Reconstructing the VOC–Ozone Research Framework Through a Systematic Review of Observation and Modeling

Abstract

Tropospheric ozone (O₃), a secondary pollutant of mounting global concern, emerges from complex, nonlinear photochemical reactions involving nitrogen oxides (NO_x) and volatile organic compounds (VOCs) under dynamically evolving meteorological conditions. Accurately characterizing and effectively regulating O₃ formation necessitates not only precise and multi-dimensional precursor observations but also modeling frameworks that are structurally coherent, chemically interpretable, and sensitive to regime variability. Despite significant technological progress, current research remains markedly fragmented: observational platforms often operate in isolation with limited vertical and spatial interoperability, while modeling paradigms—ranging from mechanistic chemical transport models (CTMs) to data-driven machine learning approaches—frequently trade interpretability for predictive performance and struggle to capture regime transitions across heterogeneous environments. This review provides a dual-perspective synthesis of recent advances and enduring challenges in the VOC–O₃ research landscape. We first establish a typology of ground-based, airborne, and satellite-based VOC monitoring systems, evaluating their capabilities, limitations, and roles within a vertically structured sensing architecture. We then examine the evolution of O₃ modeling strategies, from empirical and semi-mechanistic models to hybrid frameworks that integrate physical knowledge with algorithmic flexibility. By diagnosing the structural decoupling between observation and inference, we identify key methodological bottlenecks and advocate for a system-level redesign of the VOC–O₃ research paradigm. Finally, we propose a forward-looking framework for next-generation atmospheric governance—one that fuses cross-platform sensing, regime-aware modeling, and policy-relevant diagnostics into an integrated, adaptive, and chemically robust decision-support system.

1. Introduction

Tropospheric ozone (O₃) remains one of the most complex and policy-relevant secondary air pollutants, emerging not from direct emission but from intricate nonlinear interactions among nitrogen oxides (NOx), volatile organic compounds (VOCs), and meteorological variables under the modulation of solar radiation. Unlike primary pollutants, the formation and spatiotemporal evolution of ozone are governed by region-specific photochemical regimes, sensitive to precursor ratios, local emission profiles, and atmospheric dynamics. Among these precursors, VOCs—together with NO_x—are key agents of photochemical reactivity, but VOCs pose particular challenges due to their chemical diversity and observational complexity. They exhibit high chemical diversity, heterogeneous reactivity, and substantial spatiotemporal variability, complicating both their measurement and mechanistic representation [1,2].

Despite decades of research, the dual challenge of observing and modeling VOC–O₃ interactions remains unresolved. While conventional air quality assessments often track indicators such as NO₂ and O₃, VOCs have remained comparatively under-observed in both spatial resolution and chemical speciation. Observationally, advances in satellite remote sensing, ground-based monitoring, and airborne sampling have expanded the capacity to capture VOC-related indicators such as formaldehyde (HCHO); yet, inconsistencies in resolution, coverage, and chemical specificity persist [3,4]. More critically, these platforms often operate in isolation, lacking structural integration across vertical and spatial scales, and failing to deliver the regime-resolving information necessary for predictive modeling. On the modeling side, traditional approaches—such as observation-based models (OBMs) and chemical transport models (CTMs)—have achieved impressive chemical completeness, yet suffer from coarse spatial resolution, latency in emission updates, and limited regime adaptability [5,6]. The emergence of data-driven algorithms, particularly machine learning models, has brought new momentum to the field, but many of these methods prioritize accuracy over interpretability and often lack sensitivity awareness or chemical transparency [3].

These gaps are not merely technical—they point to a deeper fragmentation in the VOC–O₃ research landscape. Observation strategies, modeling paradigms, and regulatory frameworks have often evolved in silos, each advancing rapidly but seldom converging. As a result, current systems remain diagnostic rather than responsive and descriptive rather than functional. The consequence is a growing disconnect between scientific capacity and governance needs, especially in transitional regions where O₃ formation is highly sensitive to both precursor dynamics and meteorological anomalies [7].

In light of these challenges, this review seeks to provide a system-oriented synthesis of recent progress and enduring limitations in VOC–O₃ research [8]. We adopt a dual-structured perspective: first, tracing the evolution and classification of observational platforms used to monitor VOCs and related precursors, including ground-based, airborne, and satellite-based systems; second, examining the methodological advances and structural bottlenecks in O₃ modeling, spanning from mechanistic CTMs to structure-informed AI algorithms. Rather than treating these domains in isolation, we focus on the coupling logic between observation and inference—how platform capabilities influence model fidelity and how modeling demands shape the future of VOC measurement—ideally forming a recursive sensing loop, where models not only consume observations but also help prioritize what, where, and when to observe [9,10].

Ultimately, this review aims to move beyond cataloging techniques toward envisioning a cohesive research architecture for VOC–O₃ interaction: one that aligns observational granularity with chemical regime inference, integrates physical mechanisms into learning systems, and bridges scientific understanding with real-time governance. In doing so, we seek to clarify not only where the field stands, but also what kind of system it must become.

2. Mapping Methods and Knowledge Evolution in VOC–O₃ Studies

2.1. Methodology

2.1.1. Data Sources

To construct a comprehensive and methodologically stratified overview of VOCs (volatile organic compounds) observation across multiple platforms, a systematic literature search was conducted in the Web of Science (WOS) Core Collection. The strategy was structured around seven platform categories: (1) ground-based monitoring stations, (2) mobile monitoring systems, (3) tower-based vertical observations, (4) industrial site monitoring, (5) chamber-based calibration experiments, (6) airborne platforms (manned aircraft and UAVs), and (7) satellite-based remote sensing.

To operationalize this search, we developed a series of Boolean query strings for each platform category, intersecting three key dimensions:

(1)

VOC species terms: e.g., “VOCs”, “NMHCs”, “BTEX”, “isoprene”, and “formaldehyde”;

(2)

Observation descriptors: e.g., “in-situ measurement”, “real-time analysis”, “profiling”, “remote sensing”, and “continuous sampling”;

(3)

Platform-specific identifiers: e.g., “air monitoring site”, “mobile laboratory,”, “Sentinel-5P”, and “drone-based VOC sampling”.

To avoid semantic overlap and ensure mutual exclusivity across platforms, we incorporated negative filters (e.g., NOT TS = (“satellite” OR “ground station”)) in each query to exclude non-targeted observation types. For example, satellite retrievals excluded UAVs and towers, and UAV queries excluded aircraft or surface stations. This approach enforced platform orthogonality and minimized redundant hits across categories.

Additionally, a targeted title-based inclusion list was appended to the tower-based search to recover high-impact benchmark studies that lacked full platform descriptors in their abstracts. This hybrid approach—combining structured topic-field queries with expert-curated recall—helped ensure comprehensive coverage while maintaining domain specificity.

The combined platform-resolved search, limited to SCI-indexed journal articles and covering all years through early 2025, returned 823 records. After removing 38 duplicates, a manual screening of 785 titles and abstracts was conducted. Records were excluded if they (i) lacked VOC measurement specificity, (ii) failed to indicate observation platform context, or (iii) focused solely on meteorological processes, vegetation studies, or climate modeling without pollutant metrics. This screening yielded a final dataset of 436 articles, which were retained for full-text analysis.

Although a formal PRISMA diagram is not included, the search logic, exclusion criteria, and semantic structure of the screening process are made transparent through both narrative documentation and the visual schematic provided in Figure 1.

2.1.2. Tools

This study employs a combined methodology of bibliometric analysis and knowledge mapping to systematically examine research on VOC observation across ground-based, aerial, and satellite platforms, as well as ozone modeling and inversion using artificial intelligence techniques. Bibliometric analysis enables a structured evaluation of publication trends, platform typologies, observational strategies, and modeling frameworks. The process involves literature filtering, metadata standardization, and the application of quantitative indicators to assess platform coverage, target pollutants, modeling methods, and citation impact, ensuring thematic alignment with atmospheric pollution monitoring and data-driven environmental assessment. All bibliographic records were retrieved from the Web of Science Core Collection. Visual analytics were performed using VOSviewer (v1.6.16) and CiteSpace (v6.3.R1). VOSviewer was applied to conduct co-occurrence and clustering analysis of research keywords, helping uncover key themes such as ground-based VOC detection, satellite remote sensing of precursors (e.g., HCHO and NO₂), and machine learning applications in ozone simulation. CiteSpace was used to identify citation bursts, track the temporal evolution of VOC-related observation technologies, and map collaboration networks across institutions and countries. Additionally, graph-theoretical metrics such as degree centrality, betweenness centrality, and Citation Half-Life were computed to quantify the structural dynamics of platform integration and computational modeling trends within the VOC–O₃ research landscape.

(1) Degree centrality

Degree centrality reflects how many direct connections a given node (e.g., a country or institution) holds within the collaboration network. A higher value signifies broader participation and more extensive direct cooperation. The formal definition is:

$$C_D\left(v\right)=\mathrm{d}\mathrm{e}\mathrm{g}\left(v\right)$$

(1)

where $\mathrm{d}\mathrm{e}\mathrm{g}\left(v\right)$ denotes the total number of edges directly linked to node $v$.

(2) Betweenness centrality

This metric assesses the intermediary importance of a node by measuring how frequently it lies on the shortest paths connecting other nodes. Nodes with high betweenness serve as bridges or brokers in the information flow. It is calculated as:

$$C_B\left(v\right)=\sum _{s\ne v\ne t}{\displaystyle \frac{{\sigma }_{st}\left(v\right)}{{\sigma }_{st}}}$$

(2)

where ${\sigma }_{st}$ is the total number of shortest paths from node s to node t, and ${\sigma }_{st}\left(v\right)$ is the count of those paths that pass through node $v$.

(3) Citation Half-Life (HL)

Citation Half-Life describes the temporal persistence of a node’s impact by measuring the median age of citations. For a series of citations C over years t₁, t₂,…, t_n, HL marks the point where half of the cumulative citations have been accrued. It is defined as:

$$HL=min\left\{t_k|\sum _{i=1}^k{C}_i\ge {\displaystyle \frac{1}{2}}\sum _{i=1}^n{C}_i\right\}$$

(3)

where C_i represents the number of citations in year t_i.

Together, these indicators provide a multi-dimensional perspective on collaboration dynamics—revealing not only who contributes most, but also how structural influence is established and sustained. Within the context of this study, they help map the evolving architecture of global research efforts in VOC monitoring and ozone prediction.

2.2. Knowledge Structure and Research Dynamics in the VOC–O₃ Domain

2.2.1. The Changing Landscape of VOCs and Ozone Research

Research on platform-based VOC observation and ozone modeling has progressed notably over the past three decades, following a distinct temporal trajectory shaped by technological advancement and scientific demand. Prior to 2005, the number of annual publications remained negligible, reflecting a nascent stage characterized by isolated site measurements and limited analytical resolution. Between 2005 and 2014, the field entered a slow-growth phase, with annual outputs fluctuating below ten and minimal citation impact—most efforts were confined to descriptive analyses of VOC concentrations or ozone levels near fixed monitoring stations, with few methodological innovations. A marked inflection point emerged after 2015, driven by the urgent demand for higher spatiotemporal resolution and the concurrent maturation of sensing technologies. Publication volume increased steadily, rising from 13 articles in 2016 to a peak of 51 in 2022, representing nearly a fourfold expansion within six years. This quantitative surge coincided with the increasing adoption of high-resolution ground-based flux towers, mobile monitoring systems, and atmospheric simulation chambers, alongside a parallel rise in artificial intelligence techniques for precursor source apportionment and ozone forecasting. Although citation frequencies peaked around 2020 and subsequently stabilized, they broadly mirrored the field’s growing academic visibility and practical relevance. Collectively, these trends mark a decisive transition from exploratory, single-node measurements to a multi-platform, computationally enriched paradigm that defines the contemporary VOC–O₃ research landscape (Figure 2).

(1) Ground-Based Foundations and the Rise of Urban VOC Characterization (2000–2008)

The early stage of VOC and ozone research was predominantly rooted in fixed-site observations within urban and industrial areas, where stationary gas chromatography systems and passive sampling techniques were the primary tools. Studies during this period concentrated on characterizing the chemical composition and temporal dynamics of VOC groups such as alkanes, alkenes, and aromatics, thereby establishing a foundational understanding of anthropogenic emissions and their photochemical reactivity. These efforts provided the basis for initial assessments of ozone formation potential and preliminary source attribution, albeit with limited spatial resolution and methodological depth (e.g., Volatile organic compounds in urban and industrial atmospheres in Rio de Janeiro, 2003).

(2) Vertical Expansion and the Emergence of Multiscale Modeling (2009–2015)

This phase witnessed a methodological evolution, as VOC observation expanded beyond surface-level monitoring to incorporate vertical and spatially distributed perspectives. Satellite-based instruments such as OMI and SCIAMACHY began delivering column-integrated retrievals of key ozone precursors such as formaldehyde (HCHO) and nitrogen dioxide (NO₂), complementing an increasingly diversified ground-based monitoring network that included rural and background stations. The integration of photochemical indicators—most notably the HCHO/NO₂ ratio—enabled the first widespread classification of ozone formation regimes. Concurrently, the uptake of chemical transport models (e.g., GEOS-Chem and CMAQ) initiated a shift toward mechanistic simulations, embedding observational data within predictive, process-resolved atmospheric frameworks (e.g., Satellite-derived formaldehyde as an indicator of VOC emissions, 2012).

(3) Cross-Platform Fusion and the Era of Precision Atmospheric Modeling (2016–2020)

A notable hallmark of this period was the fusion of diverse observational platforms into cohesive, multiscale monitoring systems. Ground-based superstations, airborne sensors, and satellite remote sensing products were integrated using data assimilation and inverse modeling approaches, enabling high-resolution quantification of precursor emissions and real-time characterization of photochemical regimes. The instrumentation landscape broadened considerably with the incorporation of technologies such as proton-transfer-reaction mass spectrometry (PTR-MS), differential optical absorption spectroscopy (DOAS), and flight-mounted LIDAR systems. This technological convergence facilitated more consistent cross-platform comparisons and laid the methodological foundation for high-precision, dynamic simulation of VOC–O₃ interactions under complex meteorological scenarios.

(4) Intelligent Sensing and Predictive Atmospheric Analytics (2021–present)

The most recent phase marks a paradigm shift toward computationally intelligent atmospheric diagnostics. Artificial intelligence methods—including convolutional neural networks (CNNs), long short-term memory (LSTM) models, and gradient-boosted ensemble trees—are increasingly employed to reconstruct ozone concentrations, infer spatiotemporal emission patterns, and predict photochemical behaviors under diverse environmental conditions. These data-driven approaches are often coupled with input streams from unmanned aerial vehicles (UAVs), mobile laboratories, and cost-effective IoT sensor arrays, forming adaptive frameworks capable of near-real-time monitoring and decision support. By fusing domain knowledge with algorithmic generalization, such systems have enabled robust, scalable solutions for VOC–O₃ mapping across geographies and seasons, advancing both scientific insight and policy relevance (e.g., Machine learning approach for ozone prediction based on VOC indicators, 2022; Spatiotemporal mapping of VOCs using drone-based sensor arrays, 2023).

From isolated manual sampling to intelligent, multi-platform sensing ecosystems, the evolution of VOCs and ozone research reflects an accelerating convergence of atmospheric science, sensor technology, and machine learning. As observational granularity deepens and computational tools mature, the field is increasingly positioned at the forefront of data-driven air quality governance and climate–chemistry interaction analysis.

2.2.2. National Research Profiles and Structural Influence

The global research landscape on VOC observation and ozone modeling exhibits pronounced asymmetries in both quantitative output and spatial distribution. As shown in Figure 3, China and the United States dominate scholarly production, with 100 and 61 publications, respectively, jointly accounting for over half of all identified literature. This prominence reflects sustained investment in atmospheric monitoring infrastructure and the early incorporation of modeling frameworks in national environmental agendas. Following these two leaders, countries such as Germany, South Korea, and the United Kingdom (20–25 publications) form a transregional core linking Europe and East Asia, often serving as hubs for algorithm development and platform integration.

In contrast, countries such as France, Belgium, and Italy (10–16 publications) maintain a moderate research presence, frequently participating in collaborative campaigns or contributing to precursor regime classification. Nations such as Brazil, Finland, Spain, and India (fewer than 10 publications) show more emergent or peripheral engagement, shaped by either nascent scientific infrastructure or limited integration into global research consortia. This geographical heterogeneity sets the stage for more nuanced evaluations of how countries participate in—and shape—the evolution of VOC–O₃ research beyond publication metrics alone.

Beyond publication volume, structural indicators offer deeper insights into the roles countries play within the global research network. As detailed in

Table 1. Quantitative metrics of national contributions and collaboration in VOC–O₃ research.

Country	Frequency of Occurrence	Degree (Collaboration)	Betweenness Centrality	Citation Half-Life (HL)
China	100	13	0.22	20.5
USA	61	19	0.41	21.5
Germany	25	15	0.21	16.5
South Korea	22	16	0.36	14.5
UK	20	8	0.11	12.5
France	16	15	0.25	13.5
Belgium	13	10	0.1	14.5
Italy	10	8	0.08	11.5
Brazil	9	4	0	13.5
Finland	6	1	0	21.5
Spain	5	7	0.01	10.5
India	5	8	0.11	15.5

, the United States, while second in publication count, leads in degree centrality (19) and betweenness centrality (0.41)—a clear indication of its dual role as both a prolific contributor and a key intermediary across collaborative pathways. In contrast, China, despite producing the highest number of publications, exhibits only moderate structural connectivity (degree: 13) and a lower bridging role (betweenness: 0.22), suggesting a more internally cohesive or regionally concentrated network of collaborations.

Germany, South Korea, and France occupy intermediate positions, combining solid output with structural connectivity. South Korea is particularly notable for its high betweenness centrality (0.36) relative to output, underscoring its role as a regional bridge between East Asian and Western research institutions. Meanwhile, countries such as Brazil, Finland, and Spain contribute to the literature but exhibit limited structural influence, with low centrality scores and marginal network positioning. These disparities between productivity and structural centrality illustrate the importance of international positioning—not merely output volume—in shaping influence and research integration within the VOC–O₃ domain.

Finally, citation-based indicators offer a third analytical dimension—illuminating the efficiency and scholarly resonance of national outputs. As visualized in Figure 4, countries such as the United States (52.43), Belgium (51.08), and Brazil (51.67) achieve exceptionally high average citations per article, indicating disproportionate academic impact despite moderate or low publication volumes. Similarly, Finland and Italy (37.5 and 31.2) reflect efficient research strategies with relatively high per-publication influence.

In contrast, high-output countries such as China (100 publications, 26.28 citations/article) exhibit moderate citation efficiency—likely reflecting a broad yet heterogeneous portfolio with variable international reach. South Korea and France also display subdued impact relative to output, suggesting potential limitations in journal visibility, language accessibility, or thematic generalizability. These asymmetries point to a partial decoupling between productivity and influence, wherein smaller, more strategically aligned research efforts may achieve greater global resonance.

Together, these multi-dimensional indicators—encompassing spatial output patterns, structural network positions, and citation efficiencies—offer a holistic characterization of national roles in VOC–O₃ research. They collectively highlight that advancing this field requires not only increasing output but also fostering global integration, thematic precision, and interdisciplinary innovation.

Despite these asymmetries, expanding the geographical and institutional base of VOC–O₃ research is both feasible and necessary. For emerging research communities outside the current dominant clusters, we offer the following suggestions. First, leverage open-access remote sensing platforms (e.g., TROPOMI and OMI) and reanalysis products (e.g., ERA5 and MERRA-2), which provide high-resolution data for ozone precursors and meteorology without requiring dedicated local infrastructure. Second, prioritize low-cost sensor networks and urban pilot studies, particularly in heat-vulnerable megacities where VOC–O₃ interactions are likely to be intensified by climate change. Third, engage with global collaborative programs—such as IGAC (International Global Atmospheric Chemistry) and WMO GAW (Global Atmosphere Watch)—to access technical resources, model toolkits, and cross-national co-authorship opportunities. Finally, explore regime-sensitive AI tools that can adapt to local conditions without requiring extensive historical data. These pathways not only democratize participation but also enhance the global representativeness and resilience of the ozone monitoring and modeling ecosystem.

2.2.3. Evolving Research Focus and Influential Contributions

Figure 5 highlights the top 12 keywords with the strongest citation bursts in VOC–O₃ studies between 1996 and 2024. Among them, “isoprene emissions” displays the most intense and sustained burst (2009–2020, strength = 6.71), reflecting its pivotal role in biogenic VOCs research and ozone formation. Meanwhile, “model” and “trend” stand out not only for their burst strength but also for the exceptionally long duration of their baseline presence—with “trend” exhibiting relevance from as early as 2000 and “model” maintaining consistent citations since 2003. This indicates their foundational status in structuring both observational and predictive frameworks.

In contrast, recent bursts such as “pollution”, “OMI”, “surface ozone”, and “satellite” (primarily 2020–2024) mark the rise of remote sensing diagnostics and policy-relevant air quality assessments [11,12]. Notably, keywords such as “VOCs”, “China”, and “variability” have emerged after 2022 and remain active, pointing to a thematic transition toward region-specific source complexity and uncertainty quantification in atmospheric modeling.

Collectively, these patterns reveal two layers of research focus: long-term conceptual scaffolding (e.g., trend and model) and recent methodological surges driven by technological advances and regulatory demand. The juxtaposition of persistent and emergent keywords underscores a maturing research field balancing theoretical grounding with applied innovation.

3. Hierarchical Platforms and the Recursive Architecture of VOCs Observability

3.1. Recursive Sensing and the Architecture of VOCs Observability

The challenge of observing VOCs lies less in detecting their presence than in resolving their spatiotemporal dynamics across chemically diverse, atmospherically coupled, and source-heterogeneous environments. This complexity demands not a single sensor but a vertically integrated system—one in which platforms function not as isolated instruments but as coordinated nodes within a structured observational logic. As shown in Figure 6, the taxonomy of VOCs monitoring—ground-based, airborne, and satellite-based—is not defined by hardware per se, but by how measurement intent maps onto atmospheric structure and process. Ground-level platforms, ranging from fixed stations to flux towers and simulation chambers, serve as the empirical foundation for source detection, reaction pathway validation, and near-field quantification. Airborne systems, especially UAVs and research aircraft, mediate between surface heterogeneity and orbital abstraction, enabling vertical profiling, flux estimation, and transport tracking. Satellite-based observations, in turn, offer synoptic-scale retrievals through instruments such as TROPOMI, OMI, and IASI, while specialized missions (e.g., TES and GIIRS) add vertical depth or spectral nuance, and legacy satellites (e.g., GOME and TOMS) anchor trend continuity.

What emerges is not a hierarchical cascade, but a recursive sensing architecture, where information flows bidirectionally—laboratory mechanisms inform model inversion; satellite retrievals are anchored in ground truth; and aircraft close the spatial and process gaps. The typology table and Sankey diagram (

Table 2. Typological framework of VOCs observation platforms and their functional roles.

Observation Tier	Platform Category	Specific Configuration	Functional Role and Scientific Relevance
Ground-based Platform	Fixed Station	Urban background station	Captures long-term urban-scale VOC trends, supporting pollution assessment and policy feedback [13].
		Traffic station	Enables high-temporal-resolution analysis of vehicle emissions and their transient behavior [14].
		Rural background station	Represents regional transport background and secondary VOCs formation baselines [15].
		Industrial site station	Precisely identifies industrial sources and exposure zones, supporting emission verification [16].
		Coastal station	Reveals VOC flux and atmospheric transport dynamics at land–sea boundaries [17].
	Mobile Platform	Vehicle-based platform	Builds city-scale mobile monitoring networks for high-resolution spatial heterogeneity mapping [18].
		Portable observation device	Provides flexible and rapid monitoring tools for hotspot detection and point-source screening [19].
	Point-Source Monitoring	Factory fence-line monitoring	Enables high-resolution monitoring of industrial boundary emissions for source strength estimation and accountability [20].
		Chimney/stack emission sensor	Real-time monitoring of stack outlets is essential for emission control and compliance auditing [21].
		Source-oriented mobile survey	Combines mobile scanning with near-field detection to locate potential sources efficiently [2].
	Flux/Gradient Towers	Flux tower	Measures area VOC fluxes to support source–sink estimates and land–atmosphere exchange analysis [22].
		Meteorological tower with gradient sampling	Constructs vertical profiles to study VOC transport and dilution in the boundary layer [23].
	Laboratory and Calibration Platform	Photochemical smog chamber	Simulates typical photochemical environments to study VOC oxidation chains and ozone formation [24].
		Atmospheric simulation chamber	Controls initial conditions and reaction parameters to analyze secondary pollutant formation [25].
		Instrument intercomparison laboratory	Compares monitoring methodologies to promote observational consistency and standardization [26].
		Flow tube reactor/reaction cell	Used for single-pathway reactions and transient intermediate detection in mechanism validation [27].
Airborne Platform	UAV Platform	UAV vertical profiling	Captures VOC vertical distribution within the boundary layer, addressing near-surface observational gaps [28].
		UAV remote sensing imaging	Provides high-resolution areal data for dispersion modeling and spatial pattern recognition [29].
		UAV source identification	Rapidly detects concentration anomalies near sources, supporting emission tracing and incident response [30].
	Manned Aircraft Platform	Aircraft trajectory observation	Collects high-frequency samples along flight paths to analyze long-range transport and evolution [31].
		Aircraft flux observation	Measures VOC flux across transects to estimate regional budgets and boundary-layer dynamics [32].
		Aircraft remote sensing imaging	Acquires high-altitude wide-area imagery for model validation and pollution mapping [31].
		Aircraft vertical profiling	Constructs VOC profiles from the surface to mid-troposphere for model constraint and mechanism analysis [33].
Satellite-based Platform	Satellite Platform	Core Operational Satellites (e.g., TROPOMI, OMI, and IASI)	Provide high-resolution global retrievals of VOC precursors (e.g., HCHO and NO2), forming the backbone for emission assessment, trend analysis, and atmospheric model assimilation [34,35,36].
		Specialized Instruments (e.g., TES, GIIRS, EMI, MOPITT, and OMPS)	Address advanced tasks such as vertical profiling, hyperspectral inversion, combustion source tracking, and supplementary column observations of trace gases [34,37,38,39].
		Legacy Observation Missions (e.g., GOME, GOME-2, SCIAMACHY, and TOMS)	Established foundational VOC and ozone remote sensing datasets, enabling cross-generational continuity and long-term atmospheric change analysis [40,41].

) together reflect this shift: platform categories no longer denote instrument classes but functional roles—initiators of detection, intermediaries of integration, and synthesizers of inference. The modern VOCs observation system is thus not a sum of devices, but a designed coherence—a methodological chain through which chemical variability, atmospheric reactivity, and policy-relevant accountability become measurable. This foundational layer of the sensing system, though often overlooked, anchors the vertical architecture of VOCs observability.

3.2. Ground-Based Platforms for VOCs Monitoring

Ground-based observation remains the empirical cornerstone of VOCs monitoring—not because it is closest to the source, but because it is closest to the science. It is here that measurements acquire chemical resolution, where data speak the language of reactivity, speciation, and mechanistic constraint. Unlike satellite retrievals, which must infer, or airborne systems, which often interpolate, ground-based platforms observe with intent—anchoring, validating retrievals, and illuminating processes [42,43].

Before delving into functional typologies and case-based analysis, it is instructive to consider the statistical landscape of ground-based VOCs monitoring platforms as reflected in the reviewed literature. As shown in Figure 7, five major categories dominate the domain: laboratory-based systems (45 studies), mobile ground platforms (42), fixed ground stations (38), tower-based flux and profiling structures (18), and point-source monitoring units deployed at factory fencelines or industrial boundaries (7). While differing in configuration and application, these platforms collectively embody the empirical foundation upon which VOCs observational science is built.

Over the past two decades, the typology of ground VOCs platforms has evolved from static monitoring nodes to a diversified, functionally adaptive architecture. This architecture is not defined by geography alone—urban, rural, or coastal—but by what it enables: trend interpretation, source attribution, regulatory validation, and mechanistic experimentation [44,45]. What follows is not a catalogue, but a storyline of how these platforms extend the reach of atmospheric insight—from fixed street corners to chemical reaction chambers.

We begin with the fixed monitoring station, long considered the spine of regulatory and trend-based VOCs networks. Figure 8a presents a modular station in peri-urban France, equipped with continuous GC-FID speciation and designed for unobstructed boundary-layer sampling [46]. Beyond its instrumentation, its placement matters—it stands at the interface of urban dynamics and regional background, enabling time-resolved analysis of BTEX and other ozone precursors [43]. Such stations have become diagnostic hubs in cities such as Hong Kong and Seoul, linking local emissions with ozone exceedances and enabling model–inventory reconciliation [47,48].

Yet cities breathe, and their emissions move. Hence the rise of mobile platforms—systems that chase sources rather than await them. Figure 8b depicts a ruggedized, vehicle-adapted mass spectrometer deployed for dynamic source tracking [49]. Its real-time sensitivity, compactness, and spatial mobility enable responsive monitoring in industrial corridors, port zones, and even emergency contexts. Mobile campaigns in Beijing and Mexico City have revealed discrepancies in inventories, identified fugitive industrial plumes, and trained machine learning algorithms to recognize source fingerprints in transit [50].

Still, some truths lie not across streets but along gradients. Figure 8c shows a 365-m flux and profiling tower near Shenzhen, which captured altitude-dependent NO_x and O₃ variations during the COVID-19 lockdown [23]. These towers act as structural interpolators—constructing vertical VOCs profiles, resolving inversion layers, and feeding satellite and CTM validation. Whether over urban basins or biogenic forests, such platforms reveal how VOC fluxes evolve with meteorology, mixing, and time of day [51].

Some sources, however, do not advertise their presence. Fenceline platforms, such as the one shown in Figure 8d at a synthetic rubber facility, are designed for exactly that—emissions that evade regulation but impact nearby populations. With dual passive/active sampling nodes and spatially resolved layouts, these systems uncover emission heterogeneity, detect carcinogenic hotspots, and guide mitigation strategies [21]. Their integration into community health monitoring and EPA methods marks a shift from passive surveillance to proactive accountability [52].

Lastly, VOCs science does not stop at detection—it interrogates. Figure 8e illustrates the SAPHIR-PLUS simulation facility in Germany, a hybrid platform that simulates both biogenic VOC emissions and their oxidative fate under controlled conditions [53]. Laboratory platforms like these are where abstract pathways become measurable, where PTR-TOFMS methods are cross-compared, and where model structures are either upheld or undone [54]. They are the epistemic laboratories of atmospheric chemistry—supporting campaign design, instrument calibration, and theoretical constraint.

Figure 8. Typology of ground-based VOCs monitoring platforms: from fixed nodes to mechanistic simulators. Visualized configurations of five core ground-based platform types: (a) fixed station [55], (b) mobile monitoring unit [49], (c) stratified tower [23], (d) fenceline source tracker [2], and (e) laboratory chamber system [53].

Figure 8. Typology of ground-based VOCs monitoring platforms: from fixed nodes to mechanistic simulators. Visualized configurations of five core ground-based platform types: (a) fixed station [55], (b) mobile monitoring unit [49], (c) stratified tower [23], (d) fenceline source tracker [2], and (e) laboratory chamber system [53].

Far from constituting the lowest tier of an observational hierarchy, ground-based VOCs platforms act as both the empirical inception and the interpretive fulcrum of atmospheric sensing. They do not merely support other systems—they contextualize them. Satellites may extend scope, and UAVs may resolve gradients, but without the chemical granularity, experimental controllability, and regulatory proximity offered by ground platforms, these upper-tier systems risk floating untethered above the complexity they aim to resolve.

This is the epistemic leverage of ground-based observation: it renders atmospheric chemistry not just observable, but legible. It mediates between molecule and model, between emission and attribution. In the era of integrated, recursive observation chains, the ground is no longer a reference—it is a reality anchor. In addition, in VOCs research, it remains where measurement becomes meaning.

Nonetheless, the interpretive power of ground-based platforms is not without its geographic caveats. While chemically resolute and temporally rich, these platforms are disproportionately concentrated in urban and peri-urban locales, resulting in a spatial sampling bias that may skew broader assessments of ozone formation regimes. In particular, the underrepresentation of rural or transitional zones—where biogenic VOCs dominate and NO_x levels fall below regulatory thresholds—poses challenges for extrapolating urban-derived insights to regional scales. This spatial asymmetry can distort regime classification, leading to overemphasis on VOC-limited dynamics while underdetecting NO_x-limited transitions in downwind or forested environments. Bridging this observational gap requires not only the expansion of fixed networks into under-monitored areas but also the strategic deployment of mobile units and UAV-based samplers to trace rural plume evolution and assess background–source interactions.

3.3. Airborne Platforms and Multi-Altitude VOC Detection

Airborne platforms occupy a strategic tier in VOC monitoring, bridging the vertical and spatial resolution gaps between ground-based sites and satellite instruments. They enable researchers to track plumes in real-time, map altitudinal profiles of reactive species, and investigate atmospheric chemistry in regions where fixed stations are sparse or inapplicable—such as above forest canopies, over oceans, or during dynamic wildfire events. These platforms expand not only the physical reach of observation but also the temporal and chemical granularity at which VOC processes can be resolved.

Two main airborne systems—manned aircraft and unmanned aerial vehicles (UAVs)—form a complementary observational hierarchy. Manned aircraft offer unparalleled endurance, payload capacity, and vertical profiling, while UAVs provide agile, near-surface access at finer spatial resolutions. Rather than functioning in opposition, they occupy distinct ecological and technological niches within the VOC sensing landscape.

To clarify the operational contrasts and synergistic roles of these two platform types,

Table 3. Manned vs. UAV VOC observation platforms: comparative capabilities and applications.

Dimension	Manned Aircraft	Unmanned Aerial Vehicle (UAV)
Operational Altitude	Up to 15 km (troposphere)	<500 m (near-surface, canopy level)
Typical Duration	4–10 h (long-range campaigns)	0.5–1 h (battery-limited)
Payload Capacity	High (full-size GC–MS, PTR–MS, offline trays)	Limited (mini samplers, sensors)
Spatial Resolution	Moderate–high (regional plumes, stacked paths)	Very high (local gradients, point sources)
Application Focus	Biomass plumes, marine, regional transport	Industrial tracing, canopy fluxes, emergency response
Strengths	Vertical profiling, long-distance coverage	Low-altitude flexibility, near-field resolution

summarizes their respective capabilities, applications, and trade-offs across key dimensions such as altitude range, instrumentation payload, spatial resolution, and observational focus. This comparison lays the foundation for understanding how each system is deployed in real-world campaigns—and why both are indispensable to the atmospheric VOCs toolkit.

These complementary features are not just theoretical—they translate into mission-specific deployments across a wide array of environments. The following four case studies illustrate how manned and unmanned systems have been used to intercept biomass plumes, resolve industrial emissions, characterize forest canopy processes, and quantify offshore facility leakage. Each reflects the platform’s situational strengths and application logic within broader observational strategies [56]. These real-world deployments are visualized in Figure 9, which showcases representative configurations for high-altitude profiling, offshore attribution, industrial plume tracing, and forest-canopy gradient mapping across manned and unmanned airborne systems.

Nowhere is this interplay more evident than in tropical biomass burning regions, where reactive carbon is lofted, transformed, and transported across vast atmospheric layers. During the ACRIDICON-CHUVA campaign over the Amazon, manned aircraft performed stacked flight profiles above and around Manaus, tracing formaldehyde, glyoxal, and methylglyoxal across cloud layers and into the upper troposphere [31]. These measurements revealed not only the vertical complexity of OVOC emissions but also their systematic underrepresentation in inventories—a mismatch that only altitude-resolved, chemically detailed surveys could expose [59].

Yet not all emission sources rise to meet such aircraft. In coastal and offshore settings, for instance, plume dispersion occurs within shallow marine boundary layers, often below satellite detection limits and beyond the reach of land-based monitors. To address this, aircraft such as the Twin Otter were reconfigured for low-altitude transects across North Sea oil and gas platforms. Using isotopic tracers and mass-balance fluxes, these missions disentangled venting, flaring, and fugitive sources—resolving not just the magnitude of emissions but also their regulatory origin [56]. Here, manned aircraft did not survey passively; they isolated and attributed, bridging science with enforcement.

Where even small aircraft cannot go, UAVs extend the observational frontier. In Taiwan’s Linhai Industrial Complex, drone-mounted needle trap samplers were deployed along pre-modeled dispersion paths to trace VOC hotspots around petrochemical and wastewater facilities. Unlike broader aerial transects, these missions mapped spatial variability in real-time—detecting localized spikes in TEX compounds with a resolution unattainable by fixed stations. The drone was not an accessory; it was the only platform capable of operating at that proximity, in that context, with that degree of specificity [28].

Elsewhere, in subtropical forests of southern China, UAVs were dispatched to investigate vertical gradients of biogenic VOCs—particularly isoprene and α-pinene—across canopy layers. Measurements at 25 and 100 m revealed unexpected reversals in concentration patterns, prompting the use of transport modeling to explain eddy diffusion and slope-induced flow effects. The result was more than a novel finding; it was a correction to long-held assumptions in biogenic emission modeling, one achievable only by operating in the vertical interspace between forest towers and satellites [58].

In this way, airborne platforms do more than fill gaps in existing systems—they create new observational grammars. Manned aircraft translate regional chemistry into altitude-resolved narratives; UAVs resolve spatial heterogeneity into mechanistic constraints. Their combined choreography allows VOCs science to ask different questions, in different places, with different scales of accountability.

In VOCs research, air is not an inert transit zone—it is a chemically active theater. In addition, airborne platforms are how we learn to navigate it.

While ground- and airborne platforms provide high-resolution and chemically explicit insights at limited spatial scales, satellite platforms extend this observational framework to the planetary scale, enabling global diagnostics, emission tracking, and long-term trend analysis.

3.4. Satellite-Based VOCs Monitoring and Ozone Sensitivity Diagnostics

In the vertically structured VOCs observation system, satellite platforms occupy a singular and irreplaceable position—not by virtue of their chemical sensitivity, but through their unique capacity for spatial omnipresence and temporal continuity. Unlike ground-based networks that anchor chemical precision yet remain spatially constrained, or airborne platforms that offer vertical profiling but suffer from campaign intermittency, satellites provide uninterrupted and geographically extensive coverage. This fundamental advantage becomes increasingly critical as VOC–ozone interactions evolve into a central concern of atmospheric chemistry, air quality management, and climate-linked policy. While satellite retrievals often lack vertical resolution and species-level granularity, recent advances in spatial fidelity, multispectral design, and algorithmic inference have transformed these platforms from passive monitors into dynamic informers—capable of diagnosing chemical regimes, reconstructing emission sources, and tracing complex transport phenomena across scales unreachable by any other means.

One of the most significant contributions of satellite-based observation lies in its diagnostic capacity. Jin et al. (2017) utilized long-term OMI retrievals to develop and evaluate the formaldehyde-to-NO₂ ratio (FNR) as a satellite-derived proxy for surface ozone formation regimes [60]. Their study revealed that, with proper correction for vertical representativeness, satellite-derived FNR can closely mirror surface-level reactivity transitions and distinguish between VOC-limited and NO_x-limited conditions across urban and regional domains [61]. Unlike isolated ground stations that provide point-based measurements, satellite data allowed for continental-scale regime mapping and seasonal trend analysis, uncovering early transitions toward NO_x-limited chemistry in densely populated areas such as New York, London, and Seoul [62]. These findings underscore a unique attribute of satellites—not in resolving chemical intricacies, but in spatializing chemical dynamics into policy-relevant landscapes.

A second frontier emerges from the use of satellite retrievals to invert and constrain global VOC emissions. Bauwens et al. (2016) demonstrated how nine years of OMI-based HCHO column observations, when combined with adjoint modeling via IMAGESv2, could reconstruct global NMVOC fluxes with unprecedented granularity [63]. Their inversion revealed substantial overestimations in standard bottom-up inventories, especially across equatorial forests and biomass burning zones. By adjusting pixel-level emissions to match observed formaldehyde patterns, they not only corrected static inventory biases but also captured interannual variability linked to ENSO cycles and land-use change [64]. This application highlights a distinctive strength of satellite platforms: not merely recording environmental states, but recalibrating global understanding where ground truth is inaccessible or systematically flawed [65].

The diagnostic and reconstructive capabilities of satellite platforms converge most clearly in national-scale policy evaluation. Ren et al. (2022) employed decade-long datasets from OMI and TROPOMI to assess the evolving ozone–precursor sensitivity across China during a period of aggressive NO_x reduction [66]. While surface monitoring networks recorded declining NO₂ levels, satellite-derived HCHO/NO₂ ratios revealed the persistence and even expansion of VOC-limited regimes in multiple urban clusters. These results challenged prevailing assumptions that NO_x controls alone would suffice for ozone mitigation and demonstrated that, in the absence of coordinated VOC abatement, regime shifts may amplify rather than reduce surface ozone concentrations. Satellites, by enabling consistent nationwide diagnostics, revealed a deeper asymmetry between emission policy and atmospheric response—an insight inaccessible via local measurements alone.

Beyond structured evaluation, satellite platforms also excel under conditions of environmental disruption. During the 2018 Canadian wildfires, Alvarado et al. (2020) leveraged Sentinel-5P TROPOMI data to monitor the long-range transport of glyoxal and formaldehyde, revealing coherent plumes that persisted over distances exceeding 1500 km [67]. Contrary to standard assumptions of short atmospheric lifetimes for oxygenated VOCs, dispersion modeling required extended decay constants to reproduce the spatial footprint observed by satellites. More importantly, the data suggested continuous in-plume production from precursor species—highlighting atmospheric recycling processes invisible to ground monitors and infeasible for aircraft to track in full spatial or temporal extent. In such extreme scenarios, satellites emerge not as complementary tools but as the sole observational mode capable of capturing both the chemistry and geography of large-scale perturbations [68].

The cumulative evolution of satellite observation platforms is further evidenced in their deployment across VOC-focused literature. As shown in Figure 10, OMI remains the most widely utilized instrument, reflecting its long operational timeline and robust HCHO and NO₂ retrievals. However, the rise of TROPOMI in recent years signals a methodological shift—toward higher spatial resolution, shorter revisit intervals, and enhanced multi-gas coverage [69]. Earlier missions such as SCIAMACHY and GOME laid foundational algorithms, while emerging platforms such as EMI and GIIRS reflect the increasing demand for finer spectral discrimination and greater regional targeting [70]. This instrumental evolution parallels the scientific shift from data collection to interpretation—satellites are no longer used solely to confirm chemical presence but to infer process, identify deviation, and quantify transformation.

Together, these advances recast satellite platforms not as replacements for terrestrial or aerial systems, but as structural complements that operate in the spaces between and beyond. They are not optimized for resolution or specificity, but for continuity and totality—for the ability to observe what no other system can see and to synthesize what no other configuration can measure. Their known limitations—cloud interference, vertical averaging, and species selectivity—are steadily mitigated by algorithmic refinement and synergistic fusion with ground and aircraft data. In the recursive architecture of VOC sensing, satellites serve as the closing and opening node: they provide the final spatial context within which all local observations must be interpreted and the first global frame from which atmospheric dynamics must be inferred.

As VOCs research transitions toward integrated environmental governance—spanning air quality, carbon policy, health exposure, and climate feedback—satellite-based observations will remain indispensable. Not only because they see the most, but because they allow others to see more clearly. This integrative role is further exemplified by recent representative studies (Figure 11), which demonstrate how satellite data enable regime diagnostics, emission inversion, sensitivity tracking, and cross-continental VOC transport modeling.

4. Technological Pathways for VOC Observation and Ozone Modeling

This section provides an integrated overview of the main technological approaches for diagnosing VOC–O₃ interactions and modeling ozone formation. We outline a progression from semi-empirical sensitivity analysis to traditional observation-based and chemical transport models, followed by remote sensing inversions and emerging artificial intelligence frameworks. Each approach is presented with its theoretical foundation, methodological advances, and representative applications, highlighting how these models collectively inform both mechanistic understanding and regulatory strategies.

4.1. Nonlinear Mechanisms and Semi-Empirical Modeling of O₃ Formation Sensitivity

Ozone (O₃) formation is governed by highly nonlinear chemical reactions involving nitrogen oxides (NOx) and volatile organic compounds (VOCs). The sensitivity of O₃ production to these precursors serves as a critical basis for air quality management strategies. Early empirical and modeling studies across major Chinese urban clusters such as Chengdu, the Pearl River Delta (PRD), and the Yangtze River Delta (YRD) have consistently revealed spatiotemporal shifts in O₃ formation regimes. Notably, a diurnal transition between NO_x-limited and VOC-limited conditions has been observed, especially under high-radiation or stagnation conditions [71]. This temporal dynamism has been confirmed in smog chamber experiments and box model simulations conducted in Beijing, Zibo, and Hefei, where radical ratios such as HO₂/OH have emerged as more effective indicators than conventional photochemical markers such as P_H₂O₂/PHNO₃ [72].

$${\displaystyle \frac{{\mathrm{H}\mathrm{O}}_2}{\mathrm{O}\mathrm{H}}}\mathrm{vs}.{\displaystyle \frac{P_{{\mathrm{H}}_2{\mathrm{O}}_2}}{P_{{\mathrm{H}\mathrm{N}\mathrm{O}}_3}}}$$

(4)

Beyond individual events, multi-year and multi-regional analyses—spanning the U.S., Hong Kong, and Shanghai—have illuminated a systematic shift in O₃ sensitivity regimes as a response to prolonged NO_x emission reductions. This evolution from VOC-limited to transition or NO_x-limited regimes underlines the adaptive nature of the atmospheric chemical system. Landmark studies such as MOOSE and DISCOVER-AQ, as well as the “lockdown” scenario in Shanghai, further emphasize the dynamic feedback loop between policy interventions and chemical sensitivity landscapes. These insights have prompted the emergence of flexible control strategies centered on real-time precursor sensitivity rather than static reduction ratios.

To address the growing complexity in regime identification, researchers have advanced a range of modeling tools. Response Surface Models (RSMs) and enhanced polynomial variants such as Epf-RSM allow for efficient prediction of O₃ concentrations in high-dimensional nonlinear response fields. Such models help avoid overfitting near transition boundaries and are now widely applied in regions such as the PRD [73,74,75].

$$y_{{\mathrm{O}}_3}={\beta }_1{x}_1+{\beta }_2{x}_2+\cdots +{\gamma }_1{x}_1^2+{\gamma }_2{x}_2^2+\cdots +{\delta }_{12}{x}_1{x}_2+{\delta }_{13}{x}_1{x}_3+\cdots$$

(5)

Modified CMAQ-DDM and OSAT modules have also introduced dynamic threshold-based indicators, improving interpretability in mixed or transitional regimes. In parallel, observable proxies such as HCHO/NO₂ and H₂O₂/NO₂—now routinely retrievable via OMI, SCIAMACHY, and Sentinel-5P—have become standard metrics for regime classification on regional to continental scales.

$${\displaystyle \frac{\mathrm{H}\mathrm{C}\mathrm{H}\mathrm{O}}{{\mathrm{N}\mathrm{O}}_2}}\mathrm{or}{\displaystyle \frac{{\mathrm{H}}_2{\mathrm{O}}_2}{{\mathrm{N}\mathrm{O}}_2}}$$

(6)

More recently, hybrid methods integrating RSMs with differential modeling (RSM-DM) or adopting machine learning tools such as random forests with VOC reactivity groupings have begun to bridge the gap between model accuracy and operational utility. These techniques enable both interpretability and speed in sensitivity diagnosis and have demonstrated promising performance in multi-city deployments.

Geographically, the sensitivity regime of O₃ formation varies significantly across China’s urban and peri-urban areas. VOC-limited conditions dominate high-density eastern metropolises, whereas NOx-limited regimes prevail in central-western cities and coastal forests—an observation supported by both ground-based and satellite-derived NOx/VOC and HCHO/NO₂ ratios. These spatial disparities are compounded by influences from biogenic VOCs, biomass burning, and meteorological transport, producing mosaic-like sensitivity landscapes even within a single metropolitan region.

$$\mathrm{V}\mathrm{O}\mathrm{C}\text{-}\mathrm{limited}:{\displaystyle \frac{\mathrm{H}\mathrm{C}\mathrm{H}\mathrm{O}}{{\mathrm{N}\mathrm{O}}_2}}<2$$

(7)

$${\mathrm{N}\mathrm{O}}_x\text{-}\mathrm{limited}:{\displaystyle \frac{\mathrm{H}\mathrm{C}\mathrm{H}\mathrm{O}}{{\mathrm{N}\mathrm{O}}_2}}>4$$

(8)

Intra-city comparisons, such as Pudong vs. Jinshan in Shanghai or EP1 vs. EP2 in Beijing, further validate the need for fine-resolution classification systems capable of capturing microscale transitions [76]. These findings support the development of hierarchical diagnostic frameworks that can accommodate city-, district-, and neighborhood-level heterogeneity.

From a policy perspective, scenario-based simulations have proven instrumental in evaluating the implications of different NOx/VOC reduction ratios. Studies confirm that in VOC-limited or transitional regimes, NOx-only reduction can paradoxically exacerbate O₃ pollution. Instead, targeting high-reactivity VOC species—such as alkenes and aromatics—yields more consistent benefits. In contrast, NO_x-limited settings respond favorably to NOx mitigation alone. Industrialized urban regions such as Zibo and the Chengdu-Chongqing corridor demand tailored phase-wise strategies, with some studies recommending VOC:NOx control ratios of 3:1 or 2.4:1 to avoid sensitivity inversion.

High-temperature scenarios introduce another layer of complexity, often triggering a switch in regime type that undermines previously effective strategies. For instance, in Shanghai, heatwaves have been shown to shift sensitivity from VOC- to NOx-limited states. Emission sector attribution analyses further reveal that mobile sources and solvent-based industries dominate VOC emissions, although the latter alone cannot mitigate O₃ peaks due to their limited spatial correlation with observed maxima. Recent modeling efforts also stress the need to correct for photochemically depleted VOCs, which, if ignored, may lead to systemic underestimation of NOx sensitivity and flawed policy guidance.

VOC_eff = VOC_obs + Δ_photochem

(9)

Collectively, these insights mark a shift in paradigm: from rule-based interventions to chemically informed, region-specific, and temporally adaptive strategies for O₃ management.

4.2. Traditional Modeling Paradigms: From Observation-Based Inference to Chemical Transport Simulation

Traditional modeling frameworks have long underpinned the scientific understanding of ozone (O₃) formation, serving as essential tools for deciphering precursor contributions and informing air quality control strategies. Among these, Observation-Based Models (OBMs) and Chemical Transport Models (CTMs) represent two complementary approaches: the former emphasizes in situ photochemical diagnosis, while the latter integrates emissions, chemistry, and meteorology to simulate regional O₃ dynamics. Over the past two decades, these models have advanced in structural integration, regional adaptability, and nonlinear sensitivity diagnosis [77,78].

CTMs provide a comprehensive, mechanistically grounded platform to simulate ozone formation and assess its response to precursors across spatial and temporal scales [79]. They typically couple chemical reaction mechanisms (e.g., SAPRC, CB05, and MCM), gridded emissions inventories (e.g., MEIC, NEI, and EDGAR), and meteorological drivers (e.g., WRF) to simulate pollutant production, transport, and deposition. Notable CTMs include GEOS-Chem, WRF-Chem, CMAQ, MATCH-MPIC, and TOMCAT. For instance, GEOS-Chem sensitivity analyses conducted during the ARCTAS campaign identified the aerosol uptake of HO₂ radicals as a dominant uncertainty in Arctic ozone formation, with marked seasonal divergence in precursor response patterns [80]. Over the Tibetan Plateau, CTM simulations revealed systematic underestimations of O₃ levels due to inadequate representation of NOx and OVOCs, particularly under background atmospheric conditions [81,82,83]. Additionally, grid resolution has emerged as a critical factor: model experiments showed that coarse grids (e.g., >120 km) tend to smooth NOx gradients, leading to substantial overestimations of tropospheric O₃, particularly in complex emission settings [84,85].

In parallel, OBMs have been extensively used for site-level photochemical diagnosis, offering high temporal resolution insights into O₃ formation regimes. These models typically ingest real-time concentrations of VOCs, NOx, O₃, and radical intermediates into simplified chemical mechanisms (often derived from MCM or SAPRC) to estimate ozone production rates and identify sensitivity transitions. In coastal and urban settings such as Hong Kong and southeastern China, OBM-based trajectory and radical analyses have pinpointed dominant oxidation pathways and the role of specific precursors (e.g., HONO and CO) in driving local ozone peaks [6,86]. In Beijing, OBM simulations during autumn haze events revealed differing radical cycling mechanisms across compound pollution regimes, while studies across the Yangtze River Delta have documented dynamic transitions from VOC-limited to NOx-limited regimes as pollution episodes evolve [87].

Notably, hybrid OBM–CTM frameworks are gaining traction, particularly in regions with complex emission topographies. For example, integrated studies in Tongling, Jinan, and Wuhan have demonstrated that OBM can provide real-time diagnostics while CTM facilitates attribution and forecasting, together revealing acute sensitivities to reactive VOC species and nonlinear O₃-precursor relationships. These synergies enable more reliable source apportionment and help inform spatiotemporally targeted emission control strategies [78,88].

Emerging innovations also aim to overcome the limitations of traditional modeling. Recent applications of deep learning algorithms have shown the ability to emulate CTM output while drastically reducing computational costs and improving representation of nonlinear atmospheric chemistry [78]. Meanwhile, comparative studies in cities such as Shanghai have attributed discrepancies between OBM and emission-based models (EBMs) to inconsistent assumptions about carbonyl compound speciation, NO₂ accuracy, and spatial averaging effects, underlining the importance of input harmonization [89].

These developments underscore the unique strengths and persistent challenges of both modeling paradigms. OBMs excel at regime classification and local photochemical insight, while CTMs offer system-level simulations across policy-relevant scales. As shown in

Table 4. Representative high-impact studies on OBM and CTM applications in tropospheric ozone research.

Article Title	Year	Authors	Affiliations	Publication /Source Titles	Cited Reference Count	Contribution
Ozone production and hydrocarbon reactivity in Hong Kong, Southern China	2007	Zhang, J et al. [6].	Hong Kong Polytechnic University	Atmospheric Chemistry and Physics	129	OBM trajectory analysis coupled with δCO/δNOy tracer diagnostics quantified that 50% of extreme O3 episodes (>100 ppbv) resulted from Guangdong-sourced precursors, while VOC-HONO constraints in the box model resolved 50–100% local photochemical amplification.
Global tropospheric ozone modeling: Quantifying errors due to grid resolution	2006	Wild, O et al. [84].	Japan Agency for Marine-Earth Science and Technology (JAMSTEC)	Journal of Geophysical Research: Atmospheres	104	CTM multi-resolution experiments (T21–T106) exposed artifactual O3 overproduction (up to +27% at 5.6°) from spatiotemporal smearing of NOx gradients, demanding sub-120 km grids for fidelity in nonlinear chemistry.
The effects of lightning-produced NOx and its vertical distribution on atmospheric chemistry: sensitivity simulations with MATCH-MPIC	2005	Labrador, LJ et al. [90].	Max Planck Society	Atmospheric Chemistry and Physics	86	MATCH-MPIC CTM parametrization established free-tropospheric lightning NOx injection (98%) as critical for accurate O3-OH-HNO3 covariance—elevating trans-Pacific NOy export by 24–43% versus convective schemes.
Atmospheric oxidation capacity and ozone pollution mechanism in a coastal city of southeastern China: analysis of a typical photochemical episode by an observation-based model	2022	Liu, TT et al. [86].	Chinese Academy of Sciences	Atmospheric Chemistry and Physics	79	OBM-MCM radical tracking decoupled daytime/nighttime AOC drivers (HONO: 33% ROx initiation), with CTM sensitivity runs confirming VOC-limited O3 suppression under suppressed NOx scenarios.
Global chemical transport model study of ozone response to changes in chemical kinetics and biogenic volatile organic compounds emissions due to increasing temperatures: Sensitivities to isoprene nitrate chemistry and grid resolution	2009	Ito, A et al. [91].	Japan Agency for Marine-Earth Science and Technology (JAMSTEC)	Journal of Geophysical Research: Atmospheres	74	CTM-simulated ΣANs-O3 correlations demonstrated that NOx recycling parametrization governs BVOC–O3 climate sensitivity as profoundly as grid resolution (+40 Tg burden uncertainty).
Seasonal variability of secondary organic aerosol: A global modeling study	2004	Lack, DA et al. [92].	Queensland University of Technology (QUT)	Journal of Geophysical Research: Atmospheres	71	CTM intercomparison (bulk-yield vs. partitioning algorithms) exposed oxidant-driven SOA biases over biomass-burning regions, reducing global flux uncertainty from 39% to mechanistically resolved regimes.
The Mediterranean summertime ozone maximum: global emission sensitivities and radiative impacts	2013	Richards, NAD et al. [93].	University of Leeds	Atmospheric Chemistry and Physics	67	TOMCAT CTM emission-tagging isolated near-surface O3 sensitivity to local NOx (9× > global sources) versus Asian-sourced UT O3 radiative dominance—reconciling TES/GOME-2 column discrepancies.
Parameterization of secondary organic aerosol mass fractions from smog chamber data	2008	Stanier, CO et al. [94].	University of Iowa	Atmospheric Environment	59	Derived CTM-ready surrogate-VOC basis set with fixed saturation concentrations (C = 0.1–103 μg/m3), eliminating enthalpy-driven volatility uncertainties in aerosol modules.
Sensitivity to grid resolution in the ability of a chemical transport model to simulate observed oxidant chemistry under high-isoprene conditions	2016	Yu, KR et al. [85].	Harvard University	Atmospheric Chemistry and Physics	48	GEOS-Chem 0.25° simulations resolved mesoscale NOx-isoprene segregation, suppressing high-NOx pathway errors by 5% and rectifying SEAC4RS HCHO validation failures.
Comprehensive Insights Into O3 Changes During he COVID-19 From O3 Formation Regime and Atmospheric Oxidation Capacity	2021	Zhu, SQ et al. [87].	Fudan University	Geophysical Research Letters	48	Source attribution CTM decoupled lockdown O3 surges into AOC-enhanced local chemistry (60%) vs. imported pollution, triggering OBM-validated VOC→NOx-limited transition.

, high-impact studies leveraging these models have shaped current understanding of ozone production, regional transport, and precursor sensitivity. Their integration represents a critical pathway toward improved model realism and decision-making relevance, especially under the dual pressures of climate change and urban air quality degradation.

4.3. Remote Sensing-Based Inversion of Precursors and Modeling of Ozone Formation Mechanisms

The integration of satellite remote sensing with chemical transport modeling has revolutionized the diagnosis of ozone (O₃) formation regimes, allowing a transition from station-bound empirical methods to spatially continuous, mechanism-driven analyses. Among the key breakthroughs is the utilization of tropospheric column densities of formaldehyde (HCHO) and nitrogen dioxide (NO₂), which serve as proxies for volatile organic compounds (VOCs) and nitrogen oxides (NOx), respectively. Their ratio, ΣHCHO/ΣNO₂—commonly referred to as the formaldehyde-to-NO₂ ratio (FNR)—has become a satellite-accessible metric for delineating ozone formation sensitivity across VOC-limited, NOx-limited, and transitional regimes.

Initial efforts employed static FNR thresholds, often derived from model-based benchmarks over North America. However, the expansion of satellite platforms such as OMI, TROPOMI, and GOME-2, alongside the rise of data assimilation techniques, has enabled more nuanced applications. In particular, chemical transport models (CTMs) such as GEOS-Chem and CMAQ have been coupled with satellite-derived FNR data through adjoint sensitivity analysis and four-dimensional variational (4D-Var) inversion, supporting the dynamic reconstruction of precursor emissions and chemical response fields. Furthermore, the emergence of source-resolved indicators such as the glyoxal-to-formaldehyde ratio (RGF) and sector-tagged ozone formation potential (OFP) has enriched the attribution of ozone regimes to specific anthropogenic or biogenic sources [95]. These methodological advancements collectively support a shift toward emission-resolved, vertically stratified, and policy-relevant ozone regime assessment frameworks.

Three representative studies exemplify this methodological evolution from different perspectives:

In a national-scale study of China’s ozone trends from 2013 to 2021, Ren et al. systematically derived region-specific FNR thresholds using high-resolution TROPOMI observations, validated against surface O₃ measurements across over 1400 monitoring sites. Their results revealed a spatially heterogeneous distribution of O₃ sensitivity: megacity clusters such as the North China Plain, Yangtze River Delta, and Chengdu–Chongqing Basin remained VOC-limited, while rural and western regions were predominantly NOx-limited. More importantly, they linked the failure to suppress VOC emissions to sustained or worsening ozone levels, despite notable NOx reductions under the Clean Air Action Plan. By quantitatively estimating the VOC/NOx reduction ratio required for effective ozone control in key cities, this work exemplifies the use of remote sensing to directly inform emission policy adjustment. These spatial patterns and multi-year dynamics are illustrated in Figure 12, which depicts the temporal evolution of high-percentile MDA8 ozone concentrations across major Chinese regions from 2014 to 2021.

Focusing on the 2018 Canadian wildfires, Alvarado et al. employed TROPOMI retrievals of HCHO and glyoxal (CHOCHO) to investigate the long-range transport and source identity of ozone precursors [96]. By coupling FLEXPART dispersion modeling with spectral decomposition, they inferred that CHOCHO and HCHO observed over distances exceeding 1500 km were not only primary fire emissions but also products of continuous oxidation of long-lived VOC precursors. The glyoxal-to-formaldehyde ratio (RGF) was utilized to distinguish pyrogenic sources from biogenic backgrounds. This approach expanded the diagnostic resolution of remote sensing from emission regime mapping to source differentiation and chemical aging in dynamic plumes. These spatiochemical insights are visually captured in Figure 13, which illustrates the co-distribution of CHOCHO, HCHO, and FLEXPART-modeled tracers across the wildfire-affected domain.

Coupling meteorological process dynamics with O₃ prediction using STRMF-informed deep learning [97]. While not a conventional inversion model, Hu et al. demonstrated how meteorological forcing could be integrated into data-driven ozone prediction by using CNN-LSTM architectures informed by the spatiotemporal evolution of key variables such as temperature, humidity, solar radiation, and wind vectors. By encoding regional meteorological fields as “weather videos,” their model accurately predicted daily MDA8 ozone concentrations in the North China Plain and Yangtze River Delta, with over 85% explained variance. Although primarily a machine learning study, it reinforces the necessity of accounting for meteorology-induced variability when interpreting satellite-inferred ozone sensitivity and hints at future possibilities for hybrid physical-AI modeling paradigms. The underlying architecture of this STRMF-informed CNN–LSTM framework is shown in Figure 14, highlighting the sequential encoding of meteorological drivers into spatiotemporal ozone prediction.

Collectively, these studies demonstrate a convergence toward integrated remote sensing frameworks where vertical sensitivity analysis, emission tagging, and dynamic meteorological coupling enable refined diagnostics of ozone production. As satellite sensors continue to evolve in resolution and spectral fidelity, these methodological innovations offer a robust basis for predictive, regionally tailored, and source-resolved ozone control strategies.

4.4. Artificial Intelligence for Ozone Modeling: From Data-Driven Generalization to Mechanistic Inference

Over the past decade, artificial intelligence (AI)—especially deep learning (DL)—has profoundly reshaped the landscape of surface ozone (O₃) modeling. Traditionally dominated by chemical transport models (CTMs) and satellite-driven statistical inversion techniques, the field has long grappled with computational inefficiencies, limited spatial resolution, and insufficient adaptability to complex, nonlinear atmospheric interactions. In contrast, AI-based approaches now offer a scalable and flexible alternative, capable of capturing intricate spatiotemporal dynamics, integrating heterogeneous data sources, and even quantifying uncertainty in policy-relevant scenarios.

At the core of this paradigm shift lies the ability of AI models to absorb multisource inputs—including satellite radiance, meteorological reanalysis, emission inventories, and land surface characteristics—and learn nonlinear mappings across space and time. Foundational architectures such as convolutional neural networks (CNNs), long short-term memory (LSTM) networks, and graph convolutional networks (GCNs) have laid the groundwork for more sophisticated hybrid models that integrate attention mechanisms, autoregressive dynamics, and probabilistic inference [98,99]. This trajectory has moved the field from mere data fitting toward mechanistic approximation and decision-support system development.

(1) From Single-Pollutant Estimation to Joint Inference

A key advancement has been the development of joint inversion frameworks that treat O₃ not in isolation but as part of an interdependent pollutant system. The SOPiNet model exemplifies this shift by simultaneously retrieving PM_2.5 and O₃ concentrations from MODIS satellite observations, meteorological reanalysis, and CTM forecasts using a multi-head attention deep learning architecture. Unlike conventional LSTM models, SOPiNet dynamically prioritizes historical inputs across a 20-day window, achieving a 3.8% gain in temporal R² for O₃ and enabling accurate inference even in data-sparse regions. Moreover, by embedding spatial-temporal interaction features (e.g., province-month coupling) and allowing masked loss for missing labels, the model facilitates near-real-time, nationwide 5 km resolution mapping with robust generalization capabilities. The structural composition and operational workflow of the SOPiNet model are illustrated in Figure 15, including its multi-head attention module and temporal sequence extraction for pollutant inference.

(2) Scaling Up: Full-Coverage O₃ Fields with Deep Spatiotemporal Convolutions

Another milestone is the construction of full-coverage, high-resolution O₃ datasets using 3D convolutional neural networks (3D-CNNs) [101]. A notable study developed a pipeline that first imputes missing OMI-NO₂ tropospheric column data via LightGBM, then feeds the completed dataset—together with ERA5 meteorology, NDVI, land cover, and population density—into a 3D-CNN architecture. The network captures both spatial proximity and temporal lag effects over a 7-day × 5 × 5 grid structure, producing a 0.1° daily MDA8 O₃ dataset for China from 2016 to 2020. Compared with TAP and CHAP benchmark datasets, this model achieves lower RMSE (15.65 μg/m³) and exhibits minimal interpolation artifacts, making it a valuable tool for epidemiological risk assessments and national pollution control evaluation. Figure 16 illustrates the complete workflow of this 3D-CNN-based framework, including preprocessing, data integration, and spatiotemporal inference of surface ozone concentrations across China.

(3) Structural Learning and Model Explainability

Beyond accuracy, AI models are increasingly embedded with structural priors and interpretability modules to support regulatory and scientific applications. Res-GCN-BiLSTM, for instance, constructs topological representations of air quality monitoring networks and learns inter-site transport patterns using residual GCN layers, while capturing temporal autocorrelations through bidirectional LSTMs. Likewise, hybrid systems combining GCNs, multilayer perceptrons (MLPs), and Gaussian process regression (GPR) enable both high predictive fidelity and uncertainty quantification [103,104].

Explainable AI methods such as SHAP (SHapley Additive exPlanations) have further deepened the field’s interpretive capacity. PMF-SHAP frameworks have quantified nonlinear interactions between VOC emissions and meteorological modifiers, revealing, for example, a negative SHAP value of −2.96 for NO_x-suppression zones and +1.38 for industrial ozone enhancements. Similarly, GraphSAGE-based spatiotemporal models have identified solar radiation as the leading 6-h predictor for Houston’s ozone levels (SHAP = 0.38), guiding targeted control interventions.

(4) Toward Real-Time Applications and Generalizable Frameworks

The frontier of AI in ozone modeling is rapidly extending toward real-time decision-making and operational deployment. Models such as SOPiNet and LightGBM-3DCNN not only produce spatially continuous fields but also outperform traditional models in speed and robustness. IoT-integrated platforms combining LSTM/RF pipelines enable predictive control of industrial emissions. Meta-learning systems employing stacked generalization strategies demonstrate superior forecasting for air quality indices (AQI) and air quality grades (AQG), achieving F1 scores exceeding 0.97 [105].

Moreover, algorithmic innovations such as genetic algorithm-tuned LSTMs and SHAP-integrated CatBoost regressors allow for adaptive optimization and transparent factor attribution, reducing overfitting and enhancing trustworthiness [106]. These frameworks signal a transition from isolated model-centric workflows to ecosystem-level AI architectures that interlace data assimilation, feature engineering, explainability, and stakeholder utility.

While recent AI architectures have made notable progress in modeling ozone dynamics, a critical frontier remains under-addressed: the explicit incorporation of extreme temperature events, such as urban heatwaves, into predictive frameworks. Temperature not only modulates VOC reaction rates and boundary-layer mixing but also interacts nonlinearly with humidity, solar radiation, and NO_x availability to shift ozone formation regimes. However, many existing AI models rely on smoothed reanalysis inputs or historical climatology, which limits their responsiveness to short-term thermal anomalies. Moving forward, future frameworks should integrate heatwave indicators, diurnal temperature extremes, and land surface temperature (LST) anomalies as dynamic input features—ideally coupled with regime-aware modeling strategies that embed photochemical regime logic into model design, either through dynamic thresholding or chemically informed constraints, to better capture the thermodynamic thresholds of ozone sensitivity. Such enhancements are essential for ensuring model reliability under evolving climate extremes, especially in pollution-prone megacities where VOC abundance intersects with thermal stress.

5. Discussions

5.1. Platform Interoperability and the Observation–Model Divide

Although the VOC–ozone research community has witnessed remarkable progress in observational granularity and computational modeling, the architecture of the field remains structurally fragmented. At the observational level, the vertical stratification of platforms—ground-based, airborne, and satellite—has enabled coverage across scales but has yet to coalesce into a fully interoperable system. Each platform contributes data within its designed altitude, resolution, and sensing constraints, but mutual calibration and operational synergy remain limited. Satellite retrievals, while spatially extensive, often lack the vertical resolution and chemical specificity required to resolve boundary-layer VOC chemistry; yet these limitations are rarely compensated by concurrent airborne or ground-based profiles. Conversely, aircraft-based missions offer high-fidelity vertical snapshots but are campaign-bound and episodic, with limited capacity to anchor long-term trends. Ground stations provide rich chemical speciation and temporal continuity but remain geographically sparse and biased toward urban settings. Despite their conceptual complementarity, these platforms often operate in silos—linked more by retrospective validation than by real-time functional integration. As a result, the observational system functions less as a cohesive sensing network than as a collection of parallel instruments, each illuminating a portion of the atmospheric puzzle but seldom reconstructing the whole.

These platform-specific strengths and limitations are not merely observational constraints but critically shape how different data streams contribute to the broader VOC–ozone research architecture. In our synthesis, we explicitly accounted for such heterogeneities by aligning platform roles with their operational advantages: satellite data were primarily weighted in trend detection and spatial anomaly mapping, aircraft and UAV data in vertical structure reconstruction and episode diagnostics, and ground-based observations in chemical regime classification and precursor–O₃ sensitivity analyses. This functional weighting ensures that platform data are interpreted not as interchangeable but as complementary inputs with domain-specific validity and epistemic roles.

Nevertheless, real-world efforts at platform integration often expose systemic interoperability gaps. For example, the assimilation of satellite-derived total-column HCHO into urban-scale CTMs frequently fails to constrain surface-level chemistry due to coarse pixel resolution and lack of vertical discrimination, especially in heterogeneous urban terrains. UAV-based campaigns, while offering low-altitude flexibility, are often limited in sensor payload and temporal coverage, making their datasets difficult to co-register with satellite overpasses or aircraft transects. Similarly, aircraft profiling missions rarely coincide with satellite or ground station windows, limiting their value for direct intercalibration. These mismatches are not isolated incidents but recurring structural failures that reflect the absence of synchronized deployment logic, common data standards, and platform-aware fusion frameworks.

This disjunction extends and indeed deepens when observations are interfaced with models. Whether mechanistic (e.g., CTMs and OBMs) or data-driven (e.g., CNNs and LSTMs), most modeling frameworks engage with observational data only intermittently and abstractly—often relying on preprocessed or downsampled proxies such as total-column HCHO, grid-averaged NO₂, or reanalysis-based meteorology. The fine-scale heterogeneity, regime shifts, and source-specific behaviors captured by high-resolution measurements are frequently lost in this translation. In turn, models calibrated to these smoothed inputs may achieve statistical fidelity but fail in diagnostic robustness, particularly near sensitivity thresholds or under regime perturbation. This is acutely evident in rapidly evolving domains—such as the transition from VOC-limited to NOx-limited chemistry following emission control interventions—where modeling outputs diverge from observational indicators not due to data scarcity, but due to representational misalignment.

What is required, therefore, is a structural convergence—both across observational platforms and between observation and modeling systems. Internally, the sensing architecture must evolve toward recursive interoperability: satellite retrievals anchored by tower-derived gradients, aircraft profiling guided by real-time surface diagnostics, and ground campaigns designed to validate orbital anomalies rather than operate in isolation. Externally, models must cease to treat observations as static inputs and instead engage them as dynamic constraints—informing model initialization, sensitivity tuning, and predictive uncertainty quantification. In turn, observational strategies should be designed not merely for data collection but for modeling utility: prioritizing chemically informative tracers, temporally aligned deployments, and regime-sensitive indicators. Only through such recursive, task-aware coupling can the field transcend the current observation–model divide and achieve a genuinely integrated framework—one capable not only of describing the atmosphere but also of simulating its chemical evolution and supporting actionable intervention.

5.2. Spatial Mismatch, Sensitivity Drift, and the Fragility of Predictive Modeling

Even as models grow in complexity and accuracy, their operational reliability remains undermined by two persistent challenges: the mismatch between modeling and observational scales and the inherent instability of ozone precursor sensitivity regimes under shifting environmental baselines. Most state-of-the-art modeling frameworks—whether mechanistic or data-driven—operate at spatial resolutions that are coarse relative to the heterogeneity of emission sources and meteorological dynamics. While satellite retrievals and CTM outputs typically span 1–10 km grids, effective ozone control requires diagnostics at sub-kilometer scales, where the influence of roadways, industrial corridors, and topographical features becomes non-negligible. This disconnect is not merely one of resolution but of misalignment: key processes such as VOC speciation, chemical aging, and turbulent exchange occur on scales finer than most models resolve, leading to systematic dilution of gradients and blurring of sensitivity boundaries.

The implications of this mismatch become particularly acute near regime transition zones. In regions characterized by overlapping anthropogenic and biogenic emissions—such as peri-urban belts or monsoon-influenced basins—ozone formation can toggle between VOC-limited and NO_x-limited states within hours, depending on insolation, temperature, and boundary-layer dynamics. Yet most models, especially AI-based predictors, tend to encode these dynamics as static or slowly varying features, leading to brittle predictions when underlying conditions shift. In policy evaluation scenarios, this can result in perverse outcomes: emission reductions that are forecasted to improve ozone levels may in fact exacerbate them, as the regime structure adjusts in ways the model fails to anticipate. The 2020–2022 episodes in several Chinese megacities—where post-lockdown NO_x reductions coincided with unexpected ozone rebounds—serve as cautionary cases of this modeling fragility.

Adding to the complexity is the temporal instability of chemical sensitivity regimes in response to both anthropogenic control and climatic fluctuation. The progressive substitution of solvent-based VOC sources, the re-emergence of biomass burning plumes, and the intensification of extreme heat events all act to reshape the reactive balance of the lower atmosphere. Yet most models are calibrated to historic regimes and offer limited adaptivity to these emerging conditions. Even when observational data suggest a shift—such as rising HCHO/NO₂ ratios or altered oxidant partitioning—models may continue to forecast responses based on outdated sensitivity fields. In effect, the atmosphere evolves faster than the models that seek to describe it.

Addressing these limitations requires a paradigm shift in both model architecture and deployment logic. Grid-based chemical models must move toward dynamically adaptive resolution, incorporating nested domains, localized emission perturbation tests, and real-time assimilation of sensitivity indicators. AI models, for their part, must embed regime-aware loss functions, temporal segmentation strategies, and structural priors that reflect known photochemical pathways. However, perhaps more fundamentally, predictive modeling must cease to be treated as a closed-loop system. It must instead function as a conditional scaffold—updated recursively by high-resolution measurements, stress-tested across extreme regimes, and continuously validated against the shifting boundaries of chemical sensitivity. Only under such a framework can ozone modeling transition from static forecast generation to dynamic, policy-relevant scenario testing.

5.3. Artificial Intelligence in VOC–O₃ Modeling: Structural Capabilities, Functional Gaps, and Systemic Remedies

Artificial intelligence has emerged as a transformative force in surface ozone modeling, particularly under the pressure to generate high-resolution, scalable, and near-real-time predictions. Models based on convolutional neural networks, long short-term memory units, and spatiotemporal fusion architectures have shown great success in reproducing observed O₃ concentrations, often outperforming traditional CTMs and OBMs in statistical accuracy. Yet despite their impressive numerical performance, AI-based models remain structurally under-equipped to address the specific demands of VOC–O₃ interaction modeling. This mismatch stems not from algorithmic weakness per se, but from a deeper incongruity between the epistemology of AI systems and the atmospheric processes they seek to approximate.

The first and most fundamental limitation lies in the inability of current AI architectures to internalize the regime-sensitive nature of ozone formation. Unlike physical models that simulate the chain of VOC oxidation, radical cycling, and NO_x photolysis, deep learning models operate as end-to-end regressors—mapping precursor fields, meteorological inputs, and land-use attributes directly to O₃ concentrations without decomposing the underlying chemistry. As a result, they offer little traction on the most policy-relevant question: which precursor reductions will actually lower ozone under a given regime? This becomes critical in urban or peri-urban zones where regime transitions occur over short spatial or temporal scales and where model misclassification—e.g., interpreting a VOC-limited regime as NO_x-limited—can lead to control policies that are not only ineffective but counterproductive. Such failures are not hypothetical: they have been documented in data-driven forecasts for cities such as Jinan, Suzhou, and Wuhan, where regime toggling across seasons or emission patterns caused pronounced divergence between model guidance and observed ozone outcomes.

A second structural deficiency concerns the representation of VOC sources and reactivities. While many AI models include satellite-retrieved HCHO columns or land-surface descriptors as proxy features, few distinguish between VOC classes or sources, and even fewer weight them by reactivity or yield. This abstraction limits the model’s ability to respond accurately when source composition changes—such as the rise of halogenated solvents, biofuel additives, or secondary oxygenated VOCs under high-radiation conditions. Furthermore, the spatial resolution of VOC inputs is often coarser than the scale of critical emission zones, such as industrial fence-line clusters or traffic corridors, leading to diluted source signatures and over-smoothed regime boundaries. Models trained under these constraints may perform well on global metrics but fail at source disambiguation, sector attribution, or high-variability scenarios.

In particular, the emergence of novel VOC sources—such as halogenated solvents, advanced biofuels, or thermally unstable additives—poses additional challenges for generalization, as these compounds are often underrepresented or entirely absent from historical training data. To address this, recent studies have begun incorporating proxy features such as source-tagged emission inventories, photochemical yield indices, or satellite-inferred reactivity gradients into model inputs. Transfer learning and domain adaptation techniques also offer a promising avenue for extending model validity to chemically evolving emission landscapes. While still in early stages, these approaches highlight the need for structural flexibility and continuous updating mechanisms in AI-based models if they are to remain policy-relevant under rapidly changing VOC profiles.

Third, AI models are inherently vulnerable to generalization failure under meteorological or emission perturbations that fall outside the training distribution. Ozone production is known to respond nonlinearly to extreme heat, stagnant air masses, biomass burning events, and shifts in VOC speciation induced by regulatory intervention. Yet in most AI workflows, such perturbations are not explicitly encoded or stress tested. Models trained on climatologically stationary datasets exhibit degraded performance under anomalous events, precisely when predictive insight is most needed. The 2022 heatwave episodes over Eastern China, for instance, exposed this fragility: multiple deep learning models underestimated the persistence of nocturnal O₃ accumulation due to unmodeled inversion and radical trapping mechanisms—features that physical models could simulate, but AI systems failed to recognize due to feature blindness.

A fourth and often overlooked limitation lies in the model objectives and loss function design. AI-based ozone models are typically optimized for overall RMSE or R² across broad spatial domains, but these metrics obscure performance at regime boundaries or near peak exceedance events. As a result, models may fail to penalize false negatives in VOC-limited zones or underestimate peak concentrations that are epidemiologically significant. Without loss structures tailored to regime fidelity, policy threshold detection, or precursor attribution accuracy, these models risk converging to high accuracy with low policy utility.

These limitations are not intrinsic to AI itself but to how it is currently implemented. Corrective pathways are increasingly being explored. On the input side, new models now incorporate precursor reactivity metrics, emission sector tags, and photochemical surrogate variables (e.g., O_x production efficiency) to encode mechanistic priors. On the architecture side, hybrid frameworks—such as physics-informed neural networks, CTM-AI fusion layers, or graph-based representations of monitoring networks—have shown promise in combining chemical interpretability with data-driven flexibility. On the output side, model explainability tools such as SHAP, LIME, and attention heatmaps are helping translate numerical predictions into chemically meaningful inference—allowing users to trace which inputs most strongly drive ozone forecasts under different conditions. However, these methods remain largely peripheral to model structure and are seldom deployed in operational pipelines.

The way forward lies in elevating these adaptations from post hoc augmentations to design principles. Future AI systems for ozone modeling must be constructed not simply to fit data but to solve chemically meaningful tasks—such as regime identification, sensitivity boundary detection, and source-contribution mapping. Their architectures must reflect the asymmetry and nonlinearity of atmospheric reactions; their training must prioritize sensitivity coherence over brute-force accuracy; their outputs must support policy evaluation, not just concentration reproduction. Most critically, they must function not as black boxes, but as intelligible intermediaries—integrating the vastness of observational data with the discipline of mechanistic constraints.

In the context of VOC–O₃ research, such models will not replace physical simulation but will redefine its boundaries. They will absorb observational anomalies, resolve spatial discontinuities, flag regime instabilities, and provide adaptive counterfactuals in ways that traditional models cannot. However, to do so, they must be designed as atmospheric reasoning systems—not regressors. The transition from predictive power to atmospheric relevance is not a matter of scale but of structure.

5.4. Toward Systemic Integration: From Structural Correction to Functional Governance

The limitations of current VOC–O₃ research frameworks—whether rooted in platform isolation, modeling misalignment, or algorithmic abstraction—ultimately converge upon a larger structural gap: the absence of a coherent, feedback-rich system linking observation, inference, and policy action. While each component—satellite sensing, airborne campaigns, ground-based monitoring, mechanistic modeling, and AI-based forecasting—has advanced substantially in isolation, they remain functionally fragmented. Observation feeds modeling, and modeling supports scenario evaluation, but rarely do these processes operate in synchrony, and even less frequently do they interact with real-time regulatory mechanisms. As a result, the system remains diagnostic rather than responsive, descriptive rather than adaptive.

This fragmentation is particularly problematic in the context of ozone governance, where regime transitions are nonlinear, emissions are rapidly evolving, and meteorological baselines are shifting under climate pressure. Policy effectiveness increasingly hinges on the ability not only to anticipate ozone exceedances but also to interpret their drivers and adjust mitigation pathways dynamically. Yet existing workflows—often built on static inventories, coarse-resolution simulations, and retrospective evaluation—lack the structural elasticity to support such responsive governance. Even where observational data are abundant, they are seldom assimilated into decision-making systems in a form that captures sensitivity structure, source attribution, or impact trajectories in near real-time.

What is needed is a paradigm shift—from modular advancement to systemic integration. At the core of this shift lies the construction of a recursive, task-oriented governance architecture in which sensing, simulation, and strategy are continuously interlinked. Observational data must no longer terminate at model boundaries but re-enter them through assimilation, regime validation, and anomaly detection. Models must cease to be passive forecasting engines and instead evolve into active scenario generators—flagging instability zones, simulating intervention outcomes, and quantifying the uncertainty range of precursor control options. Policy tools must likewise move beyond threshold triggers and embrace regime-informed logic, where emission reduction ratios are not fixed but dynamically allocated based on prevailing chemical sensitivity, meteorological context, and source profile evolution.

The foundational elements of such an integrated system are already emerging. Hybrid platforms that combine IoT-enabled VOC sensors with AI-based inference engines now allow sub-hourly updates on precursor distribution and regime hotspots. Machine-learning-driven control strategies, informed by real-time FNR maps and historical emission-responsiveness profiles, are being tested for adaptive traffic or solvent use restrictions. Regional atmospheric command systems—modeled loosely on carbon monitoring infrastructures—are being explored to link satellite-derived diagnostics with localized intervention protocols. However, these efforts remain embryonic, hindered not by lack of technology but by lack of system logic: an architecture that defines how observation and inference translate into coordinated action.

To build such logic, VOC–O₃ governance must adopt a systems perspective. It must treat air quality not as a product of emissions alone, but as the emergent outcome of interacting observational, modeling, and regulatory subsystems. It must prioritize structural redundancy—overlapping sensor types, multi-model triangulation, and cross-scale sensitivity verification—so that no single point of failure skews policy judgment. In addition, it must redefine response not as static compliance, but as dynamic interaction: a system that learns from its own outcomes, adapts to new conditions, and iteratively refines its strategy in step with atmospheric evolution.

Only by reimagining the governance landscape in this way—where VOC monitoring, O₃ modeling, and precursor control are co-designed and co-evolving—can the field move beyond descriptive diagnostics toward functional resilience. In this convergence lies not only the future of ozone mitigation but also the prototype for next-generation atmospheric governance more broadly.

While VOC–O₃ interactions are often studied through the lens of regional air quality management, mounting evidence suggests that their impacts can aggregate and interact to produce global-scale atmospheric consequences. Transboundary transport of ozone precursors, hemispheric pollution corridors, and the amplification of photochemical reactions under climate-induced temperature extremes all point to a systemic vulnerability beyond individual regions. For instance, persistent ozone enhancements over continental outflow zones—such as the North Pacific or South Atlantic—are increasingly linked to cumulative precursor emissions from disparate sources, many of which originate from rapidly developing urban clusters. Similarly, climate-driven increases in biogenic VOC emissions, coupled with stagnant air masses, can elevate background ozone levels at a planetary scale. These phenomena are not merely additive but synergistic, revealing feedback loops between local emissions, regional chemistry, and global atmospheric composition. As such, VOC–O₃ research must evolve to account for cross-scale propagation mechanisms and incorporate them into integrated governance strategies that transcend administrative boundaries. Only through such a perspective can mitigation policies achieve both local effectiveness and global relevance.

6. Conclusions

The atmospheric chemistry community stands at a critical turning point in understanding and managing ozone pollution. As volatile organic compounds (VOCs) continue to mediate the nonlinear formation of tropospheric ozone (O₃), research efforts have increasingly shifted from isolated observations and static simulations toward integrated, chemically meaningful systems of analysis. This review has systematically mapped the evolution of VOC–O₃ research across three interconnected dimensions—platform advancement, modeling innovation, and methodological rethinking—highlighting both the achievements made and the structural bottlenecks that persist.

On the observational side, diverse sensing platforms have enabled unprecedented access to VOC gradients and precursor dynamics, yet operational fragmentation remains widespread. Satellite, airborne, and ground-based platforms are rarely designed with cross-compatibility or model assimilation in mind, leading to missed opportunities in regime diagnosis and precursor attribution. Critical observational gaps persist in VOC speciation, near-surface vertical resolution, and suburban-scale temporal granularity, all of which undermine the fidelity of model inputs.

Modeling efforts—whether mechanistic or data-driven—have evolved rapidly, but they remain fundamentally constrained by their representational architecture. While CTMs offer chemical completeness, they often lag in responsiveness and localization. AI models, despite impressive statistical accuracy, tend to obscure mechanistic insight and falter under extrapolative stress. In particular, most models fail to internalize the regime-sensitive nature of ozone formation, producing outputs that are numerically precise yet chemically or policy-wise ambiguous. The challenge is not just accuracy, but interpretability, sensitivity, coherence, and operational relevance.

These limitations are not marginal—they reflect a systemic decoupling: observations remain disconnected from modeling inference, and models remain isolated from responsive policy action. Despite the proliferation of sensing platforms and algorithmic innovations, the field still lacks an integrated, feedback-oriented architecture wherein observational granularity informs model design, models adapt to regime shifts, and regulatory interventions align dynamically with evolving atmospheric sensitivities.

Yet within these limitations lies a clear opportunity. This review identifies not only the technical and conceptual gaps but also a converging pathway forward: intelligent sensing infrastructures, structure-informed neural networks, and chemically transparent regime modeling now collectively point toward a new operational paradigm. This paradigm is not merely analytical but decision-oriented—designed to support real-time diagnostics, scenario forecasting, and policy calibration within complex urban and regional contexts. Expanding participation from underrepresented regions—through open data, lightweight modeling approaches, and global collaborations—is essential to ensuring the generalizability and policy relevance of future VOC–O₃ research.

Our proposed synthesis—linking vertically structured VOC monitoring with regime-aware modeling frameworks—offers a replicable conceptual model for integrated atmospheric governance. Beyond advancing ozone science, it presents a scalable blueprint for environmental reasoning systems that are interpretive, adaptive, and actionable across spatial, temporal, and institutional boundaries.