Modifying emission scenario projections to account for the effects of COVID-19: protocol for Covid-MIP

Lockdowns to avoid the spread of COVID-19 have created an unprecedented reduction in human emissions. While the country-level scale of emissions changes can be estimated in near-real-time, the more detailed, gridded emissions estimates that are required to run General Circulation Models (GCM) of the climate will take longer to collect. In this paper we use recorded and projected country-and-sector activity levels to modify gridded predictions from the MESSAGE-GLOBIOM SSP2-4.5 scenario. We provide updated projections for concentrations of greenhouse gases, emissions fields for aerosols and 5 precursors, and the ozone and optical properties that result from this. The codebase to perform similar modifications to other scenarios is also provided. We outline the means by which these results may be used in a model intercomparison project (CovidMIP) to investigate the impact of national lockdown measures on climate. This includes three strands: an assessment of short-term effects (5-year period), of longer-term effects (30 years) and an investigation into the separate effects of changes in emissions of greenhouse gases and 10 aerosols. This last strand supports possible attribution of observed changes in the climate system, hence these simulations will also form part of the Detection and Attribution Model Intercomparison Project (DAMIP).

Societal lockdown measures to contain the spread of COVID-19 have resulted in unprecendented global changes to the emissions of greenhouse gases (GHGs) and aerosols (Le Quéré et al. (2020a); Venter et al. (2020); Forster et al. (2020)). There are reports of a 36% reduction in population-averaged global NO 2 concentrations (Venter et al. (2020)) for 34 countries prior to the 15th of May, and CO 2 emissions are expected to fall by 4-8% in 2020 (Le Quéré et al. (2020b); IEA (2020); Liu et al. (2020)). Shorter-duration and localised changes have been even more extreme (Bauwens et al. (2020); Yang et al. (2020)), 25 but show nonlinear changes in air chemistry that simple, globally averaged climate models will miss (Le et al. (2020)). It is therefore desirable to explore the impact of these changes on climate change projections, both to establish to what extent simulations ignoring the effects so far need updating due to short-term changes and to investigate potential impacts of the lockdown in the long term. This is challenging because country-level emissions estimates are often generated only on a yearly basis, missing the variations between months or weeks. Moreover, detailed climate simulations require emission statistics to 30 be broken down on a higher resolution uniform grid, and these are typically only estimated several years after the emissions have occurred Feng et al. (2020); Meinshausen et al. (2020). This paper demonstrates the use of near-simultaneous "nowcasting" from open-access data on mobility, energy grids and aviation to modify pre-existing predictions on a country-and sector-specific grid. By expressing our scenario as a modification of a pre-existing scenario that diverges only at the point of interest, we have an estimation of sector emissions on a grid that 35 simulation teams know how to handle. We can also use the pre-existing runs of the baseline scenario as our point of comparison and to provide the initialisation condition for the modified run. This reduces the computational load of running a complete new model when rapid results are desired.
Following the country-level analysis of Forster et al. (2020), we apply this technique to generate four scenarios of emissions and concentrations incorporating the effects of lockdown and various different recoveries. We also process the emissions fields 40 through an atmospheric chemistry model to provide the ozone field, often required as an input for General Circulation Models (GCM). We finally describe a protocol for a model intercomparison project (MIP) assessing the impact of national lockdown measures.

Data sources
For this exercise, we change the concentration of the three main greenhouse gases (GHGs): CO 2 , CH 4 and N 2 O, and emissions 45 assumes that CO 2 emissions from agriculture, forestry and other land use (AFOLU) are reduced by the same amount as the average CO 2 emissions change from industry, whereas here we assume no difference in AFOLU emissions. Many gridded IAM models do not report emissions monthly but only on a five or ten-year average basis, and climate models simply interpolate this data for the remaining years. Typically, emissions changes are smooth and the amount of data lost in this way is therefore low. However, when a particularly strong trend occurs suddenly this is difficult to represent on this timescale. Because 2020 is a year normally reported by IAMs, if the emissions for this year were simply corrected without changing anything else then the effects of lockdown would also be felt in the interpolated years before it started, as well as in 95 following years when it is expected to have ended. It is therefore necessary to interpolate additional years onto all datasets with lockdown effects on them -we interpolate 2019, 2021 and 2023. We require data for 2019 to ensure no emissions reduction in the years before lockdown starts. We similarly interpolate 2023 before modifications are made to ensure long-term effects only happen when the model dictates. Since the years 2020 and 2021 are expected to be very different from the surrounding years, they are both interpolated and modified by the effects of lockdown. The year 2022 is defined as exactly equaling the 100 interpolation between the effects of lockdown and the baseline behaviour, so does not need to be interpolated. By request from certain groups, monthly data with every year from 2015 to 2025 is available, as is daily data for 2020. Since emissions change on a seasonal basis, interpolated years are interpolated between the same months of the years with available data on either side. This is done before imposing the effects of lockdown, except when we add data for 2022.

105
The process for handling emissions is more complicated than concentrations, and was subject to a significant change between version 3 and version 4 for shipping. This is described in the table on versioning, table 1.
The baseline SSP2-4.5 data contains emissions for nine sectors: AFOLU; energy; industrial processes; surface transportation; residential, commercial and other; solvent production and application; waste; international shipping; and aviation. Aviation emissions are subdivided by altitude, and handled separately. These mostly map well onto the sectors whose activity levels 110 were investigated by Forster et al. (2020), with two exceptions. Firstly Forster et al. model residential and public/commercial buildings separately, so we will use the emissions-weighted mean of these for each country. Secondly, Forster et al. did not have sector-specific estimates for emissions changes from solvents, waste or AFOLU (although CO 2 emissions from AFOLU are implicitly assumed to scale with industrial emissions reduction, as discussed above). We will assume that no changes occured to these sectors. We similarly assume that the small island nations and regions like Antarctica not included in the 143 nations corners are sea, the pixel is instead classified as international waters and is therefore modified by the internationally averaged change in shipping activity rather than the national change in shipping level. Using this definition, only shipping emissions are found in international waters. We emphasise that this classification scheme is purely for emissions calculations and should not be interpreted as a statement of political designation. This treatment of the seas began in version 4 -prior to this, all sea activity used the national shipping activity level of the closest country. Examples of this analysis for April can be found in figure 2, and 125 the globally averaged emissions reduction factors can be found in figure 3. An animation of the global distribution is available in the emissions modification github repository, stored in Zenodo, see Code availability.

Aviation emissions -monthly
The aviation activity level is always treated globally. The daily number of flights is taken from Flightradar24 free data. This is available from 6/01/2020 up to the time the version is defined. The "null flights" level is calculated as the average number of 130 flights per day in January, and activity level is then expressed as the daily number of flights divided by this. After the end of the available data, we project a linear trend, fitted to data collected after 1/05/2020 (not inclusive), until it reaches the long-term level. This is defined as 2 3 of the reduction factor of the last complete month of data. In equation form, with angular brackets indicating the mean over the subscript period, f (t) representing flights on the date t days past January 1st and a(t) representing activity level, f 0 = f Jan and 135 for constants m and c that are fit to the data from dates after 1/05/2020. For some versions of the data, the flight activity level is already at the 2 3 reduction level by the end of the period of collected data so no linear interpolation is seen. The monthy average of this data is then taken to produce the activity level of aviation. This is assumed to be globally uniform and the same across all altitudes. See the graphical illustration in figure 4a.

Aviation emissions -weekly
Most analyses do not use any finer-grained data than monthly, but for one project, finer-grained, weekly data is investigated for the 2020 data. For this project, using open-source data was not required, so we obtained previous years of flight data from FlightRadar'24 to better control for seasonal changes. We can then use the weekly-averaged data from 2018 and 2019 for the corresponding day as the baseline instead of the January values: 145 a(t) = 2 f 2020 (j) j=t:t+7 of 08/10/2020, the data for 2019 has also been released open-source, so later interations of the code will likely use a similar approach to this for monthly data too.

Protocol for CovidMIP
The emissions and concentrations described above are used in CMIP6 Earth system models to simulate the climatic impacts of lockdown. There are three focuses or strands to this MIP. The first is to address the short term response to the emissions 195 reductions, and the second to address the longer-term response to alternative recovery scenarios. There are sufficient differences in design and groups interested to make this split pragmatic. The third focus is on understanding processes and separating out the role of individual forcing components in contributing to changes in radiative forcing and climate.
Some model groups also have the ability to perform "nudged" simulations which force their model's physical state towards a pre-defined meteorology. This can reduce signal-to-noise issues and help identify aspects of atmospheric composition which 200 might not be apparent in "free running" model simulations. This is allowed where models have this capacity.
It is assumed that model groups have performed the SSP2-4.5 scenario simulations and we use this as a reference set of simulations (baseline) against which we will compare CovidMIP results. Any forcing or aspect of simulation not explicitly defined in this protocol (for example HFCs or land-use) should be kept unchanged from the SSP2-4.5 simulation.

Strand-1. Near-term impact of COVID-lockdown emissions reductions 205
The goal of these simulations is to assess the impact of COIVD-induced lockdown emissions reductions on climate, atmospheric composition and air quality in the near term. To achieve this, we use emissions reductions as close as possible to real emissions as reconstructed from activity data described above. A recovery to baseline emissions is assumed by 2022 and simulaitons should run for 5 years (although longer is also accepted -see sectin 6.2). This uses the two year blip forcing.
Protocol details: In order to maximise the chance of being able to extract a potentially small signal we request as many ensemble members 215 as possible.
o Ensemble size: as large as possible. We suggest at least 10 members, but there is no required minimum.
o Initial condition ensemble, with model-by-model choice how to arrive at perturbed initial conditions. Note the requirement that parallel SSP2-4.5 simulaitons exist, so we anticipate that the same ensemble technique and initial conditions can be used. -Experiment name o "ssp245-cov-strgreen", "ssp245-cov-modgreen", "ssp245-cov-fossil"

245
COVID lockdown has led to reduced emissions across a wide range of sectors and species. Some of these have competing or offsetting effects on atmospheric composition, radiative forcing and climate. For example, Forster et al (2020) show that at a global level the near term warming due to reduced aerosols may be at least partially offset by reduced greenhouse forcing from ozone. Only on longer timescales does the climate effect of CO 2 reductions become significant.
In this strand we use both detection and attribution techniques and fixed-SST diagnosis techniques to isolate and compare 250 the effective radiative forcing (ERF) from individual emission types or categories, and their full implications for regional and global climate evolution.
Two detection and attribution simulations are proposed to parallel ssp245-covid, and allow the separation of the effects of aerosols and well-mixed greenhouse gas perturbations on climate, similar to the way that hist-aer and hist-GHG simulations in DAMIP allow the separation of the effects of these forcings over the full historical period (Gillett et al., 2016). The ssp245-cov-255 aer simulation is identical to ssp245-covid, except that only aerosol and aerosol precursor emissions (BC, OC, SO 2 , SO 4 , NO X , NH 3 , CO, NMVOCs) follow ssp245-covid, while greenhouse gas concentrations, ozone and all other forcings follow ssp245.
Similarly the ssp245-cov-GHG simulation is identical to ssp245-covid, except that only the concentrations of the well-mixed greenhouse gases follow ssp245-covid, while all other forcings follow ssp245. We suggest that groups run as large ensembles of these simulations as possible, but no minimum size is required.

ERF calculations
The most commonly used methodology for estimating Effective Radiative Forcing (ERF) is to utilize simulations with fixed sea-surface temperatures (fSST) and prescribed emissions (Richardson et al., 2019;Pincus et al., 2016;Myhre et al., 2013). This allows the atmospheric conditions to rapidly equilibrate, and rapid adjustments to play out, but broadly avoid the feedbacks associated with a change in surface temperature. For example,  found thirty years of fSST simulations 265 sufficient to reduce the global 5-95% confidence interval to 0.1 W m −2 , superior to other methods.
As CovidMIP aims to quantify ERFs that are likely to be relatively weak, Diagnostics For all strands, we request model groups produce the same diagnostics as per their baseline SSP2-4.5 simulations.

285
Nudged simulations As an alternative to fSST based ERF diagnosis, some models are able to run nudged simulations where meteorological conditions (typically surface winds and temperatures) are forced to be comparable between signal and baseline.
This allows for a direct, time evolving calculation of ERF based on differences in top-of-atmosphere radiative imbalance between the simulations (Chen and Gettelman, 2016;Liu et al., 2018). Although they may not capture the full range of atmospheric adjustments , nudged ERF calculations are sufficiently comparable to fSST based calculations 290 that they will be used in CovidMIP provided they have prescribed the same emissions as described above.

Anticipated analysis
CovidMIP analysis plans include specific analysis on near term climate effects of emissions reductions. This will draw primarily on 2-yr blip simulations up to 2025. Focus will be on main climate outputs of surface temperature and rainfall, winds and basic circulation and also basic level biogeochemical diagnostics such as carbon stores and fluxes. Similar analysis is planned, 295 but focusing on temperature and precipitation extremes, with analysis based on daily tasmax and precipitation data and a focus on regional aspects.
Regional-specific analyses are possible, with East Asia a particular focus region as this is where the largest effects of emissoins have been seen in surface aerosols and air quality. The implicaitons of this on local rainfall and monsoon circulation patterns is of particular interest. North Atlantic and European circulation changes will also be investigated.

300
The effect of emissions reductions on CO2 concentrations is also of interest and may be investigated by ESMs with the capability of performing emissions-driven CO2 simulaitons. Similarly, ESMs with atmospheric chemistry schemes will be investigated to see the role of emissions reductions on surface ozone and PMs.
Model data will be made freely available via the Earth System Grid Federation (ESGF). Users of this data are encouraged to contact model group representatives and invite possible involvement in any resulting publications.

Conclusions
We have demonstrated a novel way to combine data-rich emissions nowcasting with long-term emissions projections to create a dataset suitable for investigating the impact of the large and unforeseen emissions reduction arising from lockdown. This will form the basis for a model intercomparison project to answer questions around how much climatic impact we expect to observe from lockdown measures in both the short and medium term. We also provide ozone field derived from these results 310 for models that do not produce their own estimates of this. Finally we provide a protocol for how different simulation groups can run experiments on un-initialised, coupled AOGCM/ESM.

Data availability
The output of these protocols is available from several zenodo addresses.  4.0 14/07/2020 Pixels whose four corners are in the sea use internationally averaged shipping factors Table 1. Table of noteworthy difference between versions of data. The first digit of the version number is incremented by both additional months of complete data and by major coding developments. The second digit represents significant coding changes or additional data use within the same final month of data. The third decimal place denotes changes in the times at which data is reported or minor bugfixes.

Scenario Assumptions
Baseline SSP 2-4.5 data is used without modification.
Two year blip Data is modified for all of 2020 and 2021 in accordance with observed activity levels in the sectors of different countries. This is projected to continue at 2 3 of the latest rate for the rest of the period. Activity is interpolated back towards baseline over 2022 and is equal to baseline thereafter

Fossil fuel
Follows two year blip until 2023. Thereafter, the effects of additional investment in fossil fuels during recovery are included in a globally uniform way.

Moderate green
Follows two year blip until 2023. Thereafter, the effects of small additional investment in green technology are included in a globally uniform way. Sets a global net zero CO2 target for 2060.

Strong green
Follows two year blip until 2023. Thereafter, the effects of large additional investment in green technology are included in a globally uniform way. Sets a global net zero CO2 target for 2050. Table 2. Summary table for the differences between scenarios. For more details on how these were constructed see Forster et al. (2020).