Revelation and Prediction of Atmospheric Pollutant Photolysis on Particulate Matter with Deterministic Learning Based on Infrared Spectra

Revealing nitrate photolysis on atmospheric particulate matter (PM), which generates secondary NO_x and contributes to haze, acid rain, and daytime HONO formation, offers crucial insights into regulating regional air quality. While in situ spectroscopic techniques are useful to reveal photolysis mechanisms on single-component PM surfaces, it remains challenging to infer the photolysis behavior of mixed PM in real environments and quantify the influencing factors. To mitigate this challenge, we herein utilized artificial intelligence (AI) to capture time-dependent spectral variations, enabling the prediction and analysis of nitrate photolysis in complex atmospheric systems. Specifically, the deterministic learning approach, a novel machine learning (ML) algorithm specialized in modeling and analyzing the nonlinear systems, was innovatively applied to extract critical dynamic characteristic curves from time-series infrared spectra. Taking NH₄NO₃ photolysis on mineral dust as an example, we developed a spectroscopy-based ML approach that relies on single-component PM data sets and demonstrates its broad applicability by enabling direct prediction of nitrate photolysis trends across diverse mixed-source PMs without experimental measurements. In addition to the predominant contribution of photoactive TiO₂ and Fe₃O₄ to nitrate photolysis by dynamic characteristic curve analysis, we also uncovered and experimentally validated a significant but overlooked inhibitory mechanism exerted by carbonates on these components. This work advances the integration of AI with atmospheric research, offering new perspectives for predicting and analyzing multisystem interactions in atmospheric environments while reducing the reliance on extensive experiments.