Prospective lifecycle assessment of carbon footprints. Carbon footprints (reported in g CO2 per vehiclekm) are estimated using a prospective lifecycle assessment framework based on GREET® 2  Version 201922 adapted for scenario analysis4. The framework enables estimations of future carbon footprints of passenger cars depending on climate change mitigation efforts in global manufacturing. Two scenarios for this mitigation are analyzed: Stated Policies and Sustainable Development; the results for Sustainable Development are presented in the main paper, with the results for the Stated Policies in the SM. The Stated Policies Scenario is based on currently implemented and stated climate policies by 2019 and the Sustainable Development Scenario is designed to limit global mean temperature increase to below 1.8°C, which is assumed to be in line with the Paris Agreement’s goals. The two scenarios are based on the IEA23 scenarios with the same names.
Carbon footprints per km are estimated based on
$$\text{C}\text{a}\text{r}\text{b}\text{o}\text{n} \text{f}\text{o}\text{o}\text{t}\text{p}\text{r}\text{i}\text{n}\text{t}\left({\text{t}}_{0},{\tau },\text{D}\right)=\frac{\text{V}\text{e}\text{h}\text{i}\text{c}\text{l}\text{e} \text{c}\text{y}\text{c}\text{l}\text{e}\left({\text{t}}_{0},{\tau }\right)+{\sum }_{\text{t}={\text{t}}_{0}}^{{\tau }}\left(\text{F}\text{u}\text{e}\text{l} \text{c}\text{y}\text{c}\text{l}\text{e}\left(\text{t},d\left(t,{\tau },D\right)\right)+\text{T}\text{a}\text{i}\text{l}\text{p}\text{i}\text{p}\text{e}\left(\text{t},d\left(t,{\tau },D\right)\right)\right)}{{\sum }_{\text{t}={\text{t}}_{0}}^{{\tau }}\text{d}\left(\text{t},{\tau },\text{D}\right)}$$
1
For different combinations of vehicle lifetimes (240 years), annual average driving intensities (5,000100,000 km per year), and the manufacturing year, t0, between 2020 and 2050. The average annual driving intensity, D, corresponds to the distance traveled over the whole lifetime of the vehicle, \({\tau }\). The annual driving distance, d, for year t, is assumed to decrease by b = 4.4% per year, following this equation:
$$\text{d}\left(\text{t},{\tau },\text{D}\right)=\frac{\text{D}\bullet {\tau }}{{\sum }_{t={t}_{0}}^{{\tau }}{\left(1b\right)}^{\text{t}{t}_{0}1}}{\left(1b\right)}^{t{t}_{0}1}$$
2
Vehicle cycle CO2 emissions are estimated based on manufacturing processes as implemented in GREET® for the Stated Policies Scenario, while new and innovative processes are phased in over time for the Sustainable Development Scenario based on a literature review4.
Fuel cycle as well as tailpipe CO2 are usephase emissions related to the distance traveled each year. The specific energy use of the car is determined depending on the type of car and its manufacturing year. ICEVs are assumed to use 669 Wh per km in 2020, decreasing to 492 Wh per km in 2030 and beyond, following energy efficiency improvements4. Similarly, BEVs are assumed to use 223 Wh per km in 2020, decreasing to 201 Wh per km in 2030 and beyond. ICEVs are assumed to use 100% fossil fuels acquired on average global markets, and BEVs are assumed to charge with electricity produced using average global, European, or Swedish technology mixes. Fuel cycle and tailpipe CO2 are estimated annually based on traveled distance, vehicle energy use, and appropriate carbon intensities.
The carbon intensity of electricity is based on estimates of average direct emissions for future electricity mixes of each respective geographic area, see description of sources for scenario data below. 2019 is used as a base year to avoid influence of the Covidpandemic on the carbon intensities. Upstream emissions occurring in production of fuels and power stations are accounted for by adding a weighted factor for future electricity mixes based on estimates by Pehl et al.24. We assume that Pehl et al.’s estimates of upstream emissions for each electricity generation technology can be applied regardless of geographic area and that their baseline and climate policy scenarios resemble the Stated Policies and Sustainable Development scenarios used in this study. Note that emissions for construction of water and nuclear power stations are assumed to be zero for Sweden and the European Union due to their long lifetime, the fact that they were mainly constructed several decades ago, and that few new stations are planned. Hence, we assume that the emissions from the construction of these stations are only attributed to electricity production prior to 2019. Continuing to account for these constructionrelated emissions in the carbon intensity of electricity after 2019 would not have any significant impact on the results.
For the global electricity mix used in manufacturing and for charging, future direct emissions and adjustments to account for transmission and distribution losses (based on the difference between estimated supply and demand) are based on estimates by the IEA23 for the two decarbonization pathways, Stated Policies and Sustainable Development. For the European electricity mix used for charging, direct emissions and adjustments to account for transmission and distribution losses are based on European Commission scenarios25 combined with the cap of the European Union emissions trading system reaching zero in 205826 for both decarbonization pathways. For the Swedish electricity mix used for charging, direct emissions for 2019 are calculated based on the total emissions for electricity generation divided by the enduse of electricity27,28. Direct emissions are assumed to decrease linearly to zero by 2045 for both decarbonization pathways, in line with the adopted netzero emission target and the Swedish government’s intention to reach zero for electricity generation29. Upstream emissions are based on estimates by Pehl et al.24 and projections for the future electricity generation mix by the IEA23, European Commission25, and Swedish Energy Agency30.
Swedish vehicle retirement statistics. Statistics on Swedish passenger cars retired between 2014 and 2018 are used to understand how changes in annual average driving intensity could influence vehicle lifetimes. The statistics are collected from the Swedish registry for road transport vehicles, regulated by Swedish law31. The excerpt, provided by the Swedish government agency Transport Analysis32, includes information on manufacturing year, date of registration, car manufacturer, engine type, mass in running order, total distance traveled at last inspection, date of last inspection, and date of deregistration. The excerpt only includes vehicles that were indeed retired at the date of deregistration. Hence, vehicles that were deregistered for administrative reasons or exported are excluded.
The cleaned dataset includes 365,575 observations. The cleaning performed by the authors aims to reduce bias in the results and applies the following criteria: (i) age or distance traveled must not be missing, equal to zero, or equal to 999,999, (ii) time between last inspection and date of deregistration must not be longer than 14 months, (iii) time between first registration of the vehicle and the manufacturing year must not be longer than one year, (iv) average distance traveled must not be greater than 400 km per day, (v) average distance traveled must not be less than 1 km per day, (vi) mass in running order must not be greater than 3,000 kg, and (vii) engine type is gasoline or diesel without hybridization. Details and rationale for these criteria are provided in SM 1.2.
Stratified random sampling is used to create a new dataset for analyzing the influence of increasing driving intensity since only a small share of total current vehicle retirement represents cars with high average annual driving intensity, such as taxis or other commercial vehicles. The strata and random sample size are set to maximize the amount of information about vehicles with high driving intensity while also ensuring high enough sample size to enable further statistical analysis. This results in strata for average annual driving intensity classes of 5,000 km/year increments from 0 km/year to 40,000 km/year, and three additional classes with larger increments (40,00150,000 km/year, 50,00170,000 km/year, and 70,001100,000 km/year) due to limited data availability. The random sample size in each stratum is 300 observations, except for the highest intensity class where the whole sample of 287 observations is used, see SM 1.3 and 1.4.
Semiempirical lifetimeintensity model. The semiempirical lifetimeintensity model enables estimations of vehicle lifetime probabilities for a given annual average driving intensity. The model should easily be updated with new parameters on average vehicle retirement lifetime, its standard deviation, and the average annual driving distance, as new statistics become available. The model should also easily be recalibrated based on new stratified random sampling datasets to enable use for other geographical regions. Two model designs are considered together with two assumptions on the probability distribution of the lifetime data as a result of these prerequisites.
If the data follow a Normal distribution, we assume that the probability of a vehicle manufactured at year t0, with average annual driving intensity D, being retired at year t is
In the elasticity design, we introduce a factor dependent on the quota between the driving intensity of the vehicle and the average annual driving intensity of current vehicle retirements, D0, as part of the mean,
$${\mu }\left(D\right)={{\tau }}_{0}{\left(\frac{D}{{D}_{0}}\right)}^{{\epsilon }}$$
4
that adjusts the expected vehicle lifetime of current retirements, \({{\tau }}_{0}\), dependent on the elasticity, \({\epsilon }\), that decides the level of influence of the driving intensity. An elasticity of 1 implies that the vehicle lifetime is fully determined by the driving intensity (e.g., if driving intensity is doubled, lifetime is halved), 0 indicates no influence and the lifetime is only determined by calendar age, while an elasticity above 0 would imply that the vehicle lifetime increases with driving intensity. This design benefits from easy interpretation, but it only applies for driving intensities equal to or greater than the current average.
The standard deviation,
$${\sigma }\left(\text{D}\right)={\alpha }{{\tau }}_{0}{\left(\frac{D}{{D}_{0}}\right)}^{{\epsilon }{\beta }}$$
5
is designed in a similar way to the design for the mean, where the constant \({\alpha }=\frac{{{\sigma }}_{0}}{{{\tau }}_{0}}\) is determined based on a fit of a Normal distribution to current vehicle retirement statistics. An additional elasticity, \({\beta }\), is introduced in the standard deviation to account for the distributions becoming increasingly narrow with higher driving intensity classes, see Figure 2.
In the logistic design, we instead assume that the distribution is governed by a function inspired by the logistic curve to better capture the form of the stratified random sampling. The logistic curve function is slightly altered to reduce the number of parameters to fit to the data. Hence, \({\mu }\left(D\right)\) and \({\sigma }\left(D\right)\) are defined as follows in this design.
\({\mu }\left(D\right)={L}_{0}\frac{L}{1+{e}^{\left(1D/{D}_{0}\right)}}\) and (6)
$${\sigma }\left(D\right)={\alpha }\left({L}_{0}\frac{L}{1+{e}^{\left(1D/{D}_{0}\right)}}\right)$$
7
where L and L0 are the parameters that would be calibrated based on the stratified random sampling. This design applies for all driving intensities greater than zero.
If the data are assumed to follow a Weibull distribution, we assume that the probability of a vehicle manufactured in year t0, with average annual driving intensity D, being retired at year t, is
where the scale, \({\lambda }\left(\text{D}\right)\), and shape, \(\text{k}\left(\text{D}\right)\), are defined in the same way as the mean, \({\mu }\left(D\right)\), and standard deviation, \({\sigma }\left(D\right)\), for the two model designs (see equations 47 above). Note that the average vehicle lifetime of current retirements, \({{\tau }}_{0}\), in this case represent the scale of current vehicle retirement statistics and that the constant, \({\alpha }=\frac{{k}_{0}}{{{\tau }}_{0}}\), is determined by fitting a Weibull distribution. The fact that the median is lower than the mean for higher driving intensity classes, see SM 2.2, indicates that the distribution is more positively skewed for higher driving intensity classes. This suggests that a Weibull distribution with a longer tail towards higher vehicle lifetimes would be a better fit, confirming previous research14,33.
The parameters for the different model designs are estimated using maximum likelihood estimation, see SM 1.4 and 2.3. A comparison of modeled vehicle lifetimes with the stratified random samples for different driving intensity classes is presented in Figure 5. The contour lines in Figure 5, also known as isodensity lines34, show how the points of equal probability density for a given vehicle lifetime shift depending on the assumed driving intensity (yaxis) and on the model design (panel and line type). The highest probability density level is shown around the mean of the distribution, and the distance indicates the rate of change, implying that a larger distance between the lines indicates a more spreadout distribution, analogously to on a topographic map.
The left panel clearly shows that the elasticity design deviates from the statistics at the average current driving intensity of 13,900 km/year and approaches an infinite lifetime as driving intensities decrease. The proposed correction of this issue is to use the logistic design, as demonstrated in the right panel. However, a limitation of the logistic design is that the distribution of vehicle lifetimes is assumed to be kept constant for driving intensities higher than the stratum with highest driving intensity (i.e., higher than 100,000 km/year in this study), see SM 2.4. The elasticity design instead results in vehicle lifetimes that approach zero for very high driving intensities. Regarding the choice of distribution, the Weibull distribution benefits from better reflecting the skewness of the statistics. However, it overcompensates for higher driving intensities when applied with the logistic design, resulting in longer tails of vehicle lifetimes than the statistics indicate, see the greater distance between lines in the right panel of Figure 5 and SM 2.4. This difference between Normal and Weibullbased model designs is close to negligible for the elasticity design. Benefits and drawbacks for the choice of distribution and for the model design are summarized in Table 1.
Table 1
Benefits and drawbacks with design different aspects of the semiempirical model
Design

Benefits

Drawbacks

Elasticity model

• Simple formulation
• Easy to interpret

• Applies to driving intensities equal to or greater than current average

Logistic model

• Applies for all driving intensities

• Less intuitive model design

Normal distribution

• Simple implementation

• Does not capture the skewness of the data

Weibull distribution

• Captures the skewness of the data and accounts for longer tails

• Shape parameter of Weibull more difficult to interpret
• May overestimate longer tails

Data availability
Data for all figures and additional data used in the analyses are available from the corresponding author upon request. Note that the detailed data on vehicle retirement are treated as confidential since data that could be traced back to individuals or companies are protection under the Swedish Public Access to Information and Secrecy Act (SFS 2009:400). Hence, requests for access to these detailed data should be made directed to the Swedish governmental agency Transport Analysis.
Code availability
The computer code used to generate the results reported in this study are available from the corresponding author upon request.