Reconstructing Late Palaeozoic Land-Ice Distributions: A Machine Learning Framework for Model-Data Comparison

Deep-time glacial intervals provide critical benchmarks for assessing Earth System Model (ESM) performance under past climate states. However, most paleo-simulations lack dynamic icesheets, leaving this key component poorly constrained. Here, we introduce a machine learning approach for reconstructing global glacial extent across the Phanerozoic, integrating paleoclimate simulations, paleo-topography, and a global stratigraphic database of glacial deposits. This framework generates spatially explicit, probabilistic reconstructions that enable quantitative comparison between geological archives and climate model ensembles, highlighting regions of agreement and mismatch.

The Late Palaeozoic Ice Age (LPIA), a >100-million-year glaciation variously attributed to declining atmospheric CO₂, palaeographic changes, and tectonic activity, provides an ideal case-study considered here. A persistent enigma concerning the LPIA is its hemispheric asymmetry, whereby preserved glacial deposits are abundant in the Southern Hemisphere but sparse in the Northern Hemisphere. Whether this bipolarity reflects genuine climate asymmetry or preservation bias remains unresolved. We address this by modelling the distribution of land-ice using environmental predictors such as temperature, precipitation, transpiration, and topography, derived from HadCM3L simulations that do not include dynamic icesheets. This analysis yields time-slice specific probabilistic reconstructions that can be directly compared with the preserved sedimentary record. We calibrate our framework against modern glaciers and LPIA glacial deposits, and subsequently applying it to other Phanerozoic ice ages, producing a consistent reference dataset for model-data comparison. While our approach does not replace fully coupled ice-climate simulations, it highlights some key discrepancies between models and geological evidence and allows climate asymmetry to be distinguished from preservation bias. By quantitatively bridging paleo-archives and climate models, our framework provides a new means of evaluating ESM performance across diverse climate states, strengthening constraints on ice-climate feedback relevant to future projections.