A Data-Driven Approach to Modelling a Global Distribution of Ice Nucleating Particles

Abstract. Ice nucleating particles (INPs) are rare aerosols essential for cloud ice formation in the mixed-phase temperature range between -38 °C and 0 °C. Due to measurement challenges and limitations in instrument capabilities, the availability of atmospheric observations of INPs remains scarce in time and space. Consequently, no observation-based global distribution of INPs exists so far. This study applies a machine learning (gradient boosting) algorithm to predict the INP concentration over the mixed-phase temperature range across the globe using aerosol mass concentration reanalysis and observed immersion-mode INPs. This proof-of-concept exercise demonstrates that even with limited measurements, the occurrence of INPs can be estimated, following the spatial pattern of key aerosol species. Point-based evaluation metrics, R² and log-based RMSE, reach 0.83 and 0.71 for the gradient boosting model, compared to 0.75 and 0.86 for the linear regression benchmark. Regional INP spectra with temperature extracted from the machine learning model agree within one order of magnitude with observations, except over Antarctica (mean bias factor of 150). INPs in continental (oceanic) regions are well predicted within a mean bias of 2.4 (overestimated within a mean bias of 4) by the machine learning model. The prediction is driven by the strong temperature dependence followed by mid-sized dust particles. This approach can resolve a long-standing source of INP prediction uncertainty in regional weather and climate models.