Scholarlink

<p>Aerosol particle number size distribution (PNSD) measurements are fundamental for quantifying aerosol dynamics, as they govern processes such as new particle formation (NPF), coagulation, and the contribution of particles to cloud condensation nuclei (CCN), thereby influencing aerosol–cloud–climate interactions. However, in the Himalayan region of India, such measurements are particularly challenging due to complex terrain, strong vertical mixing, variable boundary layer dynamics, and limited observational coverage, further compounded by instrumental constraints and uncertainties in inversion techniques that can lead to incomplete or unrealistic datasets. Therefore, modeling approaches are essential for complementing observations and improving size-resolved aerosol characterization. In this study, observations were carried out at the high-altitude Himalayan Cloud Observatory (HCO; 30.34°N, 78.40°E, 1706&#xa0;m AMSL) in Uttarakhand from 1 January to 31 December 2021. A feed-forward neural network (FFNN) model was developed using configurations with two hidden layers and 2–25 neurons, trained on 5-minute, hourly, and daily datasets to estimate particle number concentrations across different size bins using meteorological parameters. The optimal configuration (two layers with nine neurons) yielded moderate performance (R² = 0.22–0.45), with the highest accuracy at 27.4&#xa0;nm and the lowest at 48.7&#xa0;nm. The model demonstrates better performance for nucleation-mode particles, likely due to their relatively simple formation mechanisms associated with NPF, whereas Aitken- and accumulation-mode particles are influenced by multiple sources and complex growth processes. Diurnal analysis reveals a bimodal pattern in particle number concentrations, characterized by a short-lived morning peak and a more sustained evening peak. The model tends to overestimate concentrations from midnight to midday, while relatively improved performance is observed during the evening hours (16:00–21:00). This enhanced performance during the evening is likely associated with the prolonged increase in total particle concentrations, along with comparatively smoother variations in meteorological parameters, which together contribute to more stable and consistent model predictions. Seasonally, better model performance is observed during relatively stable winter and pre-monsoon periods, whereas increased fluctuations in meteorological parameters during the monsoon and post-monsoon periods introduce greater uncertainty. Stable wind speeds (1–2&#xa0;m s⁻¹) from southeast to southwest directions further support improved predictions. Lower performance for finer particles is attributed to their dynamic behavior and higher removal rates in the atmosphere. Overall, further improvement in model performance requires incorporating longer-term and higher-resolution observational datasets, which would enhance training robustness and better capture the complex variability of aerosol processes in the region.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Machine learning estimation of aerosol size distributions from meteorological data in the Western Himalayas

  • 期刊论文

  • 发表日期：

    2026年5月7日

查看全文
  • Alok Sagar Gautam,
  • Aman Deep Vishwkarma,
  • Karan Singh,
  • Sanjeev Kumar,
  • Yasti Panchbhaiya,
  • Sneha Gautam,
  • Chang-Hoi Ho,
  • Yong-Sang Choi

摘要

Aerosol particle number size distribution (PNSD) measurements are fundamental for quantifying aerosol dynamics, as they govern processes such as new particle formation (NPF), coagulation, and the contribution of particles to cloud condensation nuclei (CCN), thereby influencing aerosol–cloud–climate interactions. However, in the Himalayan region of India, such measurements are particularly challenging due to complex terrain, strong vertical mixing, variable boundary layer dynamics, and limited observational coverage, further compounded by instrumental constraints and uncertainties in inversion techniques that can lead to incomplete or unrealistic datasets. Therefore, modeling approaches are essential for complementing observations and improving size-resolved aerosol characterization. In this study, observations were carried out at the high-altitude Himalayan Cloud Observatory (HCO; 30.34°N, 78.40°E, 1706 m AMSL) in Uttarakhand from 1 January to 31 December 2021. A feed-forward neural network (FFNN) model was developed using configurations with two hidden layers and 2–25 neurons, trained on 5-minute, hourly, and daily datasets to estimate particle number concentrations across different size bins using meteorological parameters. The optimal configuration (two layers with nine neurons) yielded moderate performance (R² = 0.22–0.45), with the highest accuracy at 27.4 nm and the lowest at 48.7 nm. The model demonstrates better performance for nucleation-mode particles, likely due to their relatively simple formation mechanisms associated with NPF, whereas Aitken- and accumulation-mode particles are influenced by multiple sources and complex growth processes. Diurnal analysis reveals a bimodal pattern in particle number concentrations, characterized by a short-lived morning peak and a more sustained evening peak. The model tends to overestimate concentrations from midnight to midday, while relatively improved performance is observed during the evening hours (16:00–21:00). This enhanced performance during the evening is likely associated with the prolonged increase in total particle concentrations, along with comparatively smoother variations in meteorological parameters, which together contribute to more stable and consistent model predictions. Seasonally, better model performance is observed during relatively stable winter and pre-monsoon periods, whereas increased fluctuations in meteorological parameters during the monsoon and post-monsoon periods introduce greater uncertainty. Stable wind speeds (1–2 m s⁻¹) from southeast to southwest directions further support improved predictions. Lower performance for finer particles is attributed to their dynamic behavior and higher removal rates in the atmosphere. Overall, further improvement in model performance requires incorporating longer-term and higher-resolution observational datasets, which would enhance training robustness and better capture the complex variability of aerosol processes in the region.