Earth system models are essential for understanding and projecting future climate change. However, these models operate at spatial resolutions of tens to hundreds of km, whereas many of the observations and
the processes they aim to represent, occur at much finer spatial scales. Bridging this gap is a fundamental challenge
for Earth system science.
For biosphere processes large-scale models simulate variables such as leaf area index, vegetation productivity, and carbon uptake on grid cells covering thousands of km². In contrast, observations of phenological events at individual plants, eddy-covariance flux measurements towers, or high-resolution remote sensing describe behaviour at much
finer scales. The understanding of these processes, derived from experiments and theory, operates at individual or plot scales.
Aggregation of nonlinear responses make the response to mean conditions differ from the mean of individual responses, while spatial averaging across heterogeneous processes smooths their contrasting sensitivities.
Addressing these scale mismatches requires advances across multiple fronts: sub-grid parameterizations in mechanistic models, observational networks designed to capture spatial heterogeneity, and methods to link fine-scale processes to emergent large-scale behaviour.

Machine learning is being combined with mechanistic Earth system models, offering possibilities and risks. Models trained on plot-level data may predict grid-cell averages without accounting for sub-grid variability, sampling biases, or nonlinear aggregation. Approaches that represent multiple scales, incorporate spatial structure, or learn mappings between individual-level processes and emergent large-scale variables show promise for improving how biosphere processes are represented.
The session aims to foster knowledge exchange between mechanistic modelers, machine learning practitioners, observational scientists, and experimentalist to develop more reliable, scale-consistent approaches for representing and
evaluating biosphere processes in Earth system models. Hence, the session invites contributions on topics including, but not limited to, the following:
• Trait-based upscaling beyond PFTs
• Sub-grid scale modelling, nonlinear spatial aggregation and emergent behaviour
• Upscaling in situ observations to grid scale
• Phenology across spatial scales
• Hybrid ML-physics modelling
• ML solutions to sampling bias
• Learning emergent grid-scale variables