Global warming amplifies atmospheric formaldehyde exposure and widens spatial inequality Authors: Jingkai Xue, Qihou Hu, Tiantian Li, Qihua Li, Wenjing Su, Hongri Hou, Ting Liu, Yizhi Zhu, Yizhou Xu, Xiaohan Wang, Chengxin Zhang, Peize Lin, Cheng Liu Published: 2026 ABSTRACT Formaldehyde (HCHO) is a globally distributed ambient air pollutant associated with respiratory disease, cancer, and premature mortality, with evidence of adverse health effects even at low ambient concentrations. As a major oxidation product of volatile organic compounds, atmospheric HCHO is strongly affected by natural sources such as vegetation and wildfires, which are highly sensitive to temperature and drought. However, its response to climate warming remains poorly constrained. Here, we develop a deep learning framework that extracts empirical climate sensitivities from more than a decade of daily satellite observations to project how ambient formaldehyde will evolve in a warming climate. We estimate that global mean HCHO increases by approximately 2.3% per degree of warming, largely associated with enhanced biogenic and wildfire emissions. Approximately one-fifth of the global population is already exposed to potentially risky HCHO concentration levels under present-day conditions. Climate-driven increases are projected to substantially expand this exposure burden. Under a mid-range emissions scenario (SSP2-4.5), global person-month exposure to concentrations exceeding 5 ppb approximately doubles by the end of the century. The largest increases occur in Africa and South America, while exposure in low Human Development Index (HDI) regions globally rises by more than a factor of three, indicating a pronounced amplification of environmental inequality. We further suggest that warming-driven increases may contribute to higher cancer-related health concerns in these regions. Our findings identify formaldehyde as a climate-sensitive yet under-recognized health-relevant pollutant and show that warming amplifies this neglected exposure burden and widens global environmental inequality. Introduction Climate change and air pollution are among the most pressing challenges in the Earth’s atmosphere, and their interaction represents a critical frontier in atmospheric science [[1], [2], [3]]. Climate-driven shifts in meteorology alter the emissions, transport, and chemical transformation of atmospheric constituents, reshaping pollutant burdens globally [[4], [5], [6], [7], [8], [9]]. Among air pollutants, formaldehyde (HCHO) has emerged as one with both high toxicity and strong climate sensitivity [[10], [11], [12], [13], [14], [15], [16], [17], [18], [19], [20], [21], [22], [23]]. Even at low concentrations of 5 ppbv, short-term exposure elevates cardiopulmonary disease risks and premature mortality, underscoring its public health relevance [10,[19], [20], [21],24]. Unlike primary pollutants, ambient HCHO is predominantly produced through the oxidation of volatile organic compounds (VOCs) from biogenic and anthropogenic sources, as well as biomass burning, and its abundance is strongly modulated by meteorological factors such as temperature, solar radiation, relative humidity, planetary boundary layer (PBL) height, wind speed and direction, and precipitation frequency [6,7,19,[25], [26], [27], [28], [29], [30]]. As warming accelerates biogenic fluxes, photochemistry, and fire activity, anticipating HCHO responses under future climate change is critical for assessing air-quality-related health risks [16,20,21,[31], [32], [33], [34]]. Despite its toxicological relevance, the role of formaldehyde in climate–air quality interactions remains far less understood than that of PM2.5 and ozone, largely due to observational constraints. Whereas PM2.5 and ozone benefit from extensive, long-standing global monitoring networks [6,35,36], HCHO measurements are sparse, fragmented, and geographically biased, with little coverage in tropical and developing regions [19,37]. Consequently, global HCHO assessments depend primarily on satellite retrievals, which offer broad spatial coverage but suffer from cloud interference, retrieval biases, and limited vertical sensitivity [[36], [37], [38]]. Numerical models have been widely applied to compensate for the lack of direct observations; yet their estimates diverge substantially due to differences in emission inventories and chemical mechanisms, leading to large inter-model discrepancies in simulated formaldehyde. For instance, global isoprene emission trends differ markedly across models [[38], [39], [40]], reflecting the absence of observational constraints essential for credible prediction. This fundamental lack of constraint hampers reliable global-scale understanding of HCHO’s future response to climate change and, consequently, limits accurate projection of evolving health risks under different warming trajectories. To overcome these limitations, data-driven approaches that learn directly from atmospheric observations have emerged as a promising complement in air quality research [[41], [42], [43], [44]]. Such methods can leverage observational data to capture the complex, nonlocal processes that govern HCHO distributions, without relying solely on potentially uncertain emission inventories or chemical mechanisms. However, most existing approaches are either point-based or locally constrained, or do not fully account for the spherical geometry of the Earth, which can introduce spatial inconsistencies in global-scale simulations [[45], [46], [47], [48]]. Moreover, while reanalysis or chemical transport model outputs can partially alleviate observational sparsity, they inevitably inherit uncertainties from emission inventories and chemical mechanisms [[49], [50], [51], [52]]. Therefore, to fully exploit observational data and mitigate these uncertainties, methods capable of learning from unlabelled observations through self-supervised strategies are needed. AI-based weather and climate models such as Pangu, GraphCast, and Spherical Fourier Neural Operator (SFNO) have provided new opportunities to learn global atmospheric dynamics efficiently from observations [42,46,53,54]. In particular, SFNO respects the Earth’s spherical geometry and learns nonlocal operators in spectral space, enabling rapid, physically consistent representation of global meteorology–chemistry interactions without relying on empirical parameterizations, while naturally supporting multi-resolution adaptation suitable for large-scale simulations [[55], [56], [57]]. This backbone can be readily adapted for self-supervised learning to extract patterns directly from unlabelled observations [[58], [59], [60], [61]]. Here, we present a satellite-driven framework that leverages spherical artificial intelligence to diagnose the sensitivity of ambient HCHO to meteorological variability, using a decade of Ozone Mapping and Profiler Suite (OMPS) observations (2013–2023). This record encompasses about 3600 daily snapshots and over one billion retrievals, spanning surface temperatures from sub-zero to >40 °C and HCHO column densities ranging from background to episodic peaks. The nonlinear learning capability of the model enables it to capture the influence of extreme meteorological conditions and complex interactions among drivers, revealing relationships that are often missed by conventional linear or process-based approaches. We then integrate these learned sensitivities with Coupled Model Intercomparison Project Phase 6 (CMIP6) climate projections to evaluate future changes in HCHO burdens under different warming scenarios. This framework provides the first global, data-driven assessment of climate-driven HCHO dynamics and their health consequences, offering new insights into climate–chemistry interactions and strategies to mitigate emerging health inequities in a warming world. Materials and methods AI framework for learning daily HCHO–weather relationships To maximize the scientific value of direct satellite measurements, we developed an artificial intelligence framework that learns the sensitivities of ambient formaldehyde to meteorological and emission drivers from sparse daily satellite retrievals. Our primary observational data consist of tropospheric HCHO vertical column densities (VCDs) from the OMPS between 2013 and 2023. These Level 2 data are retrieved using the Differential Optical Absorption Spectroscopy (DOAS) technique to infer column abundances of formaldehyde and other trace gases [20,[62], [63], [64]]. Daily OMPS HCHO retrievals were subjected to standard quality-control procedures to remove non-physical retrieval artifacts, including screening for cloud contamination (effective cloud fraction < 0.5) and poor spectral fitting quality [62]. These procedures preserve the intrinsic spatial and temporal sparsity of the satellite observations, which is explicitly accommodated by the AI framework rather than artificially gap-filled. To represent key drivers of HCHO variability, we incorporated meteorological variables from the European Centre for Medium-Range Weather Forecasts (ECMWF) Reanalysis v5 (ERA5), including wind speed, temperature, humidity, atmospheric pressure, and geopotential height at 1000 and 700 hPa. Emissions were drawn from the Copernicus Atmosphere Monitoring Service (CAMS), covering both natural (vegetation, wildfires) and anthropogenic (industry, traffic, residential) sources. We further included emissions of VOCs and nitrogen oxides (NOx), together with static and semi-static covariates such as population density, Leaf Area Index (LAI), and elevation (Table S12; detailed input variable list in Table S13). The model architecture is based on the SFNO, chosen for its ability to represent spatially continuous processes on a global sphere using Fourier transforms, while naturally supporting multi-resolution representations across scales [45,65,66]. To address the challenge of disjoint observational coverage, we implemented a self-supervised masked-modeling framework [59,61]. Detailed model parameters and training procedures are provided in Text S1. In this setup, the model does not require a pre-existing gap-free target. Instead, for an arbitrary date t, the model ingests the sparse OMPS observations (from adjacent days t−3 to t+2) alongside continuous meteorological drivers. Crucially, the valid satellite retrievals serve simultaneously as contextual inputs and as training targets. The training objective minimizes the reconstruction error exclusively at locations with valid observations, while missing regions are explicitly masked out from the loss function. This forces the model to learn the underlying physical dependencies between meteorological drivers and HCHO distributions to infer the missing values, effectively “filling in the blanks” based on learned spatiotemporal structures. Through this setup, the model yields both (i) reconstructions of daily global HCHO fields where satellite data are missing and (ii) implicit sensitivities of HCHO to meteorological drivers, which form the basis for subsequent future projections. To ensure close alignment with actual measurements while preserving spatial continuity, we applied incremental retraining (similar to data assimilation) in an online learning manner using the automatic differentiation feature of neural networks to continuously update the model parameters as new data became available [67,68]. Training was performed on 2013–2019 data and evaluated on 2020–2023 data. On the test set, the model achieved a coefficient of determination (R2) of 0.58 for daily reconstructions and 0.92 for monthly averages, performance that is within acceptable limits given the noise and retrieval errors of satellite observations (Fig. S1). We validated both the native and AI-enhanced OMPS HCHO fields against ground-based Pandora measurements from 26 Pandonia stations (Fig. S5, Table S8). Pandora multi-axis tropospheric columns (fuh5) were processed in Python 3.8, retaining only high/medium quality flags (0, 1, 10, 11). Hourly Pandora values were computed as medians within ±30 min of each OMPS overpass and collocated to a 0.1° × 0.1° grid using inverse-distance weighting; extreme outliers beyond ±1.5 × IQR were removed. Fig. S5 compares (a) native OMPS, (b) full AI-reconstructed fields, (c) reconstructions at missing-retrieval locations/times, and (d) reconstructions where OMPS had valid retrievals, all against Pandora. Overall, the AI framework substantially improves data coverage and reduces low-bias scatter, confirming that the learned sensitivities capture physically meaningful HCHO–meteorology relationships. Although satellite retrieval uncertainties may affect absolute HCHO levels, their impact on the inferred climate sensitivities is mitigated by the model’s utilization of temporal context and our focus on relative changes. The strong agreement with independent Pandora measurements (Fig. S5) further supports the robustness of the learned relationships against observational noise. Surface HCHO concentrations (measured in ppb) were inferred from reconstructed tropospheric column densities using surface-to-column conversion factors derived from the CAMS [20]. Specifically, CAMS daily three-dimensional HCHO fields were used to calculate grid- and time-specific ratios between near-surface concentrations (lowest model layer) and vertically integrated tropospheric columns. Conversion factors reflect the modeled vertical distribution of HCHO and its modulation by atmospheric mixing, chemical production, and boundary-layer processes. These estimates were validated against ground-based observations from the U.S. Environmental Protection Agency’s Air Quality System (U.S. EPA AQS), yielding strong correlations across many sites (Fig. S7, Table S10). While absolute concentrations are subject to uncertainties in modeled vertical profiles, temporal trends, relative differences among SSP scenarios, and large-scale spatial patterns are expected to be more robust. In areas with limited ground observations, estimates are best interpreted as reflecting relative, large-scale trends rather than exact local magnitudes. Transfer of learned sensitivities to monthly climate projections Our projection framework leverages statistically learned but physically consistent sensitivities of HCHO to meteorological drivers extracted from daily satellite observations (Section 2.1). We assume that these statistical dependencies (e.g., the temperature dependence of biogenic emissions) remain robust under future warming. By transferring these observationally constrained response functions to climate model outputs, we embed real-world HCHO–climate interactions into future scenarios without explicitly relying on free-running chemical transport simulations, thereby reducing reliance on uncertain chemical parameterizations. We utilized monthly climate predictors from an ensemble of four CMIP6 models (CESM2, MRI-ESM2, GFDL-ESM4, UKESM1) across four Shared Socioeconomic Pathways (SSPs: 1-2.6, 2-4.5, 3-7.0, and 5-8.5) [38,[69], [70], [71]], spanning a representative range of warming trajectories (Table S16). The predictor set includes surface temperature, relative humidity, and wind speed [40,69,70,72]. To capture essential biosphere–atmosphere feedbacks, we also incorporated projected changes in LAI, wildfire carbon flux, and land use. To minimize the influence of future changes in anthropogenic activity and focus on the effect of climate change on atmospheric HCHO, population density was held fixed at present-day levels throughout all projections. A critical technical challenge is adapting the daily-trained SFNO model (Section 2.1) to operate on monthly-averaged climate data while preserving its learned physical sensitivities. To address this, we implemented an adaptation strategy in which a lightweight, learnable linear projection layer was added at the model input to map the reduced vector of monthly predictors (surface temperature, humidity, winds, etc.) into a latent feature space compatible with the pre-trained SFNO. This projection maps the input predictor vector (dimension din, corresponding to the number of monthly variables) to a latent feature space of dimension dout, consistent with the channel dimension used by the SFNO backbone. This input adaptation simultaneously serves to reconcile the timescale difference between daily training and monthly projection under the assumption that HCHO responds approximately linearly to monthly-mean anomalies of meteorology, given its short lifetime. During training, the majority of the SFNO backbone parameters were frozen to preserve the nonlinear sensitivities learned from daily data, while only the input projection layer and the final two SFNO layers were updated. This strategy limits overfitting to the smaller monthly dataset. The model was trained using ERA5 monthly reanalysis meteorology (Table S14) paired with reconstructed seamless monthly HCHO column fields over the 2013–2019 period. Model training and evaluation followed a strict temporal split, with 2013–2019 used for fine-tuning and 2020–2023 reserved as an independent validation period. This design minimizes short-term temporal autocorrelation and ensures that model performance reflects its ability to predict previously unseen conditions prior to application to the CMIP6 ensemble (Table S15). The AI model was trained on daily meteorological observations spanning a wide range of conditions, including historical extremes, and was applied to predict future HCHO at the monthly scale. Although future climate simulations may exhibit localized monthly-mean extremes at individual grid points, the meteorological input space remains broadly represented by the historical daily variability (−20 °C to 40 °C) used for model training. This design helps constrain extrapolation and preserves the applicability of learned meteorology–HCHO relationships at the climatological scale considered here. At the monthly scale, averaging substantially reduces high-frequency random retrieval noise in OMPS HCHO columns [20,73] (Fig. S34). The AI-based predictions were validated against independent ground-based observations, including Pandora and Network for the Detection of Atmospheric Composition Change (NDACC), with biases in monthly means generally within 8%. While systematic biases in the satellite retrievals cannot be fully eliminated, comparison with independent observations demonstrates that the model reliably captures spatial patterns and relative long-term changes. Model evaluation against 2020–2023 observations confirmed strong skill (R2 = 0.90, Mean Absolute Percentage Error = 18%; Figs. S2 and S35–S37) in reproducing the joint influence of meteorology, emissions, and land surface conditions on monthly HCHO variability. Assuming that the biases in CMIP6 predictors are systematic and that the physical relationships hold, we then replaced the ERA5 inputs with the bias-corrected CMIP6 ensemble drivers to generate future HCHO projections under each SSP scenario. Bias correction was performed using a quantile-mapping approach [[74], [75], [76], [77]]. Further methodological details are provided in Supplementary Methods. Tables S14 and S15 summarize all predictors used in our projections. We employed a gradient-based attribution method to interpret the transferred sensitivities [78,79], computing gradients with respect to input features to quantify the contributions of individual drivers to projected HCHO anomalies. Attribution results are robust across model initializations (Fig. S16). To infer future surface HCHO concentrations, we applied grid-specific, monthly surface-to-column ratios derived from the 2013–2019 climatology. This approach assumes that the seasonal relationship between tropospheric column HCHO and surface concentrations, as characterized by a multi-year climatology, remains broadly representative under future climate conditions. This assumption is physically motivated by the short atmospheric lifetime and predominantly surface-based production of HCHO, such that column variability is largely driven by near-surface processes on climatological timescales. Fig. S8 compares the CMIP6 scenario-based projections of monthly mean surface HCHO concentrations for 2020–2023 with observational data from the U.S. EPA Air Quality System. Formaldehyde exposure assessment The formaldehyde-associated cancer risk is quantified using the EPA’s Integrated Risk Information System (IRIS), which posits that each 1-μg increase of formaldehyde per cubic meter of air elevates the cancer risk by a factor of 1.3 × 10−5 [19,21]. The uncertainty is inherent to the IRIS assessment and inherited by this study. Accordingly, the resulting cancer risk estimates should be interpreted as indicative, order-of-magnitude metrics intended to facilitate comparison with previous studies. At present, epidemiological evidence for robust population-level exposure–response relationships of ambient formaldehyde remains limited, particularly at regional to global scales. Building on a previous nationwide epidemiological analysis in China [10], we adopt 5 ppb as a high-exposure benchmark for this global assessment. Although derived from a single-country study, China spans a wide range of socioeconomic and income levels, and in the absence of better globally resolved epidemiological indicators, this benchmark provides a pragmatic empirical reference. Because our analysis uses monthly mean HCHO concentrations rather than daily values, the daily-based 5 ppb threshold may underestimate the prevalence of moderate but persistent exposure. To account for this, we introduce a lower-exposure benchmark (3 ppb) to capture extended exposure above moderate levels (Table S11). These benchmarks are not intended as universal toxicological thresholds but serve as empirically grounded reference levels to enable consistent comparison of exposure frequency, duration, and amplification under climate change across regions. Monthly mean HCHO concentrations were compared against these benchmarks to identify periods of sustained exposure. Using monthly averages yields a conservative representation of chronic conditions, as persistent exceedance of these levels reflects prolonged exposure rather than short-term episodic peaks. Population exposure was expressed in person-months by multiplying the number of exposed individuals within each grid cell by the duration of exceedance. To examine distributional and equity-related patterns, exposure estimates were stratified using independent gridded Human Development Index (HDI) data. Socioeconomic indicators were applied solely in post-processing analyses and were not included as predictive inputs in the modeling framework. This approach ensures that identified disparities reflect the spatial overlap between climate-driven changes in HCHO and population vulnerability, rather than being imposed by model design (Fig. 1).Fig. 1Conceptual framework for the reconstruction of daily HCHO from satellite observations, AI-driven sensitivity extraction, and future projections using CMIP6 predictors. SFNO, Spherical Fourier Neural Operators; CMIP6, Coupled Model Intercomparison Project Phase 6; SSP, Shared Socio-economic Pathways; HCHO, formaldehyde. Conceptual framework for the reconstruction of daily HCHO from satellite observations, AI-driven sensitivity extraction, and future projections using CMIP6 predictors. SFNO, Spherical Fourier Neural Operators; CMIP6, Coupled Model Intercomparison Project Phase 6; SSP, Shared Socio-economic Pathways; HCHO, formaldehyde. Results and discussion Climate control of global formaldehyde Quantifying the climatic regulation of atmospheric formaldehyde requires reliable long-term records that capture its fine-scale variability. To achieve this, we developed a hybrid reconstruction–prediction framework combining satellite observations from 2013 to 2023 with meteorological and emission data. The neural operator model learns spatiotemporal relationships among concentration fields, meteorological variables, and emission sources, allowing the recovery of daily HCHO distributions with high fidelity. By exploiting physical and chemical correlations, it effectively denoises satellite retrievals and fills persistent cloud-related gaps (R2 = 0.57; Table S8). Validation against Multi-Axis Differential Optical Absorption Spectrometer (MAX-DOAS), NDACC, and Pandora observations (Figs. S3–S5), together with good agreement with CAMS reanalysis (Fig. S6, Table S9), demonstrates that the reconstruction reliably captures historical daily HCHO patterns and variability across diverse regions and observing conditions, providing a robust foundation for investigating the mechanisms governing HCHO variability. Building on these reconstructions and the weighting of model predictors, we examined the climatic controls on HCHO during 2013–2023. Elevated concentrations occur in South America, Central Africa, and East Asia, reflecting strong biogenic volatile organic compound (BVOC) emissions [25,80,81]. Although the global mean remained relatively stable, regional increases of up to about 0.9% per year in South America were closely linked to rising temperatures and drought intensity. Feature-importance analysis highlights temperature and relative humidity as the two leading drivers (Figs. S13–S17), which together regulate BVOC fluxes [25], wildfire frequency [25,[82], [83], [84], [85], [86]], and atmospheric oxidation. A bivariate response analysis shows that under extreme warming (ΔT > 5 °C) and drying (ΔRH < −5%), HCHO VCDs can increase by more than 20% (Fig. 2).Fig. 2Joint sensitivity of ambient HCHO to temperature and relative humidity. Relative changes in HCHO as a function of changes in surface air temperature and relative humidity, aggregated across all grid cells and scenarios. This response surface highlights the dominant and nonlinear climatic controls on future HCHO variability. Joint sensitivity of ambient HCHO to temperature and relative humidity. Relative changes in HCHO as a function of changes in surface air temperature and relative humidity, aggregated across all grid cells and scenarios. This response surface highlights the dominant and nonlinear climatic controls on future HCHO variability. In tropical regions such as the Amazon and Congo basins (Figs. S10–S11), higher temperatures and stronger radiation stimulate BVOC emissions and photochemical production, while increasing aridity enhances wildfire activity, providing an additional source of HCHO. At mid- and high-latitudes, interannual variations arise from shifts among background chemistry, biogenic fluxes, and episodic fires [31]. Wildfires can contribute over 30% of regional column HCHO enhancements during peak fire seasons [87]. Changes in land use and vegetation also modulate these responses [88]. Moderate warming can extend growing seasons in cooler regions, whereas persistent drought reduces productivity and biogenic emissions. In densely populated areas, anthropogenic sources further amplify natural processes. Industrial and vehicular emissions, intensified by the urban heat island effect, maintain persistent HCHO hotspots [20]. Overall, the reconstructed HCHO record reveals a complex and climate-sensitive interplay among temperature, humidity, vegetation activity, and human emissions, emphasizing the compound influence of global warming on atmospheric composition. Divergent futures under climate pathways Building on the daily-scale sensitivities established in Section 3.1 and transferred to monthly predictors through our SFNO adaptation framework (Section 2.2), we quantify climate-driven changes in global HCHO concentrations under four CMIP6 SSP scenarios. In these projections, we minimize the influence of future anthropogenic changes in order to isolate the component of HCHO change associated with climate warming. This approach preserves the physically validated responses of HCHO to temperature (Fig. S28), humidity, biogenic fluxes, and wildfire emissions while enabling robust month-to-month predictions under future climate trajectories. Observation-constrained sensitivity analysis indicates that for every 1 °C increase in global surface temperature, the global land-average HCHO vertical column density rises by approximately 0.11 × 1015 molecules/cm2 [95% CI: (0.09–0.13) × 1015]. Across scenarios, surface temperature and relative humidity emerge as the principal modulators of HCHO, regulating VOC emissions and photochemical oxidation rates. Under the high-emission SSP5-8.5 pathway, the climate-driven enhancement in global tropospheric HCHO increases steadily through the late 21st century, with the largest anomalies relative to the 2010s baseline occurring by the 2090s (Figs. 3, S18). The inter-model spread, represented by the standard deviation, indicates that while all models agree on a significant upward trend, the exact magnitude of increase varies by approximately 2.8% relative to the ensemble mean. Regional enhancements are particularly pronounced in Africa, North America, and Central Asia, with increases exceeding 30% in some areas, whereas modest declines are projected for northern China and parts of South Asia. By contrast, the SSP2-4.5 pathway yields a more moderate increase, while coordinated mitigation under SSP1-2.6 keeps the climate-driven HCHO anomaly relatively small, illustrating a clear bifurcation between low-emission stabilization and high-emission amplification after 2050.Fig. 3Climate-driven changes in global HCHO column density under different SSP scenarios. (a) Temporal evolution of global annual mean HCHO VCD anomalies relative to the 2010s baseline under different SSP scenarios (SSP1-2.6 to SSP5-8.5), highlighting the climate-driven increase in HCHO concentrations. (b) Spatial distribution of HCHO VCDs in the 2090s under the SSP1-2.6 scenario, representing the baseline low-emission pathway. (c) Spatial changes in HCHO VCDs in the 2090s under SSP1-2.6 relative to the 2010s baseline. (d–f) Spatial differences in HCHO VCDs in the 2090s under (d) SSP2-4.5, (e) SSP3-7.0, and (f) SSP5-8.5 relative to SSP1-2.6, illustrating the incremental climate-driven enhancement of HCHO under stronger warming scenarios. VCD, vertical column density. Climate-driven changes in global HCHO column density under different SSP scenarios. (a) Temporal evolution of global annual mean HCHO VCD anomalies relative to the 2010s baseline under different SSP scenarios (SSP1-2.6 to SSP5-8.5), highlighting the climate-driven increase in HCHO concentrations. (b) Spatial distribution of HCHO VCDs in the 2090s under the SSP1-2.6 scenario, representing the baseline low-emission pathway. (c) Spatial changes in HCHO VCDs in the 2090s under SSP1-2.6 relative to the 2010s baseline. (d–f) Spatial differences in HCHO VCDs in the 2090s under (d) SSP2-4.5, (e) SSP3-7.0, and (f) SSP5-8.5 relative to SSP1-2.6, illustrating the incremental climate-driven enhancement of HCHO under stronger warming scenarios. VCD, vertical column density. CMIP6 Earth System Models (Fig. S33) show a wide range of amplification strengths, with substantial inter-model spread and initial states that can deviate from observations. Within this context, our AI framework reproduces the warming-driven amplification signal (reflecting only climate-driven changes, without anthropogenic VOC emissions) with a weaker increase and an observationally consistent starting point. Our projections exhibit a strong spatial alignment with the UKESM1 simulation (Fig. S9), reproducing the consistent warming-driven trends, particularly in biogenic-dominated regions (the Amazon and Africa). Mechanistically, temperature-driven intensification of isoprene and other BVOC emissions dominates the persistent rise in tropical HCHO. Under global warming, isoprene emissions in tropical regions could surge by 20%–60% by the century’s end [40]. At the same time, enhanced wildfire activity provides an additional indirect source, particularly in northern mid- and high-latitudes. Land-cover and vegetation dynamics further modulate these patterns, with warming extending growing seasons at high latitudes but exacerbating drought stress in tropical and subtropical regions. The combined effects of climate extremes, BVOC fluxes, and wildfire emissions generate pronounced spatial heterogeneity, with the largest climate-driven HCHO increases concentrated in densely vegetated tropical and lower-HDI regions. Overall, the scenario-specific trajectories reflect the coupled influence of climate, ecosystems, and fire-related feedbacks on atmospheric HCHO. By integrating validated daily sensitivities into monthly projections, this framework provides a mechanistic and quantitatively robust assessment of the climate-driven amplification of HCHO under global warming, forming the foundation for subsequent exposure and health-impact analyses. Amplified exposure through prolonged exceedance Under baseline conditions in the 2010s (Fig. S12), a substantial portion of the global population was already exposed: over 1.6–1.8 billion people (about 20% of the global population) experienced long-term exposure above the 3 ppb benchmark, while more than 6 million were exposed to concentrations exceeding the 5 ppb benchmark (Figs. S19 and S24). Future projections indicate that exposure burdens will be substantially amplified under warming climates (Fig. 4). Under the high-emission SSP5-8.5 scenario, rising temperatures are projected to increase global annual exposure duration by 104%–125% (approximately 1.9 months) by the 2090s. This amplification is geographically uneven: peri-urban regions, where interactions with urban-edge vegetation and wildfires are pronounced, are particularly affected, with roughly 630 million additional people exposed to concentrations exceeding the lower-exposure benchmark by the end of the century. Under the intermediate SSP2-4.5 scenario, the fraction of the tropical population experiencing sustained concentrations above this benchmark is projected to increase from 26.1% to 32.2%, highlighting the exacerbation of regional health inequities under climate change.Fig. 4Global projections of annual formaldehyde exposure and affected land-use areas under climate scenarios. (a) Spatial distribution of annual months exceeding the lower-exposure benchmark (3 ppb) in the 2010s. (b) Projected land area exposed to concentrations above the lower-exposure benchmark under SSPs. (c–f) Projected changes in annual exceedance duration (Δmonths) by the 2090s relative to the baseline.Red denotes increased duration; blue denotes decreased duration. Global projections of annual formaldehyde exposure and affected land-use areas under climate scenarios. (a) Spatial distribution of annual months exceeding the lower-exposure benchmark (3 ppb) in the 2010s. (b) Projected land area exposed to concentrations above the lower-exposure benchmark under SSPs. (c–f) Projected changes in annual exceedance duration (Δmonths) by the 2090s relative to the baseline.Red denotes increased duration; blue denotes decreased duration. This increase in risk is largely driven by amplification from terrestrial ecosystems. In the SSP5-8.5 scenario, temperature and moisture stress are projected to expand the area of forests and grasslands falling within HCHO risk zones. For instance, the fraction of forested areas exceeding the lower-exposure benchmark could rise from 21% at baseline to 36% (24%–39%). Tropical ecosystems are disproportionately affected, with current hotspots such as the Amazon and Congo basins expected to experience more than a 200% increase in the annual area above threshold levels. Conversely, low-emission pathways offer substantial mitigation potential. Compared with high-emission scenarios such as SSP5-8.5, SSP1-2.6 could effectively slow and constrain the expansion of formaldehyde risk, keeping global exposure duration and spatial extent close to baseline levels, with only minor increases in localized hotspots such as the Congo Basin. This underscores the critical role of proactive climate action in limiting future formaldehyde exposure and associated health disparities. Widening inequalities in a warming world Global formaldehyde exposure exhibits pronounced disparities closely linked to socioeconomic development. Our analysis reveals substantial spatiotemporal heterogeneity in projected HCHO exposure and indicates that climate change could further exacerbate global health inequities. Currently, low- and middle-income countries bear over 80% of the global HCHO-related cancer risk, driven by a combination of high biogenic emissions and dense populations [20]. The areas most in need of monitoring, however, have the least validation data, revealing a global observational inequity; expanding ground networks is crucial to reducing uncertainty. Present global disparities in HCHO exposure are already stark (Fig. S26). Approximately 22.1% of the global population experiences HCHO levels at or above the 3 ppb benchmark. Specifically, in regions with a high HDI (≥0.7), this proportion is 13.7%, while in low-HDI regions (HDI < 0.7), it reaches 33.7%. Regarding exposure duration, data from the 2010s show that residents in low-HDI regions experience an annual exposure above the lower-exposure benchmark of 1.18 months. In contrast, those in high-HDI regions have only 0.32 months of annual exposure, with the exposure duration in low-HDI regions being 3.7 times that in high-HDI regions (Fig. 5, S23). Future projections indicate an alarming widening of these disparities (Fig. S27). Under high-emission scenarios such as SSP3-7.0 and SSP5-8.5, person-months of exposure are projected to rise markedly after 2050 (Figs. 5a and S20–S25, Tables S1–S6), particularly in tropical and less developed regions. Under SSP5-8.5, 35.3% of the global population is exposed to the 3 ppb benchmark by the 2090s, with the exposure duration nearly doubling compared to current levels. The global population exposed to hazardous HCHO is projected to increase by 58%. More critically, the annual exposure above the 3 ppb benchmark is projected to reach 9.8 (8.5–11.2) billion person-months. Among these, roughly 67 million person-months will exceed the 5 ppb benchmark. These values are 20% and 84% higher, respectively, than those under the stringent SSP1-2.6 scenario. Even under the moderate SSP2-4.5 scenario, exposure to hazardous levels is projected to double. What’s more, the projected increase in exposure for low-HDI regions stands at 0.06 months per decade, a rate 2.1 (1.8–2.4) times faster than that in high-HDI regions and 1.6 times the global average. Such significant disparities are likely to substantially increase preventable mortality, cognitive impairments, and other long-term health issues associated with HCHO exposure [10]. Asia and Africa alone bear the vast majority of the projected cancer burden (Figs. S31–S32).Fig. 5Spatiotemporal patterns of formaldehyde exposure and socioeconomic disparities under climate scenarios. (a) Projected person-months exposed to formaldehyde concentrations above the 5 ppb threshold by continent in the 2010s, 2050s, and 2090s under SSP scenarios. (b) HDI-stratified exposure amplification: binned scatterplot (HDI intervals Δ = 0.1) showing normalized HCHO surface concentration changes (2090s vs. 2010s baseline) against development level; circle area scales with incremental exposure duration. (c) Temporal trends in formaldehyde exposure for high- and low-HDI regions, with projections of person-months exposed to concentrations above 3 ppb and 5 ppb thresholds. (d) Wildfire modulation effects: comparison between wildfire emission changes and pyrogenic emission control scenarios. Spatiotemporal patterns of formaldehyde exposure and socioeconomic disparities under climate scenarios. (a) Projected person-months exposed to formaldehyde concentrations above the 5 ppb threshold by continent in the 2010s, 2050s, and 2090s under SSP scenarios. (b) HDI-stratified exposure amplification: binned scatterplot (HDI intervals Δ = 0.1) showing normalized HCHO surface concentration changes (2090s vs. 2010s baseline) against development level; circle area scales with incremental exposure duration. (c) Temporal trends in formaldehyde exposure for high- and low-HDI regions, with projections of person-months exposed to concentrations above 3 ppb and 5 ppb thresholds. (d) Wildfire modulation effects: comparison between wildfire emission changes and pyrogenic emission control scenarios. Low-HDI tropical regions are disproportionately affected despite experiencing smaller overall warming. Future winter HCHO levels in these areas are projected to exceed today’s summer baselines, unlike the more uniform warming in high-HDI regions (Fig. S29). A nonlinear relationship emerges between development level and exposure burden, with the largest increases projected for regions with HDIs between 0.5 and 0.7. The exposure disparity between low- and high-HDI regions is projected to grow from 2.1-fold today to 2.4-fold under SSP5-8.5 by the 2090s (Figs. S26 and S27). Consequently, the committed cancer debt in Africa and Asia escalates sharply under warming, whereas high-HDI zones maintain relatively stable and low risk levels (Figs. 31 and S32). In contrast, SSP1-2.6 reduces the global exposure rate by 5 percentage points, cuts annual exposure duration by 36%, narrows the per-capita gap to within 10%, and protects an additional 1.91% of the global population from exceeding 5 ppb. To test the robustness of our results to the selection of exposure thresholds, we performed sensitivity analyses using alternative benchmarks ranging from 2 to 6 ppb. Both the absolute exposure burden and the share of exposure occurring in low-HDI regions were evaluated across thresholds (Fig. S41). The overall temporal trends and inequality patterns remained broadly consistent across thresholds. Wildfires play a pivotal role in HCHO exposure, particularly in high-latitude and fire-prone regions such as North America [[89], [90], [91], [92]]. In these regions, background HCHO concentrations are relatively low, and direct effects of projected temperature increases on HCHO are modest. In contrast, fire emissions have the potential to drive substantial spikes in HCHO levels. Under high-emission trajectories, projected wildfire escalation compounds exposure risks. While projected climate warming exacerbates fire risks, future trajectories are not solely determined by climate; human intervention remains a critical lever. Proactive wildfire management could cut HCHO exposure by 13% under SSP5-8.5 and 5% under SSP1-2.6 (Fig. 5d). These results highlight that while our baseline projections treat fire emissions as climate-sensitive forcings, the projected amplification is not inevitable and can be substantially mitigated through adaptive human interventions. Together, these underscore the dual urgency of climate mitigation: reducing greenhouse gas emissions not only protects the environment but also advances public health equity, especially among the most vulnerable populations. The potential 50% to 95% reduction in risk under the SSP1-2.6 scenario provides compelling evidence for the adoption of integrated climate policies that combine emission control, ecosystem management, and health equity strategies to address the formidable challenges posed by future global HCHO exposure. Limitation Our framework is based on empirical relationships between HCHO and its meteorological drivers derived from historical observations, which are assumed to remain broadly applicable under future climate conditions. Most projected surface temperatures remain within the historical variability range represented in the training data (Fig. S37), and model performance during independent extreme heatwave events further supports the applicability of these relationships under high-temperature conditions (Figs. S35−S36). However, potential shifts in ecosystem structure or chemical regimes under long-term warming may introduce additional uncertainties. Future climate change may introduce conditions beyond the historical range, particularly under high-emission scenarios. In addition, long-term warming may induce structural shifts in vegetation composition, wildfire regimes, and atmospheric chemical sensitivity (e.g., transitions between VOC- and NOx-limited regimes). Such changes could alter the relationships between HCHO and its meteorological drivers. Our results should therefore be interpreted as estimates of climate-driven amplification under present-day atmospheric relationships, while potential regime shifts remain an important source of uncertainty that warrants further investigation using process-based models. We further assessed whether model reconstruction uncertainty varied systematically across levels of socioeconomic development by stratifying prediction errors according to HDI categories during the independent test period. Model performance was broadly comparable across HDI groups (Table S17), with no clear indication of systematically larger errors in lower-HDI regions. Prediction errors were slightly lower in low-HDI regions (MAPE = 7.8%) than in very-high-HDI regions (MAPE = 14.1%). These results suggest that the inequality patterns identified in this study are unlikely to be driven solely by spatial differences in model performance. However, uncertainties may remain higher in regions with limited observational constraints, and the results should therefore be interpreted primarily in terms of relative large-scale patterns rather than exact local magnitudes. Finally, population density and anthropogenic emissions were intentionally held constant in this study in order to isolate the climate-driven amplification of HCHO exposure. This design enables a clearer quantification of the climate penalty on atmospheric formaldehyde exposure independent of socioeconomic and emission changes. However, actual future exposure risks will depend on the combined effects of climate change, population growth, urbanization, and emission trajectories. Our results therefore represent the climate-driven component of future exposure changes rather than a comprehensive projection of public health risk (Fig. S40). Conclusion Our study quantifies how ambient formaldehyde responds to a warming climate by linking long-term satellite observations with climate projections. The approach relies on several assumptions. It assumes that HCHO sensitivities to meteorological drivers remain broadly applicable under future climate conditions, uses monthly-aggregated climate data that smooth short-term extremes, and applies uniform exposure thresholds. These choices provide an observation-constrained, large-scale perspective on climate-driven amplification, but uncertainties remain in absolute concentrations, inter-model variability in key climate drivers, and local exposure estimates, especially in regions with sparse observations [38,93]. Despite these limitations, the spatially coherent trends and scenario comparisons offer robust insights into emerging environmental risks. The results show that climate change is imposing a measurable climate penalty on air quality, intensifying biogenic emissions and wildfire activity. HCHO exposure is projected to increase most strongly in the tropics and the Global South, where populations face year-round risk despite contributing the least to global greenhouse gas emissions. Rising HCHO also serves as a sensitive indicator of biospheric stress, reflecting large-scale responses of vegetation, fire regimes, and atmospheric chemistry to warming and drought [16]. While future chemical regime shifts may influence the magnitude of HCHO responses, the large-scale warming-driven signal identified here is expected to remain robust. Importantly, future exposure is not predetermined. Limiting warming through low-emission pathways can substantially reduce projected increases in hazardous HCHO, and proactive adaptation, including land and wildfire management, can further mitigate climate-driven amplification. Looking forward, reducing uncertainty and improving equity in exposure assessment will require expanded ground-based and in situ observational capacity, particularly in low-income and data-scarce regions where validation of satellite-derived HCHO remains limited [94]. Strengthening these measurement networks will be critical for refining risk estimates, supporting early warning systems, and informing effective climate and air quality policies. CRediT authorship contribution statement Jingkai Xue: Methodology, Investigation, Conceptualization, Writing – review & editing, Writing – original draft, Visualization. Qihou Hu: Supervision, Resources, Conceptualization, Funding acquisition, Writing – review & editing. Tiantian Li: Project administration, Conceptualization, Resources, Supervision. Qihua Li: Project administration, Conceptualization, Supervision. Wenjing Su: Methodology, Conceptualization, Validation. Hongri Hou: Methodology, Supervision. Ting Liu: Investigation, Resources, Formal analysis, Methodology. Yizhi Zhu: Visualization, Resources. Yizhou Xu: Methodology, Investigation. Xiaohan Wang: Methodology. Chengxin Zhang: Resources, Project administration, Methodology, Conceptualization. Peize Lin: Investigation, Data curation. Cheng Liu: Supervision, Resources, Project administration, Methodology, Investigation, Funding acquisition, Writing – review & editing. Data availability All input datasets are publicly available from their respective repositories (URLs and DOIs provided in the Text S4). Declaration of competing interest The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.