the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 14 Oct 2025
Deciphering Isoprene Variability Across Dozen of Chinese and Overseas Cities Using Deep Transfer Learning
Abstract. Isoprene, the globally most abundant volatile organic compound, significantly impacts air quality. Determining isoprene concentration variations and their drivers is a persistent challenge. Here, we developed a robust machine learning framework to simulate isoprene concentrations, without requiring localized emission inventories and explicit chemistry. Temperature, radiation, and surface pressure were the primary drivers of short-term isoprene variations across Chinese cities. On climatic timescales, urban greenspace expansion and climate warming drove isoprene increases by 341 pptv in Hong Kong during 1990–2023, but traffic emission reductions in London counteracted the isoprene rise that climate warming would have otherwise caused (-755 pptv vs. +31 pptv). Driven by rising temperatures and isoprene levels, ozone would increase by up to 1.7-fold by 2100 under the high-emission scenario. However, ambitious reduction in nitrogen oxides would alleviate this growth to 1.2-fold. The study has the potential to inform air quality management in a warming climate.
- Preprint
(1693 KB) - Metadata XML
-
Supplement
(945 KB) - BibTeX
- EndNote
Status: final response (author comments only)
-
RC1: 'Comment on egusphere-2025-4644', Anonymous Referee #1, 29 Oct 2025
-
AC1: 'Reply on RC1', Nan Wang, 08 Dec 2025
Reviewer #1
General comments:
This study employs machine learning techniques to investigate the patterns and driving factors underlying the fluctuations in isoprene levels — a crucial precursor to surface ozone — using extensive historical datasets. The analysis demonstrates good agreement with previous short-term modeling results, while also emphasizing that vegetation expansion and temperature increases linked to climate change are key drivers of long-term variability. Moreover, by extending the model projections to the year 2100 and integrating them with a detailed chemical box model, the researchers estimated future surface ozone levels. Their results indicate a significant rise in ozone concentrations if NOx emissions are not effectively controlled. Importantly, these conclusions were obtained using a data-driven approach that differs from conventional atmospheric chemical transport models, highlighting the robustness and novelty of the findings. I recommend acceptance for publication in ACP after minor revisions.
Response: We sincerely thank the reviewer for the positive and encouraging evaluation of our work. We appreciate the recognition of the study’s novelty, the robustness of the data-driven framework, and the relevance of our findings regarding long-term isoprene variability and its impactions for ozone (O3). We have carefully addressed all issues raised by the reviewer and revised the manuscript accordingly. We believe that these changes have further improved the clarity and quality of the paper. Below, we provide the point-by-point responses to each comment, with our replies highlighted in blue and the corresponding revisions marked in red.
Please kindly find the attached PDF-file for our responses.
-
AC1: 'Reply on RC1', Nan Wang, 08 Dec 2025
-
RC2: 'Comment on egusphere-2025-4644', Anonymous Referee #2, 27 Nov 2025
General comments:
The manuscript presents an explainable deep transfer learning framework (PINN-ResMLP) to predict urban isoprene concentrations and attribute their variability across Chinese and international cities. It further explores long-term drivers in Hong Kong and London (1990–2023) and projects future isoprene and ozone responses under CMIP6/SSP scenarios, including NOx-control sensitivity. The study fills an important gap: robust isoprene prediction without detailed local emission inventories or explicit chemistry, and interpretable attribution that links meteorology, greenspace, and traffic to observed and modeled trends.
The approach is timely and impactful for urban air quality management in a warming climate. The integration of physics-informed constraints with transfer learning is a notable strength, as is the explicit discuss ability (SHAP-based) of model predictions. The Hong Kong–London contrast is compelling and policy-relevant.
Specific comments:
1. Equations (5–6) use sign functions over partial derivatives. Please clarify how the gradients with respect to inputs are computed for monotonicity (e.g., via automatic differentiation), and whether local monotonicity is enforced pointwise or globally. Also specify α and β values and sensitivity.
2. ℒstruct is defined as sum of squared weights per layer (W_i² + b_i²). Are there any architectural constraints (e.g., skip connections in ResMLP, layer widths) chosen to improve stability? Include a small ablation (ResMLP vs. ResMLP+PINN vs. PINN alone) if possible.
3. For overseas sites, you fine-tune on 70% and validate on 30%. Clarify whether the split preserves temporal ordering (to reduce leakage) and whether performance is robust to different splits (report variance across splits).
4. The authors showed that WRF-Chem performed poorly in isoprene simulations. Provide configuration details (chemistry mechanism, emissions, resolution, boundary conditions) and whether the MEGAN parameterization and land-use were tuned to urban greenspace. This contextualizes the performance gap and its causes (e.g., grid dilution, canopy-scale processes).
5. State whether SHAP is computed on the fine-tuned model per site, the background dataset used, and whether interaction SHAP was explored (temperature × radiation) to reflect coupled sensitivities.
6. For future projections, please explicitly acknowledge that future greenspace, urban form, and anthropogenic emissions will also evolve.
7. There are two “the” in line 280.
Citation: https://doi.org/10.5194/egusphere-2025-4644-RC2 -
AC2: 'Reply on RC2', Nan Wang, 08 Dec 2025
Reviewer #2
General comments:
The manuscript presents an explainable deep transfer learning framework (PINN-ResMLP) to predict urban isoprene concentrations and attribute their variability across Chinese and international cities. It further explores long-term drivers in Hong Kong and London (1990–2023) and projects future isoprene and ozone responses under CMIP6/SSP scenarios, including NOx-control sensitivity. The study fills an important gap: robust isoprene prediction without detailed local emission inventories or explicit chemistry, and interpretable attribution that links meteorology, greenspace, and traffic to observed and modeled trends. The approach is timely and impactful for urban air quality management in a warming climate. The integration of physics-informed constraints with transfer learning is a notable strength, as is the explicit discuss ability (SHAP-based) of model predictions. The Hong Kong–London contrast is compelling and policy-relevant.
Response: We sincerely thank Reviewer #2 for the thorough and positive evaluation of our work. We greatly appreciate the recognition of the novelty and impact of our study, particularly the explainable deep transfer learning framework (PINN-ResMLP), its ability to predict isoprene concentrations without detailed local emission inventories, and the interpretability provided by SHAP analysis. We also thank the reviewer for emphasizing the relevance of the Hong Kong–London comparison and the importance of integrating physics-informed constraints with transfer learning. The constructive feedback and supportive comments are highly encouraging and have helped us further clarify and refine the manuscript. Below, we provide our point-by-point responses to each comment, with our replies highlighted in blue and the corresponding revisions marked in red.
Please kindly find the attached PDF-file for our detailed responses and revisions.
-
AC2: 'Reply on RC2', Nan Wang, 08 Dec 2025
Viewed
| HTML | XML | Total | Supplement | BibTeX | EndNote | |
|---|---|---|---|---|---|---|
| 239 | 37 | 32 | 308 | 49 | 15 | 18 |
- HTML: 239
- PDF: 37
- XML: 32
- Total: 308
- Supplement: 49
- BibTeX: 15
- EndNote: 18
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
Please see my commnets enclosed in supplement.