Supplement of Changes in anthropogenic precursor emissions drive shifts in the ozone seasonal cycle throughout the northern midlatitude troposphere

Abstract. Simulations by six Coupled Model Intercomparison Project Phase 6 (CMIP6) Earth system models indicate
that the seasonal cycle of baseline tropospheric ozone at northern
midlatitudes has been shifting since the mid-20th century. Beginning in
∼ 1940, the magnitude of the seasonal cycle increased by
∼10 ppb (measured from seasonal minimum to maximum), and the
seasonal maximum shifted to later in the year by about 3 weeks. This shift
maximized in the mid-1980s, followed by a reversal – the seasonal cycle
decreased in amplitude and the maximum shifted back to earlier in the year.
Similar changes are seen in measurements collected from the 1970s to the
present. The timing of the seasonal cycle changes is generally concurrent
with the rise and fall of anthropogenic emissions that followed
industrialization and the subsequent implementation of air quality emission
controls. A quantitative comparison of the temporal changes in the ozone
seasonal cycle at sites in both Europe and North America with the temporal
changes in ozone precursor emissions across the northern midlatitudes
found a high degree of similarity between these two temporal patterns. We
hypothesize that changing precursor emissions are responsible for the shift
in the ozone seasonal cycle; this is supported by the absence of such
seasonal shifts in southern midlatitudes where anthropogenic emissions are
much smaller. We also suggest a mechanism by which changing emissions drive
the changing seasonal cycle: increasing emissions of NOx allow
summertime photochemical production of ozone to become more important than
ozone transported from the stratosphere, and increasing volatile organic compounds (VOCs) lead to
progressively greater photochemical ozone production in the summer months,
thereby increasing the amplitude of the seasonal ozone cycle. Decreasing
emissions of both precursor classes then reverse these changes. The
quantitative parameter values that characterize the seasonal shifts provide
useful benchmarks for evaluating model simulations, both against
observations and between models.


1 Text S1. Power Series fits to long-term changes: Derivation of LTC(t) The role of LTC(t) in Equation 1 is to quantify the average long-term changes that underlie the monthly mean ozone time series. This function is subtracted from the time series to detrend the monthly means, which isolates seasonal variations from longer-term changes. The detrended monthly means are then used as input for the Fourier Transform analysis, which quantifies the seasonal cycle; the Fourier analysis is discussed in detail in section S2. 5 The physical basis underlying the derivation of LTC(t) is that the average atmospheric composition is continuous over time; across long time scales, the atmosphere does not undergo abrupt, discontinuous changes. Therefore, a power series, i.e., a least-squares regression fit to a polynomial, is an effective functional form to detrend the time series (as opposed to, e.g., a piecewise discontinuous function such as sequential straight line segments), because a power series can effectively approximate any continuous function. When implementing a polynomial fit, a decision must be made regarding the number 10 of polynomial terms to use in the fit. In principle, any function can be exactly fit with a power series of infinite terms; for our purposes it is important to choose a number of polynomial terms that effectively quantifies the long-term changes without over-fitting the time series or including statistically insignificant terms. Figure S1 shows an example monthly mean ozone time series with several LTC(t) functions derived from polynomial fits with between 2 and 12 terms. The polynomial parameters in these fits are statistically significant at the 95% confidence limit (i.e., the confidence interval does not include 15 0). Therefore, the LTC(t) can be approximated by a polynomial up to at least 11 th -order. Table S1 shows that all of these fits effectively detrend the time series, giving seasonal cycle parameters that agree within their statistical confidence limits, regardless of the number of polynomial terms. Therefore, fits with as few as 2 polynomial terms and as many as 12 polynomial terms can effectively detrend the monthly mean time series. This result shows that the long-term change and the seasonal cycle of ozone in monthly mean time series approximate orthogonal functions, mathematically speaking. 20 As is apparent in Figure S1, there is negligible difference between LTC(t) fits with 5 polynomial terms (indicated by the green trace) and 12 polynomial terms (indicated by the purple trace). There is a small, gradual improvement in the percentage of variance captured by the fit to Equation (5) and in the confidence limits of the derived parameters as the number of polynomial terms increases. In our analysis, we choose to use a 5-term, 4 th -order polynomial for LTC(t). Figure   S1 and Table S1 show that this choice, although arbitrary, returns nearly optimal confidence limits for seasonal cycle 25 parameters without the complexity of additional polynomial terms. Figure S2 shows a comparison of LTC(t) functions derived from six model simulations (shown in colors as indicated in the legend) and observations (shown in black) at Hohenpeissenberg, where the measurement record is longest. This comparison highlights the differences in long-term average ozone concentrations between model simulations, and between models and measurements. The models, except for GISS-E2-1-H, do not reproduce the strong curvature in long-term average 30 ozone concentrations in later years. However, three of the six models (CESM2-WACCM, GFDL-ESM4, and MRI-ESM2-0) do accurately simulate modern-day absolute ozone concentrations (within ± 15%) over the 1971-2015 period. Figure S1. Comparison of LTC(t) derived from fits of polynomials of orders ranging from 1 (a straight line) to 11 (a 12-term polynomial fit), which overlay the example time series that is fit. Table S1. Comparison of parameters of the fundamental term of Equation 4 derived from fits to the GFDL-ESM4 Jungfraujoch 5-6 km time series illustrated in Figure S1 with LTC(t) polynomial fits with varying number of terms. Table includes parameters defining both the preindustrial fundamental frequency (A1 and ɸ1) and the Gaussian function that describes the shift of the fundamental frequency (r, m, s, rφ, mφ, and sφ), as well as the percentage of the variance in the original time series captured by the 40 fit. We include one more significant figure than is statistically justified in the table entries to show the  First, the fundamental harmonic is typically larger than the second harmonic. Exceptions are the two lower-elevation North American sites, Lassen NP and Trinidad Head at 1 km, for which the first two harmonic frequencies are approximately equal in amplitude. It is important to note that when the seasonal cycle shifts take place, the fundamental amplitude grows so that it 55 is larger than the second harmonic at all sites, even Lassen NP and Trinidad Head at 1 km. A second and more important similarity in the Fourier analysis is that together the fundamental and second harmonic frequencies capture most of the variance associated with the seasonal cycle. Harmonics higher than the second consistently make only smaller contributions across all locations and models; for most cases, higher harmonics are not statistically significant at the 95% confidence interval. Hence, we consider only the fundamental and second harmonic as the constituents of the seasonal cycle in our analysis. 60 In the paper, we illustrate the preindustrial seasonal ozone cycle for the two free troposphere locations, as they are expected to be least influenced by continental influences, and thus most representative of the background troposphere. Here, we examine 70 the preindustrial seasonal cycle for the four other, lower elevation locations included in our analysis, which are more influenced by surface emissions and ozone loss. Figures S4 and S5 show plots similar to Figure 3 of the paper; they quantitatively examine the preindustrial seasonal cycle at the lower-elevation sites. There are two important takeaways from Figures S4 and S5. First, at the European sites (Jungfraujoch and Hohenpeissenberg), the preindustrial seasonal cycle is largely determined by the fundamental (which is 75 larger in magnitude than higher-order frequencies) and reaches a maximum in the spring or early summer for most models. In contrast, the seasonal cycle at the North American sites (Lassen NP and Trinidad Head at 1 km) is more varied between models.
Some models simulate a springtime ozone maximum and summertime minimum, and most models agree on a second harmonic with spring and autumn peaks, both of which are characteristics of ozone in the MBL (Parrish et al., 2016). Therefore, the lower-elevation North American sites are somewhat different in character from their European counterparts because they are 80 more impacted by MBL conditions due to their near coastal locations than are the European sites, which are at more nearly mid-continental locations. Second, despite apparent MBL influence on the North American sites, the overall preindustrial seasonal cycle at low-elevation sites in both Europe and North America (panels (c) and (f) Figure S6), a location more isolated from continental influences, with the continental locations. We take model results from the same two altitudes considered in the manuscript for Trinidad Head: one near the top of the MBL at ~1 km, and a second within the free troposphere between 5-6 km. We present this comparison because the coarse spatial resolution of the models give grid cells that typically include 10 4 km 2 ; therefore, the model cell containing Trinidad Head, located on the west coast of 100 California, includes both significant ocean and land areas. The offshore simulation results are expected to be more representative of the baseline troposphere without direct continental emissions. There are no measurements from this offshore location to compare with the model simulations.
Similar seasonal cycle shifts are seen in the free troposphere model simulations over the Pacific. The upper right graph in Figure S6 compares the phase of the fundamental frequency in the free troposphere throughout the simulation time period at 105 both locations; each of the six models simulates similar behaviour at the two sites. The amplitude of the fundamental in the free troposphere (upper left graph in Figure S6) also exhibits similar behaviour at both locations, although there are noticeable offsets in the magnitudes of the amplitudes. The similarity of the simulations at these two locations in the free troposphere again emphasizes that the shifts in the ozone seasonal cycle are common throughout the baseline troposphere at northern midlatitudes. 110 At 1 km altitude the model simulations of the fundamental of the seasonal cycle are strikingly different between the continental and marine locations. It is likely that this altitude at the coastal location of Trinidad Head is particularly challenging for models to accurately simulate, since it represents both horizontal and vertical transitionsfrom the MBL to the free troposphere and from the marine to the continental environments. The contrasting behaviour shown in the lower panel of Figure S6 emphasizes that the seasonal shift is driven by temporal changes in anthropogenic precursor emissions (see discuss 115 in Section 3.4 of the paper) , which are primarily located at the continental surfaces, but the resulting seasonal shifts are seen throughout the troposphere.

125
Text S5. Dominance of anthropogenic activity in emissions and seasonal cycles of emissions.
Anthropogenic activity is predominantly responsible for rising overall emissions in the latter half of the 20 th Century. We consider the combined anthropogenic and biomass burning emissions; these are the emissions that are common between all 130 models. While emissions from all sources influence ozone formation and can potentially cause shifts in the seasonal cycle, Figure S7 shows that biomass burning and natural emissions are approximately constant over time; it is anthropogenic emissions that substantially vary. This figure also emphasizes that anthropogenic activity makes a much larger contribution to overall emissions at northern midlatitudes than does biomass burning, especially after the mid-20 th Century.
There is seasonal variation in anthropogenic and biomass burning emissions, as shown by the plot of monthly mean 135 emissions in Figure S8 in amplitude; however, seasonal NOx variations are small (~ 5% of the total emissions) compared to the amplitude of the seasonal ozone cycle (which is between 20% and 40% of the ozone concentrations). It is not evident in Figure S8, but the phases of NOx and NMVOC emissions remain constant over the 1850-2014 period. Seasonal NMVOC emissions are at their highest in the summer (owing to summer wildfires), while seasonal NOx emissions peak in the winter, due to peak energy usage associated with heating buildings. We can find no close correlation between the seasonal cycles of the precursor 145 emissions and that of ozone. Based on this analysis, we conclude that the seasonality of emissions is not related to the changing seasonal ozone cycle.

Text S6. Model-specific natural emissions and methane
All models share the same emission inventories for anthropogenic and biomass burning emissions; these common emissions are detailed and analyzed in the paper. However, each model has unique parameterizations and physical processes that determine precursor emissions from natural sources, so natural emissions differ between models. Table S2 describes the natural  160 sources of ozone precursor emissions in four models; and the emission totals integrated from 30° to 60° N are shown in Figure   S9.
The VOC emissions are derived and apportioned between component species through different approaches, and are given in varying units. As a consequence, we cannot quantitatively compare either natural emissions with the anthropogenic and biomass burning emissions or the magnitude of natural VOC emissions between different models. Figure S9a shows natural 165 VOC emissions from four models. Notably, the units in which they were specified vary; natural VOC emissions from GFDL-ESM4 are in units of mass of carbon emitted, while emissions from other models are in units of total VOC mass emitted. For example, the GISS-E2-1-H model specifies most natural emissions (DMS, isoprene, etc.) in total mass of emitted species, but emissions of entire families of compounds (e.g., alkenes) are given as moles of compounds. Therefore, Figure S9a gives only a qualitative comparison of natural VOC emissions between models and with the anthropogenic and biomass burning precursor 170 emissions given in Figure 8 of the manuscript. Nevertheless, Figure S9a does indicate that natural VOC emissions are largely constant over time, which implies that anthropogenic emissions are primarily responsible for the seasonal cycle shifts. Figure S9b shows natural NOx emissions, which include lightning and soil emissions. Natural NOx emissions are far smaller in magnitude than the anthropogenic and biomass burning NOx emissions given in Figure 8 of the manuscript, especially after the mid 20 th Century. Figure S9b indicates that natural NOx emissions are nearly constant over time in contrast to the 175 anthropogenic and biomass burning emissions, which again implies that anthropogenic emissions are primarily responsible for the seasonal cycle shifts.
Methane is treated differently from the other VOCs in the model simulations due to its much longer lifetime than other VOCs and its relatively uniform concentration in the troposphere. Figure S9c gives the time series of methane concentrations specified for northern midlatitudes for all of the ESMs. Methane concentrations approximately doubled between 1850 and 180 2015, with the most rapid increases between 1950 and 1990. Methane changes may also influence the seasonal cycle shift, but we do not attempt to quantitatively analyze that impact.   18 ± 2 days 4.7 ± 0.2 ppb 31 ± 2 days 5.0 ± 0.2 ppb 52 ± 2 days 5.2 ± 0.2 ppb 1990 ± 2 1986 ± 1 1991 ± 1 1987 ± 1 1986 ± 1 1985 ± 1 29 ± 4 28 ± 1 35 ± 2 26 ± 1 47 ± 2 26 ± 1 GISS-E2-1-H Phase GISS-E2-1-H Amplitude MRI-ESM2-0 Phase a MRI-ESM2-0 Amplitude 11 ± 1 days 6.7 ± 0.2 ppb 29 ± 4 days 9.5 ± 0.2 ppb 1999 ± 3 1987 ± 1 1994 ± 4 1988 ± 1 32 ± 5 27 ± 1 54 ± 11 34 ± 1 UKESM1-0-LL Phase b 12 ± 3 days 2020 ± 23 49 ± 20 UKESM1-0-LL Amplitude 7.1 ± 0.1 ppb 1990 ± 1 40 ± 1 a Results from Hohenpeissenberg, the only data set with statistically significant phase parameters b Results from European free troposphere 5-6 km, the only data set with statistically significant phase parameters 43 ± 3 13 ± 2 32 ± 5 26 ± 3 47 ± 2 28 ± 2 28 ± 4 30 ± 2 59 ± 9 29 ± 3 51 ± 6 32 ± 1 a In the free troposphere, the amplitude decreased rather than increased b Includes results from two higher elevation locations; phase shift was not statistically significant at Trinidad Head 1 km location c Includes results from two lower elevation locations; phase shift was not statistically significant at Trinidad Head 5-6 km 205 location d Includes results from two lower elevation sites, amplitude shift was not statistically significant at Trinidad Head 5-6 km location    Figure S12 shows similar plots for two surface sites in the two 235 hemispheres -Cape Grim, Australia and Hohenpeissenberg, Germany. Similar to Figure 1, during the last ~75 years of the Hohenpeissenberg time series, annual maximum ozone moved from primarily the spring months of March and April (pink, purple, or blue) before 1900 to primarily June (blue-green and light green) by the 1980s; this seasonal shift was accompanied by a pronounced increase in the magnitude of the seasonal cycle. As discussed in the manuscript, these changes are more pronounced at continental boundary layer sites such as Hohenpeissenberg, compared to the free troposphere location in Figure  240 1. The gold curves in the graphs at the right illustrate the quantification of those shifts as discussed in the text. ln contrast, at Cape Grim in southern midlatitudes, annual maximum ozone occurred consistently in the austral spring month of September (red) with no clear shift in the seasonal maximum or increase in its amplitude. Similar results are found for all six ESM simulations of ozone at southern midlatitude sites. The green curves in the graphs at the right illustrate the quantification of seasonal cycle shifts at Cape Grim through linear fits to the amplitude and phase, also as discussed in the text. The blue curves 245 show the analysis of the available (1982-2018see Parrish et al., 2020) Cape Grim measurements; these results indicate only small shits in the measured seasonal cycle at Cape Grim.
These marked differences in the simulated and measured ozone seasonal cycle between these two sites in northern and southern midlatitudes are consistent with our hypothesis that the ozone seasonal cycle shift at northern midlatitudes is driven by changing anthropogenic ozone precursor emissions, because these emissions have increased only to a much smaller extent 250 at southern midlatitudes compared to southern midlatitudes.