts the lowest error on four state variables under high emissions

Synopsis

Process-based terrestrial biosphere models can compute global carbon budgets, but their cost limits global, high-resolution and scenario-rich simulation, and purpose-built deep emulators require task-specific design and training. This work recasts 40-year Ecosystem Demography forecasts as tabular regression and lets the pretrained tabular foundation model TabPFN-3.5 predict directly by in-context learning, with no weight updates. On CarbonGlobe's 50,779 test sites it has the lowest height error of the three models without training and soil-carbon error within 4% of DeepEDLSTM, but larger errors for AGB, LAI and the carbon fluxes; under SSP5-8.5 it is the most accurate model for height, AGB, soil carbon and LAI.

Source-provided article image: Tabular Foundation Models as Emulators for Forecasting Terrestrial Ecosystem Dynamics
arXiv

Interpretation

A tabular foundation model never trained on the task can directly predict forest state 40 years ahead from an ecosystem model, and is more accurate on height than two purpose-trained deep emulators. Earlier DeepED-style emulators required designing a recurrent network or Transformer for the task and training and tuning it; TabPFN-3.5 only places training sites in context and updates no weights. In CarbonGlobe's overall evaluation, models learn from a uniform 1/16 sample of sites and are tested on the other 50,779; TabPFN-3.5 has a height error of 1.62 m, against 1.67 m for DeepEDLSTM and 1.86 m for DeepEDTransformer.

The zero-training route is not uniformly better: on AGB, leaf area index and the three carbon fluxes, the trained DeepEDLSTM is more accurate. TabPFN-3.5's soil-carbon error is within 4% of DeepEDLSTM's, but its AGB error is 24% higher, LAI 23% higher, and the three fluxes 40% to 49% higher. Across the seven outputs of the same overall evaluation, DeepEDLSTM is the most accurate of the three models for six of the seven, and TabPFN-3.5 leads only on height.

Moved to future climate scenarios, models trained only on present-day data lose accuracy by several-fold, yet under the high-emission scenario TabPFN-3.5 is the steadiest on four state variables. At 1,858 sites in six continental forest regions, year-2060 height and AGB errors grow 2 to 8 times under SSP1-2.6 and 3 to 12 times under SSP5-8.5; under SSP5-8.5 TabPFN-3.5's AGB error is 2.7 kg C m-2, less than half of DeepEDLSTM's 6.5 and DeepEDTransformer's 6.2. None of these sites is in the training or validation split, and the models see only the present-day 1/16 training sample; under SSP5-8.5 TabPFN-3.5 has the lowest AGB error in all six regions, while under SSP1-2.6 no model is best overall.

Perspective

The evaluation is confined to CarbonGlobe's ED v3.0 simulations: the present-day part trains on a 1/16 sample of sites and tests on the other 50,779, and the future part uses GFDL-ESM4-forced SSP1-2.6 and SSP5-8.5 scenarios at 1,858 sites in six regions, comparing only year 40.

TabPFN-3.5's future advantage appears to come from responding little to the change in forcing rather than from following ED: its predicted 2060 biomass is about the same under both scenarios, whereas ED's is higher under SSP5-8.5, and every model underestimates the biomass ED gains by 2060. The future test uses one Earth system model and two scenarios and changes the source of the forcing as well as the climate; TabPFN-3.5 sees only 40-year means of each driver, not the order of events or the final year's weather.