Atmosphere

Ozone (O₃) pollution exhibits complex, scale-dependent variability across short-term (ST), seasonal (SN), and long-term (LT) temporal scales. However, most existing studies examine individual scales separately and focus on static within-scale relationships, leaving cross-scale response relationships insufficiently understood. This study focuses on eastern and central China (105° to 125° E, 20° to 42.5° N) and combines daily gridded fields of O₃, NO₂, HCHO, and isoprene from the ECMWF EAC4 reanalysis with ERA5 meteorological data for 2017–2025. Treating O3 and its major precursors as a coupled system, a dual-layer Gaussian low-pass filter was applied to standardized pollutant fields to separate ST, SN, and LT components. Joint Empirical Orthogonal Function (EOF) decomposition was then used to identify spatial modes and principal component time series characterizing the co-evolution of ozone and its precursors at each scale. A Physics-Informed Neural Network (PINN) was incorporated to explore potential physical drivers underlying each mode, while rolling multivariate Ordinary Least Squares (OLS) regression and a Convolutional Neural Network–Bidirectional Long Short-Term Memory (CNN–BiLSTM) model were combined to quantify cross-scale response intensity (β) and its temporal variability. Results indicate that (1) modal coupling relationships across temporal scales are continuously evolving. When primary modes (M1) serve as targets, responses generally weaken (→: slope = −0.053 yr⁻¹, →: slope = −2.82 yr⁻¹). When secondary modes (M2) serve as targets, responses strengthen (→: slope = +4.285 yr⁻¹, →: slope = +0.252 yr⁻¹). (2) Even if the long-term upward trend in concentrations is curbed, the continued strengthening of the negative modulation of (−0.053 yr⁻¹) may still cause short-term extreme event risks to increase rather than decrease. (3) Only a marginally significant upward trend is shown (Theil-Sen slope = +10.23 yr⁻¹, p = 0.048).

Dust storms frequently affect Saudi Arabia, disrupting transportation and agriculture, threatening public health, and incurring high economic costs. This study assessed whether elevated aerosol loading during dust storms was accompanied by consistent changes in precipitation. Data sources included Terra MODIS Deep Blue aerosol optical depth (AOD), GPM IMERG precipitation, METAR observations, and ERA5 mean sea-level pressure and wind data. AOD and precipitation were analyzed across Before, During, and After phases within the event footprints and nationwide, with the phase changes compared to earlier and later low-dust control periods. AOD consistently increased from the Before to During phase, exceeding the corresponding changes in the control periods. Precipitation responses varied across events and spatial scales, frequently occurring outside regions of enhanced aerosol loading, and pressure and wind patterns also differed. Therefore, elevated AOD was not accompanied by consistent precipitation enhancement or suppression. These findings suggest that precipitation responses to dust storms are event-specific and spatially varying.

Estimating Global Solar Radiation (GSR) is crucial for various engineering applications, including agriculture, residential planning, and energy systems. Reliable ground-based GSR measurements are limited in desert climates due to sensor degradation, sparse coverage and harsh environmental conditions. This study evaluates four ensemble-based machine learning (ML) models—Support Vector Machine (SVM), Random Forest (RF), Gradient Boosting (G-Boost), and Extreme Gradient Boosting (XG-Boost)—alongside two deep learning (DL) models, Multi-Layer Perceptron (MLP) and hybrid Long Short-Term Memory–MLP (LSTM-MLP)—for predicting daily GSR in Kuwait, one of the world’s hottest regions. The models incorporate 13 daily meteorological and time-related input parameters, including temperature (average, maximum, and minimum), potential sunshine duration, cloud cover, humidity, wind speed, dew point temperature, sea-level pressure, and time-based factors (day of month, day of year, month, and year). Data were collected from a ground weather station between March 2015 and December 2020. Performance was assessed using five metrics: Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE), Mean Squared Error (MSE), Root Mean Squared Error (RMSE), and coefficient of determination (R²). Results showed strong predictive capabilities across all models, with G-Boost and XG-Boost demonstrating the best predictive performance (R² = 0.996 and 0.994, respectively), while SVM achieved the highest held-out test R² (0.9729) among the ensemble-based ML models. An additional hourly validation using the XG-Boost model demonstrated excellent agreement between observed and predicted hourly solar radiation, confirming that the proposed framework accurately captures short-term variability and is applicable for both daily solar resource assessment and short-term photovoltaic forecasting under extreme desert climate conditions. Among DL models, MLP outperformed LSTM-MLP. The models maintained consistently high accuracy across seasonal variations, including peak summer periods characterized by extremely high ambient temperatures. Feature importance analysis identified potential sunshine duration as the most influential factor, followed by maximum temperature. These findings demonstrate that the proposed framework provides accurate and reliable prediction of global solar radiation across multiple temporal resolutions, making it suitable for both long-term solar resource assessment and short-term photovoltaic forecasting in extreme desert climates.