Talk on Air Quality Forecasting given at the UKCA training course in December 2022
okay morning everybody thanks for coming make start just to say we’ll be doing a group photo uh after the break today it’ll be outside the department I think like we’ll go around there kind of nice bit with the the plaque of the department kind of the on done that taken so um we’ll do that after the tea bre um but this morning um Prov a nice to say we’ve got Paul agne who’s been sitting amongst us doing tutorials this week um to talk about air quality for forecasting application thank you Luke morning everybody um so I’m Paul agu I’m from the Met Office uh where I work in the atmospheric dispersion and air quality group uh I’ll be talking about applying ukca to air quality forecasting uh I haven’t met anybody yet who is working on a quality but increasing numbers of climate modelers are are also starting to look at air quality and you never know what direction your phds might go in so uh so some of this could well if if you think it might not be immediately relevant to you it might be at some point in the future so I’m going to just go over some of the basics of air quality say a little bit about uh air quality observations in the UK my my talk will be really focused on the UK and Northwest Europe because that’s uh that that’s the domain that our forecast model operates on um but it will be relevant to to other parts of the globe as well I’ll describe our uh operational forecast system uh how it’s configured including ukca uh I’d like to say a little bit about how to validate an air quality model to check that the model is giving sensible results um and then I’ll say a little bit about uh we we’ve all been become familiar the last few days with the different chemistry mechan mechanisms within ukca not all chemical mechanisms are alike uh and I’ll say a little bit about the performance of the different mechanisms for air quality applications and then finally I’ll come to uh some work that’s in progress uh where we’re coupling uh a different model not not the UN but the name dispersion model to ukca uh that works at a relatively early stage but uh I’d like to just say something about it and then finally I’ll just mention a couple of data sets which uh may be of interest to people uh at some point if you ever come to work on applications of air quality so I want to just start with a headline which is that um ukca has been used to deliver the the UK National air quality forecast uh since April 2014 so it’s really it’s you know it’s not it’s not just a research application it’s been used to deliver this operational forecast uh every day of the week for for for eight years now so uh if you ever driven around um parts of the UK you may see one of these strange boxes on the roadside um uh this is an air quality monitor uh a roadside air quality monitor so when we talk about air quality clearly lots of different pollutants can find their way into the atmosphere but uh in terms of air quality we’re generally talking about uh a particular subset of the commonly emitted pollutants which arise in the course of uh normal human activities so that uh that includes NO2 uh I put carbon monoxide in brackets for a reason I’ll come back to so SO2 ozone and then the two PMS pm1 PM 2.5 and of course the uh the importance of these pollutants is that elevated concentrations of them can affect human health um what’s deemed an acceptable concentration varies from country to Country um but governments certainly in Europe Europe and and across the world really are required to warn the population uh if uh if elevated levels of pollutants are expected the legislative background for the UK um despite brexit uh is still uh a directive of the um European Commission on ambient air quality and and clean air for Europe uh which gives a detailed a very detailed specification of the regulations that governments have to adhere to both in terms of controlling emissions uh controlling air concentrations of pollutants and uh what steps they must take to warn the population in the event of uh air quality episodes and another relevant uh uh aspect of this is a report by the committee on the medical effects of air pollutant comeap It’s a UK government sponsored um committee which advised government on the effects of Health on uh the impacts of pollutants on on human health so since 2014 the Met Office has been responsible for delivering the national forecasts in the UK um oddly enough the the Met Office itself isn’t responsible for delivering an air quality forecast it falls to the government Department defra um and I’ll come back to the defra website in a moment but the thing that we actually forecast is something called The Daily air quality index or the Dai clearly I’ve mentioned the uh the five pollutants which contribute to routine air quality and it would be it would be ridiculous to expect the public to interpret you know air concentrations of multiple pollutants and try to figure out what that meant for human health so the idea of the dcky is that it takes suitably time averaged values of of the air concentrations of all those pollutants and combines them to make a single index which uh which goes from 1 to 10 it’s very badly named actually uh the daily air quality index uh it’s actually an air pollution index you’d think that for a daily air quality index that that 10 would be the best air quality uh but of course it’s the other way around it’s uh 10 is the worst air quality and that causes problems when you’re trying to communicate to the public um you you won’t I don’t expect you to look read this chart in detail but you can see this on the Met Office website or the defro website but it’s the recipe for how you actually compute the uh the daily air quality index so taking time averages of the different pollutants so for ozone it’s a rolling up8 hour mean uh for the for the 2 PMS it’s a it’s a daily average of the 24-hour hourly values for NO2 it’s uh it’s it goes on an Audi mean for SO2 it’s actually based on a 15minute mean um carbon monoxide used to be a pollutant of concern in the UK uh but since 2012 um levels had fallen significantly and it was it essentially it it ceased to become a significant prodct so in 2012 the Dai was redefined uh Co was taken out of it uh but PM 2.5 was added into it there weren’t prior to about 2009 there there had not been that many observations of pm2.5 routinely in the UK uh but that that increased in 20 in 2009 uh with the addition of a lot more sites to the uh to to the to the UK Network so the um the the main place for uh seeing what the what the forecast is in the UK again oddly enough it’s not the Met Office uh but it’s the uh it’s the defra website and if you just Google UK air this is what you’ll find um but the Met Office is responsible uh for supplying the information that goes on to this so we produce uh maps of the of the Dai out to four days ahead uh there’s a text forecast which does a number of things it it helps uh it interprets the maps um if for any reason the forecast is not performing particularly well it’s a chance for a human forecaster to intervene and to add some uh disclaims or qualifies or caveats uh and uh it’s also tweeted to uh to to anybody who wants to receive the Tweet we’re very fortunate in the UK with h with a a good uh Network for routine measurements of uh the air quality pollutants that’s funded by defra it’s referred to as the automatic urban and rural Network the a uh you can see a map uh on the right hand side there um coverage pretty good over England uh bit more sparse in uh in Wales especially Western Wales uh coverage not particularly great over the highlands of Scotland but then um you know much smaller populations up there uh and less less pollution uh there are different types of sites uh roadside Urban background and Rural locations because the characteristics obviously of the pollutants differ according to uh what what and how much pollution is being produced in different places and how mean values are available and you can see these uh very very easily on the defa website if you go and click on a particular site uh you can get you can get the current recorded value and you can also get a nice time series graph of what’s been happening for the last seven days over London there’s an additional Network called the London air quality Network operating by Imperial College and many of the local authorities over the UK also have a roadside monitoring which they’re required to put in place particular uh tra traffic hotspots where pollution values are higher uh one thing just to be aware of is that not all Sites measure all pollutants so if you just if you just glance at this map and take it at face value there is a kind of a sampling bias built into it because you you’ll sometimes uh say these take take these three sites um on the uh East anglian Coast here um you might see this is Norwich Lake and Fields you might see that showing elevated values of ozone uh and an a daily equality index of be being displayed of I don’t know five or six but the s’s right next to it you’re only showing too when you think okay what’s going on there well actually those those other sites may not even be measuring ozone so just just a slide to describe the characteristics uh of of UK air quality compared to many other countries um pollution levels are relatively low in the UK nothing like on the scale of pollution that you get seeing parts of China and parts of India uh and in terms of the Dai most of the time in the UK um it the daily equality index falls in the band one two or three which is grouped together and and just called Low um there are times during the year when we have air quality episodes uh typically 10 to 15 or so a year when elevated levels of pollutant uh will be uh will come almost all of those episodes are driven by uh e either or both of ozone and pm2.5 sometimes you might get a PM 10 episodes say if there’s like Sahara and dust or something where you’ve got very coarse aerosol being transported to the UK um most of the time the anthropogenic aosa over the UK is dominated by it’s in the pm2.5 component but uh for something like a Sahar undust episode uh the course component could dominate ozone episodes um typically uh appeared May to September obviously ozone’s photochemically uh driven um the episodes in in the UK will typically be somewhat less intense than uh than Southern European countries of course um in terms of PM episodes PM 2.5 episodes typically uh occur in in Spring and autum uh they’re usually driven by a low pressure a high pressure synoptic system to the east or Southeast uh of the UK which is drawing air from uh from the low countries uh and and all the pollutant that pollutants produced there and all the prec curses as they pass over the over the Over the Sea they can uh form secondary uh inorganic aerosol in particular over the UK uh and in the in those sorts of episodes uh ammonium nitrate is a very important usually the dominant uh pollutant so we have a a forecast system at the Met Office uh it comprises a number of components there’s the uh there’s the aqm model that’s air quality in the unified model be saying more about that uh but that’s a configuration of the um which uses ukca um the weather forecast systems at the office I mean globally there’s you’ll know there’s there’s a there a very large very sophisticated global system of uh meteorological observations which are collected by um met offices across the globe and uh including the UK met office and those observations are used to uh to feed into the model via the data simulation framework to improve the initial conditions of the model we don’t use air quality observations to they’re not assimilated in into the model but we use them in two ways we have a bias Corrections postprocessing system so after the after the model has run and produced its output we make a a statistical adjustment to the model based on the most recent observations and we also use the observations for um verification so working out how well the uh the forecast T performed compared to what’s actually observed and then that all comes together to produce the uh the the dcky forecast so aqm itself it’s a limited area configuration of the uh of the um we uh we started develop developing it Circa 2005 um it became operational in 20 0 uh it runs that a relatively coarse resolution for a limited area model these days the ukv model which provides the weather forecast for the UK runs at 1.5 km resolution and uh we are running at um roughly 11 km resolution so it’s still a relatively coarse resolution um there is it’s there are various problems um the way in which just increasing the the model resolution brings an improvement in skill to weather forecasts seems not to translate so easily to air quality forecasts so just increasing the resolution is not is not the Panacea that it is for a a weather forecast uh it’s a whole atmosphere model it goes up um to to a well okay okay not into the stratosphere but uh it’s got uh well we we don’t do prognostic chemistry in the in in the stratosphere model top um has is is around 39 kilm uh with 63 uh vertical levels uh so so many of you um probably working with global models and not not so familiar with limited area models so a complication for running a limited area model is that you need to prescribe uh conditions at the boundaries of your model domain whether that’s conditions of meteorological variables or of uh composition chemistry in aerosol um so for our our lbcs our lateral boundary conditions for meteorological um parameters we take those from the Met Office Global model and for the composition so the chemistry and the aerosols we take that not oddly enough not from a Met Office model because the Met Office doesn’t run a full Global chemistry model um in the forecast Suite but the European Center does um ecmwf formerly based in Reading um now moved to bolognia uh in Italy uh but they run the cifs the the chemistry ifs model uh and that provides the lbcs for our for our chemistry aqm um I said it runs ukca and it does for gas phase chemistry so we we don’t use any of the mechanisms which we’ve been working with this week the the the the chry Strat or the Strat trop mechanisms we developed in the early days of a and we developed this Regional air quality mechanism the rack mechanism um I I’ll come later to to talk about some of the differences between the chemistry mechanisms uh but it has uh 40 transported species 16 emitted uh 116 gas phase reactions 23 photolysis reactions we used the fat JX uh photolysis scheme that that when we initially started developing we were using the uh the offline 2D fotsis scheme but going to fast JX brought us a substantial Improvement in forecast skill uh it has representative alkanes alkenes and arines uh and importantly since 2015 2016 we uh it has the it includes the heterogeneous hydrolysis of n205 on the surface of aerosol particles which is is a very important mechanism in nighttime chemistry for producing nitric acid uh which can go on to rea react with ammonium to form ammonium nitrate secondary aerosol we are still using an old an older um aerosur scheme called the classic aerosur scheme which was the uh the scheme developed for climate modeling um prior to Glow map becoming available it’s a single moment scheme so it just um the traces are uh aerosol Mass uh and the aerosol types are sulfate black car organic carbon uh biomass Burning uh dust uh and and nitrate aerosol the emissions are are absolutely key so the uh aqm is is predicting we run with a 5 minute time step uh and the model is is producing hourly mean values of pollutant concentrations so it’s have having a a good representation of your uh your your emissions is absolutely key we’re very fortunate in the UK to have something called the national atmospheric emissions inventory again funded by defra uh which is a uh a bottomup inventory of uh all of the anthropogenic all of the important anthropogenic pollutants in the UK uh and that’s free freely available on the Nae website um our domain main goes outside of the UK um it covers much of Northwestern Europe so so we need more than just UK emissions so many of you you might be familiar with the emap uh European emissions ep’s a European organization which uh also provides a European uh wide emission statea set now at 10 km resolution um and then these These are these infantries are all annual total emitted values so so inevitably uh you have to provide your your own estimates of uh of temporal profile so so how the emissions vary through the day so for example for for for nox emissions um obviously Rush Hour Peak traffic times um emissions during those periods can be very different to the to the middle of the night right and also vertical profiles for certain sectors the emissions are divided into a number of different sectors um depending on what the origin is whether it’s for example emissions from traffic or or industry or or domestic uh and it’s often relevant to consider the injection height of those pollutants into the atmosphere so you have to add a lot of additional information to the to the bare inventories sorry that there’s the European emissions the EAP and you can see the the shipping tracks uh shipping plays an important uh contribution to uh to the European emissions in terms of the the way the model is run um so just as we’ve been doing this week running a row Suite aqm runs in the in the Met Office operational Suite under a cycling row Suite the forecast model is free running there’s no assimilation data assimilation either of weather variables or of uh chemistry and aerosol variables but I mentioned this postprocessing bias correction um which uses the observations the initial conditions uh I mentioned from for meteorology coming from the the uh the glob model for composition uh so the the initial conditions actually come from the the previous days t+ 24 forecast predictions and then I’ve mentioned the lbc’s which we take again from the global uh Met Office Global model for for met variables and from the cifs for composition we we forecast out to five days ahead uh so out to t plus uh 120 hours and uh we relax to a set of climatological because the C CFS only goes out to three days ahead um we run out of after three days we run out of uh lbc’s and so we we have a scheme for relaxing to climat climatological lbcs so just to say that the um the Met Office invests um a lot of effort and energy in verifying both its weather forecasts uh and in our case the air quality forecast uh to see how how good it was how good were our forecasts we um in terms of model evaluation for meteorological uh models and for climate models when often just uses metrics like um a whole field uh bias or root means Square error correlation for air quality metrics it has a for an a quality model it has a fundamentally different character if it’s being used to provide warnings you’re you have a particular interest in knowing if uh when when the for forecast for levels uh exceed certain thresholds which are linked to uh either to human health or to statutory reporting Valu so for ozone when the concentrations go above 100 microgram per cubic meter that’s that’s a significant threshold and so we use uh a different verification methodology which employs a a contingency table to say uh did we forecast an exceedance of that threshold yes or no in the observations was that threshold exceeded yes or no and there are various tests you can you can use to uh to evaluate your skill in the prediction of the exceedence of those thresholds if you’re comparing different pollutants um say you want to know how your model how good is my model um performing the ozone prediction uh the the oone predictions compared to say the uh the SO2 predictions well the ambient concentrations of those pollutants in UK a are very different typically ozone um ambient values typically say 80 to 100 micrograms per cubic meter for for so SO2 much probably in the range from from 1 to 10 micrograms per cubic meter so so if if you want to compare skill for those very different uh magnitudes you’re forced into using normalized metrics uh and there are various options for doing that I won’t go into too much detail now um but um there are some specific metrics um which which which are well adapted to um for evaluating bias uh and and error of the forecast in a symmetrical way treating under predictions and over predictions in a in a consistent manner uh and these these uh the the these metrics which normalize with respect to the sum of the forecast uh value and the observed value uh are particularly useful I mentioned that um for an air quality model just increasing the model resolution is is no guarantee of of improved performance and I just want to illustrate that with this comparison here of let’s just call them the red model and the blue model and on the on the left hand side I’m looking at the uh a plot of the of the model bias so that’s that’s the the averaged error and you can see that um generally the the red model has a small positive bias the blue model has a small negative bias uh but when you come to the air quality episode The the red model over responds to it and gives excessive predictions of of uh of this pollutant in in this case PM 2.5 the the blue model um underpredicts if you just look at the uh root mean square error of these two of these two models you can see that the red model uh has a much higher Peak rmsse the blue model has a substantially lower rsse so if you were just looking at the the the whole field error in the forecast you’d you’d look at this and you’d say well the the blue model is performing better so bias is the is is is is the averaged error uh means square error this fact 2 uh metric is uh it’s it’s a measure of the the fraction of of the of the of the observations of the model predictions which are within a factor of two of the uh of of The observed values and again you can see the blue model actually performs better there um but if we just go back in terms of a warning system the blue model model would be completely hopeless because when you need it when you need advanced warning of uh elevated levels of pm2.5 the blue model completely fails to respond to those episode conditions so actually the the red model would be a far better tool for providing warnings uh and this is just a particular time series at one site uh this is for ozone measurements again the blue model the black dash line is the uh observations the red model um doing much better in terms of the peak value there uh I won’t go into detail of this but it’s this this this is the contingency table that I was talking about with where where you essentially uh populate this table uh if if your event was forecast your your your event being an exceedence of a threshold whe was it forecast yes or no was it observed yes or no you add up the number of um times you met those conditions populate that table and based on that you can compute these uh conditional probabilities uh so the hit rate uh and the false alarm rate um and you can see if you do that going back to the to to to the mean field metrics the bias and the rmsse these are the this is where we said before looking at these the blue model looks like it’s performing best but in terms of these categorical metrics the hit rate um you can see the um the red model so so a perfect forecast would be have a value of unity here for the hit rate um a completely rubbish forecast uh has a very small hit rate you can see the blue model has a very small hit rate there so those are the sorts of metrics that you need to be considering if you’re uh validating an air verifying an air quality forecast uh and if you if you dig a little bit deeper into the two models here you can see that the this is um this is the frequency distribution of ozone levels uh given by the by the by the two models look drilling down into into the EDF um and what’s happening at these at these at these higher values is a very important uh diagnostic for understanding how you how skillful your model is at predicting um predicting elevated pollution values so so just to say that the summarizing the standard metrics often used by to gauge the skill of a weather forecast um need to be supplemented by categorical metrics when you’re validating an air quality forecast there are some uh interesting ways to visualize performance of a model um in an earlier slide I talked about these these normalized normalized metrics a normalized mean and and a fractional gross error um these I don’t know if people have come AC cross these plots before they we call them soccer plots uh you just plot the mean bias against the fractional gross era for each of the of your observations and observation sites uh a perfect forecast is a point here at the uh at the at the origin um this is typically what you would get for ozone you get a somewhat random um set of distribution of points here if you look at this for PM 2.5 P pm10 rather you get this this linear linear feature out this way a kind of a linear feature out this way and what this is telling you is that in this case with with the with the random distribution of points your your model errors are dominated by uh by random errors whereas in this case there’s a systematic uh either e either negative bias in this case here or positive bias in this case here so um those different ways of looking at the uh verification metrics can can add value in helping you very quickly uh understand the the character of the errors in your model um we maintain uh uh a near real time verification system where we take all of the observations we get a data feed of the Aur observations uh and we’re we’re able to plot um in almost Real Time Time series of of observed values and compare our uh forecast predictions as well that’s an incredibly valuable um tool for us in in in in near real time monitoring the performance of the forecast model um and then it’s often useful the secondary pollutants in particular um H vary on on on a on a much bigger scale than some of the primary emitted pollutants and it’s it’s important to understand what’s happening across larger sections of the country rather than at specific sites and so just looking at field plots uh is a is a very important thing as well just to to gauge the picture over the uh over the whole of the country uh and if if you if you these are the a sites where we we color them uh with respect to the um the observations and um so this is for the daki this this middle one is for ozone this one is for PM 2.5 so and this this was our model forecast so this was clearly an ozone episode um because it it’s the ozone Dy which is uh is is is colorful the uh the pm2.5 observations are all still in the low value so so we know that this is uh this is this episode’s being driven by uh by ozone I mentioned our postprocessing uh bias correction system so that there’s a fundamental difference in in the in the in the character of um the meteorological atmosphere and and the chemical atmosphere for a weather forecast it’s very important to get the initial point to get the initial conditions of the forecast correct and there there’s a whole methodology of data assimilation to try to to get to get the the best possible initial condition because that influences the subsequent evolution of the forecast for the for for the for the atmospheric chemistry that that simply isn’t true the atmosphere loses its memory of the chemical state of the atmosphere on a much shorter time scale than it does for the meteorological variables so so for met variables that initial State can be remembered by the atmosphere for for a number of days actually for for the for the chemical state of the atmosphere the memory of that initial condition it obvious it falls with time but typically after 6 or 9 or 12 hours um the memory of that initial condition is is lost so that’s why a traditional data simulation approach to air quality parameters just just doesn’t doesn’t work work it doesn’t improve the quality of your forecast far more effective is to have a measure of the the current biases in your forecast whether your model is over predicting or under predicting according to the most recent observations and then and then nudge the the model output in a way that replicates in a sense persists that that that bias and so that’s what we do and for certain pollutants it can be uh sorry for certain pollutants it can be a it can add a major amount of skill so for pm10 in particular where the emission inventories uh are very poor generally um this post postprocessing so so the raw model is the uh the Orange Line the the postprocess corrected is the green line and you can see much closer to the uh to the actual measurements just DET so um chemistry for air quality what are the requirements for a chemistry mechanism if you want to model air quality well they’re very different to those requirements for a climate model for a future climate model you you need uh a mechan ISM which represents the longer term average chemistry of the atmosphere short-term Peak values uh far less relevant um if you’re you know a climate model you’ll typically uh take all your hourly values uh if you have them your model time step may be um maybe maybe more than an hour uh but you’ll end up you’ll end up averaging them to produce monthly or seasonal means uh with with air quality modeling and especially in a forecast model those those short-term Peaks are what you’re trying to capture and um a chemistry mechanism which fits the bill for climate modeling uh is can be very different to what fits the bill for air quality so one thing we have done is quite recently because we’re stepping back to with all the structural changes um both in ukca with it with it now coming out of the and being part of its own uh in its own repository and also all all the changes that you heard about yesterday from Ben about the Next Generation modeling system alrick um we we we wanted to set back and re-evaluate uh what models we’re using what chemistry mechanisms what aerosol mechanisms we’re using so we’ve undertaken recently a comparison of our roq mechanism against against other chemistry mechanisms so in these in these plots the um the the blue curve is the uh the straat trop mechanism which we’ve all been working with this week um the rack mechanism is the red is the red line and the the observations is the black line here and so this is this is for um a run-of the model for an ozone episode in May 2018 uh and you can see that the uh The observed values have a long tail right out to uh say about 155 micrograms per cubic meter um the almost all chemistry chemistry mechanisms would struggle to to really model this this very high end of the uh of of the PDF uh but you can see that rack gets much closer uh with about up to about 135 um compared to Strat trop which struggles to get above about 110 micrograms per cubic meter and if you look at if you look at the performance of the model versus observations on this quanti quanti plot uh you can see you can see that very clearly here in in the mid-ranges of ozone concentrations both models do fairly well the black line is is the Ono one uh model observations line but at the uh but above about 100 micrograms per cubic meter they start to diverge and the uh Strat trop um significantly underpredicts compared to to rack Rack isn’t perfect not reach the one to one line but it’s substantially better uh that’s just another ex example of it um so um so we we we we wrote this report and we concluded that overall the Strat trop scheme struggled to uh under air quality episode conditions uh often failing to show any indication of an episode which the rack scheme generally captured you can see it in the two graphs here these are the observations uh red is is up in above dy7 you can see the rack mechanism is picking up this episode here uh but it’s barely registering in the uh straat drop mechanism um in an early slide I said so what what explains this uh in an earlier slide I said that the rack mechanism has representative alkanes alkenes and and arines the strap trop mechanism uh only has alen chemistry in it there are no Al so Al only has alkan chemistry in it no alenes no arines uh the alkanes are amongst the the least reactive of uh of the VC’s and this is this is a a bar chart of the uh voc’s from the national atmospheric emissions infantry and you can see that if if you look at the um the alkenes so athine here propan they’re they’re contributing significantly and and uh arines appear they’re contributing significantly and uh not including those in a in your Chemistry mechanism is going to leave you with a mechanism which is is is less active uh and is not able to respond and produce High elevated values of uh those Peak values of ozone we also looked at the uh the CRI um it’s a much more complex mechanism um than than than both rack and Strat drop um it gave very similar performance for ozone modeling uh as as as you as you’d hope um so it Christ drat is is is would definitely be a candidate um a viable mechanism to use in in an air quality model however because of the the many more species and reactions it’s about three and a half times more expensive computationally so going forwards we haven’t yet made a decision as to what mechanism to go with uh it would be an option to to go with CRI if we could live with the um um increased computational cost uh but we haven’t ruled out sticking with the rack and doing the work that would be needed to put that into the asab framework for aerosol modeling I said we we use the the classic scheme um this is quite a busy slide but it’s just um is just comparing the um the classic scheme with with the GL glow map scheme um so we’ve undertaken an evaluation recently my my colleague Floren malaval has looked at whether glow map is is a viable aerosol mechanism for air quality modeling um Flo found that there was some interesting uh predictions where whe whe whether the two schemes gave very different predictions and one of them was in the uh in both the dry and the wet deposition between the classic and and glow map uh you can see the so for classic on the left here you can see the magnitudes of these deposition fluxes uh substantially different to what you see uh with with with glow map so that needs more work just to understand those and to to have a clear picture of which one is is in better agreement with observations but the um and then another important thing I mentioned the importance of secondary nitrate aerosol for those spring and Autumn PM 2.5 episodes um the implementation in classic can sometimes just just way over respond and produce far too much ammonium nitrate aerosol and there’s an example here um from 2020 um it’s reassuring that the glow map implementation of of secondary nitrate aerosol doesn’t seem to be quite as um overreactive as as the classic scheme does so I don’t expect you to read this but it’s just to say that the the the overall findings of our evaluation of glow map for air quality um were pretty much that it there are potentially me although there are still some things that we don’t understand like the uh the per the behavior of the of the some of the deposition um parameters certainly at this stage it looks like glow map is is a is a viable candidate as a replacement for classic for air quality uh applications just to finish off I’m going to say a little bit about a different model um you may have heard of the of of the name model it’s uh a dispersion model it’s the me it’s the model used by the Met Office uh in an operational sense to respond when when there’s when there’s a release uh into the atmosphere of um pollut due to an accident so let’s say a chemical tanker crashes on on the motorway or a a um a scrapyard catches light and a pile of um uh tires goes up in flame and is spreading choking black smoke uh over the vicinity the the the the Emergency Services have a hotline to the Met Office and the Met Office will produce a prediction of very quickly of of where the the nasty stuff is is is going to be transported and that’s used by the emergency response responders to to advise their uh their strategy for dealing with it um that was the origin of of the name dispersion model uh so in it in its for many years it was um it was it was a lonian dispersion model so what that means Ben talked a little bit about lonian and oian models yesterday in a in a in an oian model you have you have a grid over your domain and you define parameters at grid points in L Gran model that there is no grid you’re releasing notional lran particles to which you attach your pollutants the lonian particles are aded by the uh by by by the by the winds which are read into the model uh and and move around in that way and you track these lonian particles um so it’s an offline model it’s not producing its own meteorology you have to have you have to have run an nwp model and stored all the the wind data and other things in add addition to wind kind of on dis first and then you read those into the model um so it has another uh some other uh features as well which are quite useful it turns out it’s possible to run a lonian model backwards in time so you can use that to see where air has come from at a particular time so these plots here um this is the chill Bolton observatory in Hampshire and if we’re if we’re measuring um if we’re measuring pollutants here we can say over the last 48 hours where the the air that we’re analyzing at at noon on a particular day where has it been over the last 48 hours and you can see the back trajectories where the Air’s come from that’s often a very useful diagnostic for understanding um where air has come from pollutants it might have picked up so going forwards um with the transition to ngms and the alrick model um the for a variety of reasons um we are starting to look more closely at at um stopping develop the development of of aqm and developing name as an air quality forecast model recently an oian advection framework has been added to name so um so it’s still currently an offline model it still needs to be fed with nwp from an additional model um but the idea is that we would use the um or alick to generate nwp data which is then R into name which is then coupled to ukca currently name is not coupled to ukca a it has its own uh chemistry and aerosol uh scheme which is has has proven difficult to support so the idea of coupling to name um um coupling to ukca and being part of the ukca community uh is very attractive and then using that to uh deliver an air quality forecast so we have this project underway new AQ um and the the idea is that um in a couple of years from now that we will have successfully coupled to uh to ukca and that we’ve evaluated the performance of of name uh for air quality modeling that won’t stop there though because the whole concept of of an offline model of having you know have needing this this big Archive of met data that you have to Lug around with the and read into the model I mean that that’s clearly very clunky what you would really like is is to have name embedded into your nwp model which a few years from now will be the alrick model and so uh so the thing that’s missing is that that big Archive of nwp data if name can be embedded into alric directly then that that will bring a host of um of benefits you you won’t need this huge archive you won’t have the the computational expense of reading in a vast amounts of met data uh you’ll have access to to uh to met parameters chemistry and aerosol parameters uh uh model time step resolution um so many many benefits but uh a lot of work needs to happen before we get to that and then just very quickly I just want to mention two data sets which if anybody is um work interested uh and planning to work on air quality um particularly for epidemiologists trying to show relationships between air quality and uh human health um either either through um statistics of uh UK mortality or morbidity the two very valuable data sets to be aware of uh so we we maintain a data set going back to 2012 of RT plus 24 forecast data which gives for the whole of the UK um a pretty a pretty high quality representation of um pollutant air concentration values uh over that whole uh time period uh data sets available in the Met Office Mass archive uh it’s so we’ve also now recently put it onto Jasmine um and then another thing that we’re doing as part of a big um ukri funded project is to generate uh a reanalysis of air quality so so that just you may be familiar with meteorological reanalyses where people have um so you go back in time uh and you’ve got access to all the historic observations um you’ve you’ve you’ve got a consistent configuration of the model and you rerun all of the data simulation and the model runs and you generate your your best estimate of the state of the atmosphere and all the Met variables throughout the whole period so we’re doing something similar for for air quality we’re going back to 2003 um using all the observations that we can access uh and running a configuration of aqm uh to get hourly values of pollutant levels uh all the way over the UK from 2003 to present day uh we hope to have the first version of that by um the end of March this uh next year so so we’re really quite close to that uh and then just very very quickly I just mention some uh activities we are also in terms of high higher resolution modeling um there has been uh some work done as part of this ukri project to couple AQ to uh what’s called a gaussian plume model um something developed by a commercial consultancy here in Cambridge uh called the adms model um which is capable of giving predictions on on a much finer SC scale I have to say the jury still out as to how much value that brings in terms of the improvements of skills skill uh into for for roadside predictions but um that remains to be seen but it’s it’s a piece of work that we’re involved with uh and uh something completely different um the Met Office does give make a pollen forecast and at the moment it’s done in a very um almost unscientific way in a sense we we’re developing using the name model we’re developing a modelbased pollen forecast uh which will replace uh the the current forecast system which just relies on Expert judgment and kind of kind of handwavy feelings about the the way that the forecast should should should be for the next few days uh so so it it it will bring a a model-based forecast with much higher resolution um over over the UK uh and and can be objectively verified against the admittedly very small number of pollen observation sites over the UK uh so that that that’s quite an exciting development it um that that that doesn’t use ukca treats pollen just as uh as iner traces but um who knows in in time we may uh we may find it worthwhile to uh to adopt ukca for that work also thank you [Applause] going back to9 K yeah does that include avation activity and is that important we we have Aviation emissions in in terms of the impact on UK surface air quality is pretty negligible more important is actually rep having a good representation of of all the cars driving to ethro yeah yes so we have a representation of of the dial variations um so we have that on a sector by sector basis um those those variations it can be difficult to to validate them and to have them H have and to be confident that that they because for example the the people who operate steel works and power plants they’re quite protective of their data and they don’t they don’t provide easy access to either the planned activities of of their sites or even recent historically so so that so the dial Cycles yeah you you make a guess you make your best estimate um and that that’s what what you have to go with chical compation time for your the chemical compation time oh okay um so running aqm for one day takes about 5 minutes um so for 5 days 25 minutes so it’s it’s pretty quick to run actually do you have to fit it window um yeah actually the most computationally expensive part of our of our suite is producing the lbc’s we have to we have to get the data from ecmwf’s archives uh we have to do various you know cut out various domains we have to do regridding um in some cases we have to ECM the ECM ecmwf model uses a different vertical coordinate system um it’s a it’s a hybrid it’s a Hy it’s a it’s a pressure based coordinate system um and we we have to for parts of it we have to we have to put that onto the onto uh onto the um geop potential height coordinate system uh it it all takes time and that’s actually the most expensive part of our of our suite but the whole Suite I mean the model runs you know round about midnight uh and results are typically available for the to the forecasters by about 1:30 in the morning yeah yeah so that was that was all chosen by this committee on the medical effects of air pollution comeap um yeah they they looked at all of the data which relates Health impacts to observable relating air concentration values to to observable Health impacts on people and that that was the basis for where to set the thresholds the yeah so the idea is that you know if if you’re put into a room with a high concentration of so SO2 you’ll probably start to be gasping after 15 minutes uh for ozone it probably take eight hours but yeah th th those considerations sorry you did mention there actually human interaction people look at for the right I supp is there anything like where you might spot an expected episode you think oh my goodness something uh yeah well it can be I mean things like um yeah Sahara and dust um coming into the domain obviously I mean we’re reliant on that being present in the boundary conditions that we get from the global model uh but you know let’s say you know an Icelandic volcano explodes and there’s there’s a flow of volcanic ash you know impinging on Scotland that would be you know we wouldn’t have a representation of those emissions in AUM so that but that would be an example where the forecasters could say yeah Northern Scotland um it looks like the maps are saying good air quality today but remember that the volcano just erupted and uh actually we’re expecting yeah a blanketing ashcloud to descend um you you’re right you you would think oh it’s a more sophisticated scheme surely it’s better um when you just look at it objectively it becomes less clear uh there’s an additional computational expense um and other other factors than you know the fact it has two moments instead of just just one other factors you know of which there are you know many parameters go into glow map I’m sure um and those other factors could you know a pro probably more important I haven’t seen any evidence to dat that having a twom moment scheme is particularly valuable for air quality modeling in the future I don’t know it’s it’s possible but would you have to adapt your bias post processing change um no I don’t think so because that that the bias correction is never interacting with with the model and the model schemes it’s really just looking at the the output PM 10 or PM 2.5 so how you get how the model gets to that is irrelevant well thanks Paul [Applause]