Research

The science in ABC3^3 generally follows a dynamical system perspective: processes are studied coupled, not in isolation, to capture feedbacks, which provides the basis for inference based on causality, not correlation. Generally, we combine process-based modelling, causal inference on Earth-observation data streams, and physics-constrained machine learning (hybrid modelling) to push the frontier of climate modeling and uncertainty reduction in future projections. The research themes below span from global feedbacks rooted primarily in the carbon cycle to regional dynamics rooted mainly in the water and energy cycle.

Currently, our four main themes address phenology–climate feedbacks, the global carbon cycle approaching the post-peak era, CO2_2 fertilisation and its observational constraints, and hybrid Earth system modelling.

Research themes

Phenology and its feedbacks to climate

Phenological shifts and their land–atmosphere feedback pathways. © Winkler et al., (in review)

The seasonal rhythm of leaf emergence and senescence — phenology — is shifting under climate warming, with spring advancing by roughly 2 days per decade and autumn retreating by nearly as much across northern lands. Using satellite and reanalysis data, and FLUXNET data on carbon fluxes, we quantify how these shifts alter surface albedo, latent and sensible heat fluxes, and carbon fluxes across biomes. In a fully coupled Earth system model experiment, we then prescribe satellite-derived leaf area changes as a proxy for observed phenological trends to study all feedbacks associated with phenological shifts. Our idealized runs show that a 10-day growing-season extension triggers a global surface cooling of −0.10 ± 0.03 K, strongest at northern high latitudes — establishing phenology as an active regulator of the climate system, not merely a passive responder to it.

Key publications

Panwar et al. (2024)  ·  Hau et al. (accepted)  ·  Winkler et al. (in review)

Global carbon cycle modelling and the post-peak era

Land carbon dynamics from CO2_2 uptake through spill-over and hysteresis effects in the post-peak era. © Winkler et al., (in prep.)

Current climate targets rest on a near-linear relationship between cumulative emissions and warming — the Transient Climate Response to Cumulative Carbon Emissions (TCRE). Using latest-generation Earth system model simulations, we show that this proportionality breaks down: for the same total emissions (~1.2 Eg C, roughly the 2 °C budget), atmospheric CO2_2 can differ by ~100 ppm and warming by ~0.3 K depending on the timing of emissions. The discrepancy arises because slow land processes — ecological succession, shifts in vegetation composition, and soil carbon turnover — carry memory across decades to centuries, so the coupled carbon–climate system does not simply retrace its path after emissions peak. Building on this, we develop a new generation of state-dependent impulse-response functions that integrate machine learning with process-based modelling, enabling spatially explicit, history-aware climate attribution. My ERC Starting Grant PostPeak takes this further by fusing the full set of global land-cycle observations into a hybrid, physics-constrained framework to project land carbon dynamics in post-peak emission scenarios.

Key publications

Winkler et al. (2024)  ·  Winkler & Sierra, (2025)

CO2_2 fertilisation effects and observational constraints

Site-level CO2_2 fertilisation sensitivity (β\beta) from eddy-covariance flux towers across Europe and North America. © Zhan et al. 2024, JGR Biogeosciences.

Rising atmospheric CO2_2 stimulates photosynthesis and reduces stomatal conductance, impacting on the global cycles of carbon and water. Yet detecting and quantifying these physiological effects in real-world data is difficult because warming, rainfall anomalies, disturbances, and memory-effects rooted in regrowing forests and bio-climatic vegetation shifts, etc. act on ecosystems simultaneously. Applying causal counterfactual attribution to decades of satellite leaf area observations and a large ensemble of Earth system models, we show that the CO2_2 fertilisation effect dominates vegetation greening in only a few biomes — temperate forests and cool grasslands — while climate change increasingly counteracts it elsewhere, producing a global slowdown in greening and strengthening browning, especially in tropical rainforests.

Moving to independent ground measurements from eddy-covariance flux towers across dozens of extratropical forest sites, we estimate a median CO2_2 stimulation of gross primary productivity of ~3.2 gC m−2^{-2} yr−1^{-1} ppm−1^{-1}, providing direct observational evidence that the fertilisation effect persists but is smaller and more heterogeneous than many climate models assume. Process-based model experiments further reveal that the CO2_2 effect on carbon fluxes emerges at lower CO2_2 increases (~20 ppm) than its effect on transpiration (>40 ppm), explaining why water-cycle signals lag behind carbon-cycle signals in the observational record.

Key publications

Zhan et al. (2022)  ·  Zhan et al. (2024)  ·  Winkler et al. (2021)

Hybrid Earth system modelling

Machine-learning-enhanced vs. original ICON model response of canopy conductance and transpiration to atmospheric dryness. © ElGhawi et al. 2025, JAMES.

Deep learning models outperform process-based models in some tasks such as weather prediction, but long-term climate change projection is fundamentally an out-of-distribution problem with underlying state changes encoded in land and ocean processes that accumulate memory; we show in our research that AI and ML methods will not be able to provide reliable climate projection for such boundary conditions and state change. We demonstrate this concretely: diverse deep learning architectures trained on historical land CO2_2 uptake data achieve strong spatial cross-validation yet show virtually no skill under future climate scenarios, systematically rooted in misinterpreting the true causal links in the system, essentially being misled by spurious correlations.

To overcome this, we develop hybrid models that embed neural networks within mechanistic equations. In one example, we learn stomatal and aerodynamic resistances from FLUXNET eddy-covariance observations while respecting physical constraints. We then integrate these observation-based ML-parameterisations into the ICON Earth system model via a Python–Fortran bridge, replacing rigid empirical schemes for photosynthesis and transpiration. The resulting hybrid ICON model substantially reduces biases in water and carbon fluxes and produces a tighter, more realistic response to atmospheric dryness across forest and grassland biomes.

Key publications

ElGhawi et al. (2023)  ·  ElGhawi et al. (2025)  ·  ElGhawi & Winkler et al. (2025)

New emerging research topics in ABC3^3

Land states and extreme events

How would past extreme meteorological events — the 2003 Central-European heatwave, the 2018 drought — have played out if the land surface had been different? How will future meteorological extremes have different impacts given the underlying changing land: through land-use and cover change, vegetation browning or greening, soil wetting or drying, shifts in vegetation composition, structural, phenological and physiological changes? Land surfaces are continuously reshaped — directly through human management and indirectly through climate-driven ecosystem change — and each shift alters the boundary conditions at which the atmosphere sees the land, hence the intensity, duration, and footprint of the extremes that unfold above it. We are working with high-resolution coupled simulations with counterfactual land states to map where land change makes future extremes materially worse for the people living there — and, conversely, where it may mitigate them.

Biospheric effects on cloud formation under rising CO2_2

Changes in the atmosphere–biosphere coupling rooted in ecosystem and land responses also operate on longer timescales, resulting in shifts between latent and sensible fluxes and surface roughness properties. These shifts in the fluxes entering the atmosphere could lead to conditions that provide less or more the ingredients for cloud formation above land; with substantial feedbacks in radiative transfer and overall energy fluxes at the surface and top-of-the-atmosphere, critically controlling warming pattern over land. We aim to quantify the role of long-term vegetation changes and their response to rising CO2_2 in energy fluxes, surface roughness, atmospheric boundary-layer development, cloud formation, and ultimately the Earth energy budget, and provide quantitative assessments in multi-model approaches for the climatic changes in the decades ahead.

See projects →