Development and Diagnosis of a Temporal-Context U-Net Deep Learning Surrogate Model for CMAQ Particulate Nitrate over China

Abstract

Nitrate is a major inorganic component of fine particulate matter, yet simulating it via chemical transport models (CTMs) is computationally intensive. Deep learning surrogates have emerged as efficient tools for approximating CTMs. Here, we developed a Temporal-Context U-Net (TC-U-Net) model to emulate hourly surface-layer nitrate from the Community Multiscale Air Quality (CMAQ) model over China in July and December (2011–2020). Built on an attention-integrated architecture, TC-U-Net uses inputs from the target and preceding 2 h to capture rapid atmospheric changes and summarizes the preceding 96 h to represent multiday accumulation and dispersion. Incorporating temporal contexts enabled TC-U-Net to reproduce the spatiotemporal variations of CMAQ nitrate and improve high-concentration simulations, yielding higher accuracy than the common baseline models. Emission-reduction experiments captured the national-mean nitrate decreases in July but underestimated magnitudes, while agreement was limited for January NOx responses outside the model-development months. Diagnosis of TC-U-Net prediction residuals revealed statistical associations with other CMAQ outputs, including gaseous NH3 and particle-phase H+ mass concentration. Adding these variables improved prediction accuracy in exploratory tests, suggesting potential benefits from additional information on atmospheric chemical conditions. These findings motivate future development using explicit chemical constraints and diverse CMAQ emission-reduction training scenarios.