Inferring Thermodynamic Histories From In Situ Ice Crystal Imagery via Conditional Diffusion Models

Ice crystal microphysical properties govern cirrus cloud radiative effects, yet their representation in climate models remains inadequate due to insufficient understanding of how thermodynamic histories control crystal morphology. Observational platforms provide only instantaneous snapshots of cloud properties, precluding direct measurement of the growth environments that determine final crystal structure. Here, we introduce a machine learning (ML) framework for probabilistically inferring WRF‐HYSPLIT derived atmospheric thermodynamic histories directly from ice crystal imagery. We use high‐resolution ice crystal observations from the 2002 NASA CRYSTAL‐FACE field campaign paired with HYSPLIT back trajectories driven by Weather Research and Forecasting simulations to construct a data set linking crystal morphology to approximately eight‐hour prior thermodynamic evolution. A conditional diffusion model, trained on variational autoencoder embeddings of crystal images and geometric attributes, generates ensemble distributions of temperature, pressure, and relative humidity trajectories consistent with observed morphology. Relative to an unconditional baseline, morphological conditioning reduces the continuous ranked probability score by 42% for temperature, 31% for pressure, and 7% for relative humidity. Results demonstrate that ice crystal morphology encodes recoverable constraints on these model‐derived thermodynamic histories, with temperature most strongly constrained, and uncertainty increasing systematically with look‐back time. This framework establishes a generalizable methodology for extracting process‐level understanding from snapshot observations and provides a pathway for improved observational constraints on ice microphysics parameterizations in climate models.