Adaptive learning rate optimization in deep recurrent architectures for precision PM2.5 forecasting under climate variability.

Accurate forecasting of fine particulate matter (PM_{2.5}) is critical for safeguarding public health; however, conventional deep learning models frequently encounter convergence instability when exposed to highly volatile meteorological inputs. This study evaluates advanced learning rate scheduling strategies to improve long short-term memory (LSTM) and gated recurrent unit (GRU) performance using Atlanta's 2023-2025 air quality and meteorological data. The research compares polynomial, piecewise constant, and cosine decay schedules, revealing architecture-specific sensitivities: the analysis identifies that LSTM achieved optimal stability with piecewise constant decay (MAE = 0.9255 μg/m3), whereas GRU performed best under polynomial decay (MAE = 1.1092 μg/m3). Although Cosine Decay reduces peak errors in isolated instances, it demonstrates insufficient robustness against stochastic gradient noise. The optimized framework yields exceptional predictive fidelity, with the GRU and LSTM models attaining R2 values of 0.9495 and 0.9443, respectively. Beyond methodological advances, this study establishes a reliable computational baseline for site-specific temporal air quality monitoring, offering actionable insights into the interplay between anthropogenic emissions and climate-driven volatility. By supporting industrial emissions management and evidence-based policy, this study contributes to the broader goals of mitigating air pollution in a changing climate and advancing ecological co-governance.