The Chilean Meteorological Service (In Spanish, Dirección Meteorológica de Chile, DMC), under the auspices of the Global Atmospheric Watch Programme (GAW) of the World Meteorological Organization (WMO), has maintained the Tololo station (30.17° S, 70.80° W, 2154 m a.s.l.) in the premises of the Interamerican Southern Astronomical Observatory since late 1995. The site is located about 50 km east of the Chilean coast, where the fast-growing conurbation of La Serena- Coquimbo is located. Despite this urban expansion, there is no evidence of increased impact from local emissions at Tololo, as discussed in later sections. Details on urban population and infrastructure in the region are provided in Fig. S1 of the Supplement.

In addition to the urban areas in the surroundings of Tololo, there is a potential influence of the large urban areas to the south of it: Santiago, the capital city, and the conurbation of Valparaíso-Viña del Mar (Anet et al., 2017) as illustrated in Fig. 1. In this case we consider whole provinces as there are smaller urban areas functionally connected to Santiago and Valparaíso-Viña del Mar cities. While in 1992 the population of Santiago (Valparaíso-Viña del Mar) was 5 258 000 inhabitants (808 000 inhabitants), in 2024 it is estimated to be 8 421 000 inhabitants (1 104 000 inhabitants), corresponding to a 60 % (37 %) percentage growth rate. We show evidence that this influence is sporadic.

Over the years, this station's topography and atmospheric circulation have been described in some detail (Anet et al., 2017; Gallardo et al., 2000; Kalthoff et al., 2002; Rondanelli et al., 2002). The area surrounding Tololo is characterized by highly complex topography, with deep valleys and the Andes cordillera located just 30 km east of Cerro Tololo, reaching elevations up to 6 km a.s.l. Most of the time, Tololo is immersed in the free troposphere and influenced by the subsidence regime of the South Pacific high, which brings clear sky conditions – a factor that has contributed to the establishment of numerous astronomical observatories in the region. In winter, the subtropical jet stream (STJ) is located on average at 30° S. From time to time, cutoff lows and deep troughs from higher latitudes may reach the Tololo area inducing tropopause breaks and/or upward mixing of marine boundary layer air. Above 4 km a.s.l., large-scale westerly winds prevail, while northerly winds are observed in a band between 2 and 4 km a.s.l., which results from the blocking effect of the westerly flow by the Andes. In addition to these features, a radiatively driven circulation that is intensified in summer is present with up-slope (down-slope) winds during the afternoon (night/early morning) at Tololo.

Surface O₃ has been measured in Tololo since November 1995. The technique used is UV absorption as detailed in Anet et al. (2017). This station is part of GAW, and data used here (hourly averages between November 1995 and December 2023) were downloaded from the EBAS database (https://ebas.nilu.no/, last access: 1 August 2024), thus it is subject to quality standards and rigorous scrutiny.

The station was equipped with a new ozone monitor and a CO, CO₂, and CH₄ analyzer (Picarro Inc. G2401) by the Swiss Federal Laboratories for Materials Science and Technology (EMPA) in 2013. CO data are available until 2022 and CH₄ until 2023. Data were downloaded from https://gaw.kishou.go.jp/ (last access: 1 August 2024). Calibration procedures are detailed in the metadata.

Standard meteorological data – wind, temperature, pressure and solar radiation – are also measured at the station. The data for the period between 1995 and 2023 were retrieved from the data base of the Chilean Weather Service (https://climatologia.meteochile.gob.cl/application/informacion/fichaDeEstacion/300034, last access: 1 October 2024). As explained later, these data were used to correct the reanalysis data at 775 hPa. We did not use the observed data because there were significant data gaps, particularly regarding humidity.

We used ERA5 data (Hersbach et al., 2020) instead of the local time series for temperature, humidity, and winds, extracting values at 30.25° S, 70.75° W and 775 hPa, the grid point closest to Tololo's actual location. A bias was identified in the ERA5 time series when compared to available in situ measurements (see Fig. S2). To correct this, we applied a Quantile Delta Mapping bias correction to the ERA5 data for Tololo. The correction was performed separately for each month and hour: for example, to adjust values at 00:00 LT in January, all observational and ERA5 data for January at that hour, as well as data from 2 h before and after (a 5 h window), were used to calculate the adjustment, which was then applied to the January 00:00 LT data. This procedure was repeated for each hour and month, resulting in a specific correction for every month–hour combination.

We also used ERA5 data to identify synoptic conditions consistent with stratospheric intrusions or downward transport of upper tropospheric air. To this end, we utilized geopotential height and vertical velocity at 500 hPa to estimate horizontal anomaly composites, as well as potential vorticity and specific humidity between 900 and 200 hPa to estimate a vertical cross section (longitude-pressure) along 30.25° S. To calculate anomalies, firstly we estimated the climatological annual cycle between 1995–2023 based on daily data. Thereafter we applied a Fourier filter to eliminate high frequency fluctuations (sub seasonal). Then, we calculated the anomaly as the difference between the 24 h running mean of the time series and the smoothed climatology.

To assess how climate variability affects long-term trends at Tololo, we considered ENSO, QBO and MJO in our GAM analysis described later. For El Niño–La Niña we used the Multivariate ENSO Index version 2 (MEI). It corresponds to the bi-monthly mean of the principal component of the combined empirical orthogonal function of five observed variables over the tropical Pacific: sea level pressure, surface temperature, surface zonal wind, surface meridional wind, and outgoing longwave radiation. This index is provided by NOAA Physical Sciences Laboratory (https://www.psl.noaa.gov/enso/mei, last access: 30 March 2024).

For QBO, we used the monthly mean zonal wind measured by meteorological soundings in Singapore (01.22° N, 103.55° E) in 50 (QBO50) and 30 (QBO30) hPa. Monthly means are calculated from daily data in Singapore. Data were obtained from the National Aeronautics and Space Administration (NASA), at the Goddard Space Flight Center (GSFC) (https://acd-ext.gsfc.nasa.gov/Data_services/met/qbo/QBO_Singapore_Uvals_GSFC.txt, last access: 25 October 2024).

In the case of MJO, we used the Real-Time Multivariate (RMM) MJO daily index (Gottschalck et al., 2010; Wheeler and Hendon, 2004). It corresponds to the first two principal components of the combined empirical orthogonal functions between 15° N–15° S of anomalous 200 and 850-hPa zonal winds and outgoing longwave radiation. It was retrieved from the Bureau of Meteorology Australia (http://www.bom.gov.au/climate/mjo/, last access: 29 May 2024).

As indicated before, CH₄ has been recorded at Tololo for the period 2013-2023. To complete the time series to the period before 2013, we used data collected at Rapa Nui (Easter Island, 27.8° S, 109.8° W, 51 m a.s.l.), which is an ozone sounding monitoring station under GAW (Gallardo et al., 2016). In addition to ozone soundings, the National Oceanic and Atmospheric Administration (NOAA, https://gml.noaa.gov/dv/site/?stacode=EIC, last access: 29 May 2024) maintained, i.a., weekly CH₄ flask measurements on the island between 1994 and 2019. Firstly, we compared both time series over the period 2013–2019, and we found slightly higher values at Tololo (see methane at Rapa Nui and Tololo in Fig. S3). As the differences in methane between both sites were deemed minor – possibly due to oceanic upwelling near Tololo (Weber et al., 2019) – and given the long turn-over time of CH₄, we decided to extend Rapa Nui CH₄ by means of a simple linear regression of simultaneous measurements. We took Tololo measurements at 22:00 UTC on the same days as they are available at Rapa Nui. Once the regression had been performed, we used it to extend the CH₄ time series for Rapa Nui until 2023 and used them for Tololo. These data were then used in our GAM to assess the influence of methane on the long-term trend of ozone measured at Tololo. Lastly, we estimated a smooth time series (with annual seasonality) and a smooth trend (without seasonality) to represent the new time series for CH₄ according to (Thoning et al., 1989). We used the code available at NOAA (https://gml.noaa.gov/aftp/user/thoning/ccgcrv/, last access: 15 October 2024).

We use outputs from a global atmospheric chemistry model to estimate the contribution of biomass burning at Tololo. The model is TM4-ECPL as described in Daskalakis et al. (2022). The model has a horizontal resolution of 3° longitude × 2° latitude, with 34 hybrid vertical layers up to 10 hPa. We use assimilated meteorology from ERA interim. The model was run over the period 1980 to 2015, and we use the period 1995–2015. Differently from Daskalakis et al. (2022) the model has been re-run using updated emissions as used in AR6, i.e., emissions as described in Hoesly et al. (2018) and van Marle et al. (2017). The model was run with and without biomass burning to estimate the contribution of this source of precursors to ozone.

Despite its resolution, the model captures the seasonal and, partly the day-to-day, possibly synoptical in origin, variability as inferred from the comparison with daily mean observations of CO. However, there is a positive model bias (∼ 7 ppbv), and a larger distance between model and observations of CO in the upper tail of the data distribution (see Fig. S4). Hence, we performed a bias correction of the model data. To do so, we followed a similar approach to that used for ERA 5 data, and as also applied by Staehle et al. (2024) and implemented by Schwertfeger et al. (2023). Thereafter, we applied the bias correction to the rest of the model data set. We assumed that the adjusted model CO kept the same ratio with biomass burning CO as the original data, which allowed us to also adjust the biomass signal. The characterization regarding the influence of biomass burning is restricted to the period between 1995 and 2015 for which we have TM4-ECPL data.

Air of stratospheric or upper tropospheric origin, hereafter SUTO, is characterized by relatively high O₃ and low water vapor mixing ratios, thus we used the hourly time series of O₃ and specific humidity at Tololo to identify SUTO events. Firstly, we calculated enhancements of O₃ over a 10 d running mean of O₃ and a simultaneous reduction in specific humidity over a 10 d period similarly to Cui et al. (2009). Differently from Cui et al. (2009), we also averaged specific humidity and did not use relative humidity but specific humidity. After careful visual inspection of the data, we defined the arrival of SUTO as a continuous period where simultaneously hourly O₃ was 10 % above the 10 d running average, and 24 running average of specific humidity was less than 15 % of the 10 d running mean. The minimum duration of an event was set to be 18 h to exclude fast passing synoptic and sub synoptic perturbations. To avoid identification problems due to missing data, high frequency variability or errors associated with using specific humidity from ERA5, we relaxed the continuity criterion. Once an event was identified, we allowed up to a maximum of 18 h after the event in which the above criteria might not be met in all hours and still be considered part of the same event, reducing the total number of those. We recognize that many criteria may be established to identify SUTO and that the chosen thresholds are somewhat arbitrary.. Also, as TM4-ECPL has a dedicated stratospheric ozone tracer that was used to estimate the ozone of stratospheric origin arriving in Tololo, we were able to get an independent estimate of the stratospheric contribution to ozone at Tololo.

The term trend is not uniquely defined in statistics (Capparelli et al., 2013), and calculated trends are method dependent (Franzke, 2012). There are multiple statistical approaches to estimating trends in time series and they must be carefully chosen and interpreted (Chang et al., 2021, 2023; Cooper et al., 2014, 2020; Gaudel et al., 2018), particularly when dealing with nonlinear and nonstationary data, with marked seasonality and interannual and decadal variability such as surface ozone at Tololo. As the climate system is complex, many processes interact with each other at multiple temporal and spatial scales leading to nonlinear responses to external and internal changes (Snyder et al., 2011; Wang et al., 2023). Intrinsic climate variability plays an important role when assessing trends at the local and regional scales (Franzke, 2012). According to the same author, intrinsic climate variability leads to a so-called “stochastic trend” as it is expressed in autocorrelation (variables show temporal correlation), whereas a so-called “deterministic trend” emerges due to an external forcing. In general, detecting trends depends upon the size of the trend, the time span of available data, and the magnitude of variability and autocorrelation of the noise in the data (Weatherhead et al., 1998). In the case of atmospheric chemistry variables, one must add uncertainties derived from instrument detection level and sampling, change of sensors, the influence of extreme events, etc. (Chang et al., 2021).

The Tropospheric Ozone Assessment Report (TOAR) initiative has made a substantial contribution to standardizing the methods to visualize and calculate trends for data as those collected at Tololo. Hence, we use the Quantile Regression algorithm that allows piecewise detection of change points by Chang et al. (2023). Again, the concept of change point is not well-defined in statistics, nonetheless it is a useful indicator of significant changes in a time series over time, for instance change of instruments. In the case of Tololo, the ozone sensor was replaced in 2013 (Anet et al., 2017). We will assess whether this change is apparent in the data.

Generalized Additive Models (GAMs) are an extension of linear models, allowing nonlinear and nonparametric fittings of complex dependences of response variables on explanatory variables. A GAM adopts a sum of arbitrary functions of variables – potentially nonlinear – that represent different features via splines, which altogether describe the magnitude and variability of the response variables (Hastie and Tibshirani, 1986; Molnar, 2025). The choice of explanatory variables is key. It must be based on physical reasoning about potential contribution of variables that one can logically expect to explain the response variable. Statistically, one should choose variables that contribute to explaining the response variable but are largely independent from one another (Kovács, 2024). Redundancy results in unstable parameter estimates in GAMs and makes the marginal effect of features harder to interpret. In the case of atmospheric variables, one cannot assure orthogonality among explanatory variables as they are generally autocorrelated, but an effort must be made to avoid redundancy. In this study, we applied expert knowledge, and partial dependence plots (Hastie et al., 2009) to assess the physical meaning and functional dependences of the estimated relationships and thereby interpreting the GAM results. Additionally, we used the SHapley Additive exPlanations (SHAP) approach – based on game theory – to determine the ranked contribution of each explanatory variable in GAM (Lipovetsky and Conklin, 2001). In a simplified manner, a GAM is of the form (Eq. 1): 1O3=ϵ+f1v1+f2v2+…+fNvN where ϵ represents an error term, and fi, i=1, N are spline functions of variables vi, i=1, N such as specific humidity, temperature, wind speed, etc. (Hastie and Tibshirani, 1986; Molnar, 2025). The error term has a stochastic component, so we run our GAM 50 times. We utilized the following Python software: pyGAM available https://pygam.readthedocs.io/en/latest/ (last access: 15 June 2025) and shap available at https://shap.readthedocs.io/en/latest/ (last access: 15 June 2025).

Seguel et al. (2024) analyze the ozone time series at Tololo over the period (1995–2021). They identify two change points in ozone, one in early 2006 and another in early 2014, both within uncertainties of several years. As the authors note, the 2006 change point coincides with the global increase in methane, also observed at Rapa Nui (Easter Island). In this work, we extend the analysis to include 2022 and 2023 and explore the role of several atmospheric variables. Thus, we repeat the change point analysis for ozone as well as for methane, temperature (T), specific humidity (q), dewpoint temperature (Td), and geopotential height at 500 hPa (Z500).

Methane was chosen due to its role as O₃ precursor in the background atmosphere (Crutzen, 1988; Crutzen et al., 1999). We included T because, in addition to typically explaining a significant part of the variance of near surface O₃, key reactions leading to O₃ production depend on temperature (Pusede et al., 2015). Specific humidity was selected since the principal global sink of tropospheric O₃ is the photolysis by near ultraviolet radiation (λ < 310 nm) in the presence of water vapor (Jacob and Winner, 2009). We chose specific humidity instead of relative humidity to separate the effect of temperature from that of humidity. Dewpoint has been used as a meteorological explaining factor at Mauna Loa, allowing the separation between dry air associated with higher altitude and latitude air masses, and moist conditions associated with more tropical air masses, when applied to nighttime values avoiding upslope data (Gaudel et al., 2018). At Tololo, higher latitude air masses can be moist as they reach this subtropical site (30° S) in connection with deep troughs and cutoff lows (Fuenzalida et al., 2005; Rondanelli et al., 2002), while subtropical air masses are typically dry. Geopotential height at 500 hPa is used as an indicator of synoptic scale variability. These times series and their change points are illustrated in Figs. 2 and 3.

Figure 2 shows the strong seasonality in O₃, CH₄ and CO. In addition to a marked seasonality, the most prominent characteristic of ozone is the maximum values that are observed in spring (Anet et al., 2017; Gallardo et al., 2000). Carbon monoxide also peaks in spring, except for some summer events. This maximum in ozone and CO is generally attributed to biomass burning (Anet et al., 2017; Daskalakis et al., 2022). But as we will see later, many SUTO events occur in connection with synoptic conditions consistent with STE, which may also play a role in late winter and spring for ozone (Daskalakis et al., 2022; Rondanelli et al., 2002). The variability in methane, on the other hand, is less marked with a February–March minimum, and an August–September maximum, which is largely driven by the hydroxyl radical sink (East et al., 2024). ENSO also plays a role in methane interannual variability (Rowlinson et al., 2019), which will affect its seasonal variation too. The influence of ENSO, QBO, MJO has been found to be relevant for the variability of tropospheric ozone and its precursors (Daskalakis et al., 2022 and references therein).

Except for methane, change points calculated for all species are subject to multi-year uncertainties, which is to be expected given the relatively large variability of the time series (Muggeo, 2017), including long-term modes of variability (e.g., Pacific Decadal Oscillation, PDO) that may not be captured by the relatively short times series. The change point for ozone is estimated to occur near August 2005 with an uncertainty range between 2001 and mid-2010, which largely coincides with the upward acceleration of methane growth by mid-2006. The addition of the years 2022 and 2023 resulted in a sole change point near 2006, instead of two as shown by Seguel et al. (2024). Interestingly, ozone shows a positive median trend of the both before and after 2006, the rate of growth being larger after 2006 than before 2006. The near 2006 change in ozone coincides with a marked increase in the rate of change in methane, which is suggestive but difficult to reconcile with the long turnover time of methane, and the absence of enough NO_x in the region of interest, except perhaps in connection with long-range transport of biomass burning. The 5th and 95th percentiles of O₃ indicate decreasing trends before 2006 and growing trends thereafter suggesting a more significant influence of extreme events in the later period such as summer wildfires along central and Southern Chile or during the COrona VIrus Disease pandemic (Seguel et al., 2024). Over the period 2006–2023, ozone (median) shows a trend of about 2.1 ppbv per decade. Thus, over 17 years, the corresponding change in ozone is ca. 3.6 ppbv. Assuming this is well mixed, the corresponding change in ozone mass burden is 30 Tg O₃ or a yearly average of 1.8 Tg O₃. Over the same period methane volume mixing ratios have changed by 145 ppbv, which corresponds to a change in mass burden of 408 Tg CH₄. Thus, the mass burden ratio between ozone and methane is ≈ 0.08 Tg O₃ / Tg CH₄, which is in lower end but consistent with previous global model estimates, i.e., a range between 0.12 and 0.16 (Fiore et al., 2008) and 0.09 to 0.19 (West et al., 2007).

In Fig. 2, CO shows two points of change around 2009 and 2014, which, we hypothesize may be linked to an increased frequency and extent of fires in Central and Southern Chile (Carrasco-Escaff et al., 2024; González et al., 2018). However, we cannot rule out that the use of a combination of model and instrumental data may yield spurious results. Our fire hypothesis is based on the occurrence of large fires whose emissions of CO may reach Tololo, like in the summer of 2017 (Lapere et al., 2021).

Methane shows three change points around 2000, 2006 and 2016 (cf. Fig. 2). In late 1999 a plateau in CH₄ mixing ratios started, following a period of steady increase and that remained for a few years until late 2006. This was observed around the world but the reasons that explain these changes are still subject to debate due to the complex interactions between emissions and chemistry that drive atmospheric methane (Saunois et al., 2020). We identify a trend change in 2016 which coincides with observed but still not fully explained changes in CH₄ growth rate (Nisbet et al., 2019).

A strong seasonality in all meteorological variables is apparent in Fig. 3. The strong contrast between summer and winter values is characteristic of the subtropics (Garreaud et al., 2009). Except temperature that shows a change point around 2010 but with a broad uncertainty span, none of the variables shown in Fig. 3 show clear changes in trends. Nevertheless, Central and Southern Chile experienced a continuous drought from 2010 to 2022 (Garreaud et al., 2020), which was interrupted by significant precipitation events over the latest winters. The changes in temperature and humidity observed at Tololo also reflect these conditions. This is in fact consistent with the warming and drying trends driven by climate change over Central and Southern Chile (Bozkurt et al., 2019).

As methane is the variable with the most noticeable upward trend, we assess the changes in hourly ozone distribution over time when methane changes. To this end, we consider four periods as inferred by the change points of methane. This is shown in Fig. S5. Until 2006, the mean and median of ozone only show a slight difference, while after 2006 these statistical indicators clearly increase. Moreover, the growth in the indicators appears to accelerate further after 2016. This suggests that methane is a major driver of ozone changes at Tololo.

Next, we provide an update of the trend estimate presented by Seguel et al. (2024), including the determination of change points, all according to TOAR II recommendations (Chang et al., 2023). This is shown in Table 1 and Fig. 4. Table 1 shows that O₃ trends were negative or slightly positive before August 2005 for all percentiles, albeit with low reliability as defined by TOAR II. After the change point, a clear upward trend emerges for all percentiles (very high certainty). As previously indicated, trends at 5th and 95th percentiles are somewhat higher than the median trend, which may be indicative of the impact of extreme events, such as fires.

We identified events of air of stratospheric and upper tropospheric origin (SUTO) as those characterized by high ozone and low humidity, possibly linked with stratospheric intrusions or downward transport of upper troposphere air as described in Sect. 2.6. Using said method, we found 252 SUTO events over the period 1995–2023, of which 94 (37 %) last 24 h or less, 103 (40 %) last between 24 and 48 h, 35 (14 %) last between 48 and 72 h, and 20 (8 %) last more than 72 h (cf. Fig. 5). Median anomalies in ozone are around 5 ppbv and -2 g kg⁻¹ in water vapor, which are significant magnitudes compared with typical ozone and water vapor values at Tololo.

While SUTO events can occur any month of the year, most of them take place during the cold season and early spring, between May and October. One can observe a distinct interannual variability in the number of events per year. Typically, there are fewer (more) events in connection with the cold (warm) phase of ENSO, which is consistent with a stronger (weaker) South Pacific High that readily hinders (allows) the arrival of mid-latitude synoptic disturbances such as cutoff lows and deep troughs. There is no trend in the number of events per year nor in their duration. Furthermore, the number and duration of intrusion events do not appear to be related to the intensity of the ENSO anomaly (not shown).

At first sight, the fact that there are more SUTO events in El Niño years might seem contradictory with the fact that the spring maximum in ozone at Tololo is typically larger during La Niña years rather than during El Niño years as firstly stated by Anet et al. (2017), and reproduced here (see Fig. S6). However, one must recall that several processes are at play influencing ozone in Tololo. It is worth noticing that the model also captures the difference in ozone seasonality during El Niño and La Niña years, however the magnitude is overestimated (see Fig. S7).

In addition to counting SUTO events, we characterized their synoptic scale evolution. Figures 6 and 7 show composite anomalies in synoptic meteorological fields over a 12 d period that starts 10 d prior and ends 2 d after the event. In the horizontal we show composite anomalies in geopotential height at 500 hPa, and vertical velocity (Fig. 6). These fields show a slow passing deep trough of north-west to south-east orientation that reaches north of 30° S 10 d prior and shows a full mature stage by day -4. While the trough is approaching the Chilean coast, one observes uplifting (subsidence) west (east) of the Andes, which is consistent with the behavior of deep troughs and cutoff lows (Rondanelli, 2025). Thereafter, a subsidence zone is evident and advances eastward from day -2 to day 0, while a ridge develops and the subtropical high-pressure system is re-established. Thus, there is subsidence upwind and passing over Tololo with vertical velocity anomalies of more than 0.15 Pa s⁻¹, i.e., producing favorable conditions for stratospheric or upper tropospheric ozone mixing down to Tololo. Furthermore, such configuration is consistent with the occurrence of stratospheric intrusions upwind of Tololo. Nevertheless, the mere synoptic configuration does not assure the occurrence of STE.

It is well known that stratospheric air in the Southern Hemisphere is characterized by negative vorticity, and low humidity. Therefore, in Fig. 7, we show composite anomalies of potential vorticity and specific humidity of SUTO events. Consistently with the synoptic patterns discussed in Fig. 6 but occurring ca. 48 h prior, the development of the deep trough is accompanied by the entrance of negative anomalies of potential vorticity and humidity that propagate downwards, starting to reach Tololo already by day -2, i.e., prior to the occurrence of positive ozone and negative humidity anomalies. This is consistent with the duration of the events that typically last less than 3 d (> 90 % of the cases). Thereafter, the intruding air follows the westerlies with continued negative anomalies in potential vorticity and humidity east of the Andes.

Hence, based on the observed data, synoptic conditions corresponding to the average of 252 cases show passing troughs that reach the subtropics, which in turn are consistent with the potential occurrence of stratospheric intrusions. Said conditions lead to increases in ozone of around 5 ppbv on average at Tololo., accompanied by negative anomalies in humidity of around -2 g kg⁻¹, which in turn compare with typical ozone and water vapor mixing ratios of 30 ppbv and 7 g kg⁻¹, corresponding to 17 % and 28 % respectively. Such events occur mainly in the cold season and early spring and are favored by El Niño conditions. SUTOs typically occur in connection with deep troughs and cutoff lows that connect the subtropics with higher latitudes. On average, one finds 9 events per year but within a range between 5 and 13 events per year. The maximum ozone anomaly observed in our results corresponds to 15 ppbv, this is low compared with values observed in the Northern Hemisphere. We attribute this to the generally stronger STE observed in the Northern Hemisphere compared to the Southern Hemisphere (e.g., Holton, 1990).

As indicated in the precedent paragraph, Figs. 5 and 7 show a composite of 252 cases reflect that most of the cases correspond to passing troughs. However, the average also includes other synoptic configurations. To better grasp this diversity of cases, we performed a clustering analysis (k-means) to identify underlying structures, i.e., four groups of different synoptic configurations, all leading to SUTO, and some clearly consistent with stratospheric intrusions occurring at different locations. These clusters and the corresponding configurations are shown in Figs. S8 to S15. Three out of the four groups show approaching or passing deep troughs or cutoff lows, and one denotes the presence of a zonal flow. Deep troughs show enhanced subsidence behind the low-pressure system, some of which might be accompanied by a tropopause folding. In fact, part of the back trajectories for the representative cases (maximum ozone anomalies) of the three groups showing trough configurations, clearly indicate stratospheric intrusions, i.e., crossing of PV = -2 units (see Figs. S16 to S20). Interestingly, the highest ozone anomaly (≈ 15 ppbv) event takes place in connection with a defined jet-stream, which in addition to transporting upper troposphere air eastwards, it trickles down stratospheric air possibly by marked turbulence associated with the jet-stream.

The TM4-ECPL model provides an estimate of the stratospheric influence on Tololo ozone (Fig. 8). In these simulations, the ozone of stratospheric origin reaches Tololo all year around but in winter and early spring (June through September) this contribution surpasses 10 ppbv (on average), and in summer (December through February) it is roughly half of that. This corresponds to a percentage contribution of roughly 15 % in summer to 25 % in winter. It is worth noting that the seasonality of the stratospheric contribution found in the simulations is largely consistent with the seasonality found for SUTO events and shown in Fig. 5, i.e., the model captures the overall seasonal variability of ozone with higher values in winter and early spring than in summer and early fall. However, an overestimate of ozone at Tololo by nearly a factor of two is evident. Moreover, there are many days with stratospheric contributions that reach or surpass 50 % of the estimated ozone in winter (cf. Fig. 8). We suspect that part of the mismatch between model and observations is due to too strong a stratospheric contribution. In its current version, the upper boundary condition for O₃ is estimated by nudging O₃ concentrations above 50 hPa altitude to satellite observations with no explicit stratospheric chemistry. This was already reported when evaluating TM4-ECPL against ozone soundings at Rapa Nui, where an overestimate of ozone in the upper troposphere was found (Daskalakis et al., 2022).

In Fig. 9 we show the number of days of stratospheric intrusions and their contribution to ozone as calculated by TM4-ECPL for the period 1995–2015, categorized by season. Also, we indicate the number of events that coincide with those detected by our empirical methodology. The total number of days estimated by TM4-ECPL is typically an order of magnitude higher than the number of days of SUTO events detected, particularly in summer, which again suggests too strong a stratospheric contribution all year. Also, TM4-ECPL estimates many days with stratospheric contributions of less than 5 ppbv, whereas the concurrently observed SUTO events do not. The largest stratospheric contributions estimated by TM4-ECPL reach 40 ppbv or 35 ppbv when considering only days when we observed SUTO events, which is in any way much larger than the ozone anomalies based on the empirical method that reaches up to 15 ppbv. Lastly, the median stratospheric contributions calculated with TM4-ECPL are around 15 ppbv d⁻¹ but that of the empirical method is only 5 ppbv.

As shown in Fig. S4, CO at Tololo was overestimated by the model simulation by on average about 7 ppbv over the period when we count with both observations and simulations (2013–2015). We corrected this by applying a bias correction method (Cannon et al., 2015; Staehle et al., 2024) to reconstruct the whole model series (1995–2015). The contribution of biomass burning to CO and ozone was calculated as the difference between the runs with and without biomass burning as shown in Fig. 10. The role played by biomass burning in terms of ozone production and transport over the east Pacific – particularly during spring – has been shown in earlier work (e.g., Daskalakis et al., 2022). This influence is also clear in Tololo where an increase in CO during spring is both simulated and observed (see Fig. S2). While observations of CO do not allow attributing the biomass signal, simulations do. According to the TM4-ECPL model outputs, i.e., with all sources and without biomass burning, the seasonally averaged contribution to ambient CO at Tololo reaches up to nearly 23 % in October. The minimum contribution occurs in April when the average is about 5 %. Notice that in summer (DJF), averaged values of biomass burning contributions are lower than in spring, but one observes a secondary maximum in connection with regionally occurring fires that have become more common over central and southern Chile (e.g., Lapere et al., 2021). Regarding ozone, the biomass contribution peaks also in October, but it only reaches a median value of about 15 % of total ozone (∼ 35 ppbv).

After trying several combinations of potential explanatory variables, we chose 14 as indicated in Table 2. This set of variables were chosen based on their physical meaning and relevance for explaining ozone and trying to identify largely independent variables. Thus, our GAM has the general expression (Eq. 2): 2O3=ϵ+f1T+f2q+f3v+f4Wdir+f5BLH+f6STEduration+f7ENSO+f8QBO+f9MJO+f10ω+f11CH4+f12CO+f13DofW+f14DofY where ε represents an error term, and fi, i=1, 14 are spline functions, in this case with 9 nodes that represent the dependences – potentially nonlinear – on different variables detailed in Table 2.

The model (GAM) aims at capturing daily average values of ozone. As seen in Fig. 11, the model represents most features of the time series, including day-to-day, seasonal and interannual changes. It captures 76 % of the data variability but it tends to underestimate extreme values. Still, on average, the error is less than 3.6 ppbv. Also, the model captures the trend of the observed data, i.e., 1.0 ppbv per decade, using the Ensemble Empirical Mode Decomposition (EEMD) as described in Anet et al. (2017). This trend is 30 % larger than the one calculated by Anet et al. (2017), which is in line with increasing ozone mixing ratios at Tololo. When we run GAM without considering the influence of methane, GAM estimates a trend for the 1996–2023 period of only 0.4 ppbv per decade, highlighting the significant role of methane in explaining the observed upward trend in ozone at Tololo. Without methane both variance (74 %) and error (3.7) show a slightly worse performance.

To assess the relative importance of the different variables we used two techniques. First, we calculated the partial dependences of the GAM reconstructed ozone with respect to each variable. This is shown in Fig. 12, and discussed here:

The day of the year (seasonality) contributes the most to ozone with up to 8 ppbv in spring, and secondarily with ca. 3 ppbv in winter. Seasonality also contributes slightly (< 2 ppbv) but negatively to ozone at Tololo in fall. In summer, a small increase in ozone appears. This seasonal variability is fully consistent with the processes we have previously described: SUTO events in winter and spring, biomass burning in spring, episodical fires over central and southern Chile in summer, etc.

Notice that all 50 runs of GAM showed similar functional dependences of the variables (see Fig. 11) except for different error distributions among different variables. Thus, while the quantitative functional dependences may vary among runs, there is an overall physical consistency in those.

The previous paragraphs describe the partial dependences found in GAM. Now we present the results of the SHAP method (see Fig. S20). Again, SHAP ranks seasonality, humidity, temperature, methane, carbon monoxide, and boundary layer height as the most important variables, following the physical reasoning described earlier. The same applies for the duration index. However, according to SHAP, El Niño years have mostly a negative influence on ozone, while La Niña years have a slight positive impact, which differs somewhat from the much smoother ENSO functional dependence shown in Fig. 12. All other remaining variables are of less importance in SHAP. All in all, despite minor differences, the partial dependences of the GAM and the SHAP approach result in similar results in terms of the relative importance and role of the independent variables affecting ozone. This is encouraging as these methods have different mathematical grounds.