Integrating Simulations and Observations: A Foundation Model for Estimating the Aerosol Mixing State Index

Accurately predicting aerosol mixing states in real-world environments is crucial for understanding their impacts on climate change and human health. However, observational data inherently exhibit spatiotemporal gaps, and high costs and equipment requirements further exacerbate these limitations, particularly for in situ measurements. While particle-resolved models can simulate individual particle composition and size changes and serve as benchmarks, they face challenges in real-world applications due to a combination of factors. One of the major challenges is the limited availability of detailed input data (e.g., emission inventories) that accurately reflect actual environmental conditions. In this study, we frame the emulation of aerosol simulation as a general task and treat the estimation of real-world mixing states as a downstream task. We developed a foundation model pretrained on particle-resolved simulations and fine-tuned it using observational data from the field campaign. The fine-tuned model consistently outperformed baseline models, showing greater stability and robustness across various data sets. Permutation feature importance and sensitivity analyses revealed that aerosol species concentrations were the most critical factors for the foundation model. This approach, which involves pretraining on particle-resolved simulations and fine-tuning on limited observational data, offers a viable solution to challenges posed by limited observational data.