Explainable Climate Science

Key Takeaways

[Music] this is data skeptic time series the podcast about how to predict the future based on historical sequential data episode number climate science is yet another application where machine learning is rapidly being adopted in this episode i interview zach labe about the use of layer-wise relevance propagation as an explainable ai technique it can be used to better understand how deep learning models were able to make predictions about the climate system [Music] my name is zach labe and i'm a postdoctoral researcher at colorado state university in the department of atmospheric science can you tell me a little bit about what you're working on in your postdoc currently i really classify myself as a climate scientist i tend to think about climate change and climate variability which i'm sure we'll talk about more and my work really comes at climate change and thinking about it from the perspective of data science i'm not only using you know climate model simulations which are these huge big types of data sets but also diving into machine learning and for me i've actually only been in this sort of machine learning ai world for only about a year now i didn't really do it in my graduate school but for my post doc i was really hoping to you know start learn a new skill set to really apply these types of tools to climate data to better understand climate change well despite being relatively new to it certainly you'd been at least aware of or heard of machine learning are there any takeaways being a year in what was surprising or unexpected about it for the field of climate science in general we have lots of tools really to extract patterns and climate change that's really what we're all about is looking at maps of data and looking at patterns and we already have all of these existing tools so i think going into machine learning you know i was skeptical that the machine learning model perhaps it could be better statistical you know analysis but is it really using physical real mechanisms in the client system and i think i've been relieved and happily you know surprised to see that at least the things i've worked with the machine learning models are really using like physical things that i know as a climate scientist exist in the world to solve these problems so i think it really to me at least has added credibility to a lot of these methods for applying these two climate data problems could that be due to researchers like yourself who have done a really good job of feature engineering or is it that machine learning is really figuring things out given the right data yeah i think it's a bit of both i think one thing that's been important at least in the work that i've done to at least give me this idea that these models are doing credible physical processes is this explainability methods to really try to understand what's going on in the black box i think has been so important for me to really help as i said give credibility to the machine learning problem but also you know learn something new it also sparks me new ideas for my research what are some of the questions that you can use machine learning to help answer i think the way i have looked at machine learning so far is actually a bit different than other approaches where often machine learning at least in weather and climate in the world i live in is often used for prediction problems maybe we're trying to get a weather forecast or a climate prediction and you know they set up these advanced machine learning statistical models to make a prediction but the way i have approached it so far is actually setting up very simple machine learning problems problems where the answer necessarily is not interesting at all however using the explainability methods that's where we can actually learn something new and new being patterns i can look at these features like a map of a climate variable and use the machine learning methods to understand patterns which is what we're really interested in climate science and why is machine learning required why can't a couple of grad students find these patterns i think that's the hard question and something i grapple with is you know i often get asked what is the advantage of using machine learning methods you know why not use traditional linear regression problems you know i use artificial neural networks so what is the advantage and i'll say it's difficult to sometimes answer that one answer being that we don't know if we don't try we don't know if we can find interesting non-linearities by using these types of you know something like neural networks but other things you know these explainability methods we're doing really give us maps of data which is something climate scientists always work with and it's really nice that we can compare multiple different types of methods you know statistical toolboxes that we have and see how these ai methods compare and sometimes they may look exactly like other methods we may have used like principal component analysis but other times we might find new patterns and i think that's what's so exciting about this field combining data science and machine learning with climate science and other weather and geoscience applications is that we really don't know what we all could find and i think it's wide open for future research and new graduate phd theses well the main paper i invited you on to discuss is titled detecting climate signals using explainable ai with single forcing large ensembles could you elaborate on what the single forcing part is all about sure and i think the single forcing aspect is the unique part of this work so just a little background from climate world is we often look at climate models so climate models basically are running lots of different types of equations to try to simulate you know the climate of our planet both in the past but also the present and in the future and these climate model simulations have all sorts of things interacting they have changes in the ocean and in the atmosphere and the ice and they all coupled together so there are these huge you know simulations that require expensive super computers to run and often when we run them we include climate forcings so what i mean by that is things like changes in greenhouse gases like carbon dioxide to understand how that affects the climate but normally when we run these simulations we include all of the forcings so all the things like changes in aerosols like particles in the atmosphere that come from factories things like carbon dioxide and other forcings they all change at the same time but what's difficult is really to understand the climate change aspect of it is you know if we have a region that's warming up you know over north america how can we attribute that warming over north america to these different forces something like an aerosol or something like carbon dioxide or perhaps something like changes in the land so recently they've run these new simulations they're called single forcing simulations where they take a climate model and one of them will run it with all of the realistic forces just like the real world then they'll run another simulation that's completely identical but in this case they will actually hold carbon dioxide constant so it will not increase while everything else is changing like the real world and then they'll run another simulation where aerosols the particles in the atmosphere don't change but carbon dioxide does and the idea is by running these simulations where we kind of hold different forcings constant or let them evolve through time we can try to pick apart and actually attribute how changes in the climate are caused by these different forcings and as the title suggests it's doing these detections with single forcing large ensembles what are the ensembling techniques and why were those called for instead of just some i don't know linear regression like you suggested so for a climate model simulation we can run one simulation and that will be a realization of the world with all of the changes in carbon dioxide but we can also run that simulation again and in that case we get this idea that there's natural variability in the climate system so basically there's this unpredictable noise in the climate system that you could think of that is like changes in the weather you know tomorrow won't be the same as today and that is a very unpredictable part of the climate so a challenge for climate scientists is trying to understand is the change due to something like carbon dioxide or is it due to this sort of inherent unpredictable variability this natural variability so what we do when we run climate models is we run ensembles so basically we take the same climate model simulation and we tweak the initial conditions by just a real small round off error something like 10 to the minus ninth change in temperature and we do that like 20 times and then we run the simulation and while all of these simulations will have the same amount of warming by you know the end of the 21st century the differences between them due to this sort of random initial condition change is this natural variability so really we're trying to understand in this paper not only how different forcings like aerosol and carbon dioxide affect the climate but we're also then trying to understand how does that differ from just natural variability i think everyone's heard of this i don't know if it's a myth or a useful heuristic or whatever but the concept that a butterfly flaps its wings somewhere on the other side of the planet and that's the start of a chain of events that causes a tsunami how much variability is there in the planet system with these small tweaks yeah that's an excellent analogy and really where it comes from how to set up these experiments by just tweaking the initial conditions there's quite a bit of variability in the climate system it can vary as i said on days but it also has a variability from decades to multiple decades in time it's quite large if we take an area that is experiencing a lot of climate change like the arctic that's one of the areas that's warming faster than anywhere else in the planet if you actually go back in the early 20th century there was a relatively warm period in that arctic like not as warm as today but warmer than 30 or 40 years ago so it's this idea that there is a pretty large amount of natural variability which is why it makes it difficult to really attribute climate change to different regions so by running these types of ensembles this makes it a lot easier to understand that range of that variability and let's get into some of the explainable ai sides of things obviously when you have a complex system it stands to reason you need a somewhat complex model to describe it how can you balance that with the need to understand what the model is doing yes in in this case we're using very simple neural networks the explainability technique we're using is this thing called layer wise relevance propagation it essentially takes your features from your inputs and in outputs after you use this technique it outputs a heat map telling you the relevance or how important that feature was to the final prediction from the neural network so in every example our research group has been using generally our neural networks are very simple which allows this type of visualization technique to work really well so we've been lucky that our inputs are quite data rich and it seems that even though it's not very complex our models they seem to be still performing with high accuracy but also they're able to be useful for these explainability techniques i have not had a ton of good luck in my personal life with heat maps it seems like every time i make one i essentially make a population heat map even though i'm trying to make something else when you have a climatologist eyes and you look at the layer wise heat maps you're getting what does it tell you i guess you know going into this the first thing i wanted to look at for these heat maps are do the regions that pop out on the heat map do they correspond to features in the climate system that i know are important to me that was sort of like this credibility check the layer wise relevance propagation method is really showing that the network is using real things in the climate system so one thing that popped out in our maps is this really important area in the north atlantic ocean and that popped out real bright on the heat map suggesting that was a really important area for the neural network to learn to make its prediction and in the climate system i know that that area is very sensitive to these climate forcings so to me it was very much sort of checking with my eyes as a climate scientist with these heat maps to see what pops out to me as being important thanks to our sponsor bb edit from bare bones software if you're a mac user and you have any need whatsoever for a text editor which let's face it that's got to be all of us you have got to check out bb edit bb edit 14 provides mac users the most powerful professional grade text editor around it's known as capable and rock solid and can process files that are large enough to bring other editors to their knees this tool has been beloved for over 25 years and i can understand why if i'm just typing up some simple hello world i guess any text editor will do but when i really get into the weeds and need to do some data processing maybe i've got a big messy file i need to find all the non-ascii characters and remove them bb edit's there for me it has best of breed multi-file search and replace and great scripting support for text transformations there's a whole menu of text transformations that you'll find handy if you're a serious data scientist or software engineer you're going to bump into situations where you need a powerful text editor so download and try bb edit today you can get it at www.barebones.com see how much your productivity can increase barebones.com and download bb edit [Music] today's data lives in the cloud so your business needs a cloud-native next-generation solution for intelligent accessible data analytics estrato is a modern bi solution for modern bi teams your organization's data already lives in the cloud which means you can access the single source of truth estrado accesses your data where it already lives with a live query approach so there's no need for data movement ever you maintain the agility speed and elasticity that modern cloud data warehouses deliver estrado's gallery of examples will give you a quick impression of the power estrado can give your team the sales workbook example gives me a good indication of the quality of dashboard i can create for my collaborators and the tour de france workshop shows off a few other features on competitive cycling data that i could see being really inspiring on implementation for when i have a demanding executive that needs to really slice and dice the data in a lot of custom ways when i first started exploring estrada i connected a google sheet of one of my data sets just so i could kick the tires after that integrating with snowflake as a back end was super easy you'll find that integration to be pretty painless and definitely worth your time checking out if you're a snowflake user with estrado you can bring complex data to life with no coding skills necessary to learn more visit estrado dot io data skeptic estrado is a-s-t-r-a-t-o estrado dot io slash data skeptic [Music] and can you describe the underlying data set that you're training on what is the input and where do you get it from so the inputs are coming from these climate model large ensembles so we're taking these ensemble experiments that have these different forcings differences in carbon dioxide differences in aerosols and then one other simulation has everything that's supposedly realistic just like the real world and so my input to my neural network is actually just the map of temperature i'm taking essentially an annual mean map so one map for every year of temperature going back from the historical period so what i mean by that is like 1920 all the way to the future to 2080. so i essentially give the neural network this one map of temperature and it's all latitude and longitude points on earth so it comes to something like a thousand five hundred points or features that go into the network and then this layer wise relevance propagation essentially outputs a heat map of those features which correspond identically to the map of latitude and longitude points when i think of machine learning i'm always trying to make tabular data but the earth isn't quite tabular it's an oblique spheroid do you face any challenges around quantifying the data because it's not in such a rectangular format right yeah that's a good question it's definitely something our research group has been exploring in at least in the ways we've set up the problems so far it almost seems that the neural network is learning to weight the importance of the different latitude and longitude regions to account for this spherical nature of the grid so in our cases you know we haven't really used any waiting methods for areas like in the arctic which have a smaller latitude and longitude little grid box compared to the tropics but i know for other types of groups which think about sort of weather prediction they have definitely applied different types of weighting methods to give some latitude some pixels more weight than others just because of the spherical nature so i would say it's definitely a challenge and something i i know a lot of research groups are thinking about how to take that into account for machine learning models so when studying these forced factors you could run through a whole bunch of different examples of what if this was off by this much in this region and use these tools to get an you know understanding of how that might evolve do these layer wise maps are they a sufficient answer then when you go and look at or are there other statistics you're considering as well for this paper we're really focusing on i guess the pattern so in our case the heat maps actually correspond very nicely to these climate patterns that we're really trying to disentangle from the different forcings so that's one of the main metrics and as i've also mentioned we set up our neural network very simply so i don't think i've said but the output of our neural network is the year of that map that i input so it's essentially a very simple prediction problem coming at it from i don't really care what year of the map it is i care about more of the climate patterns so in our case the layer wise relevance propagation maps really reveal these patterns that we're interested in you know how does the neural network learn that that map of temperature comes from that year and then what does it look like between these different climate models with different forcings so in our case the visualization explainability part is really the heart of this work i'm not sure if i fully follow on predicting the year just yet so it i could definitely understand where the system could learn historically that you know something was brewing in the atlantic and it can see a later stage of that event that could be something it could learn i guess a time series over years but as you forecast these aren't they drifting further and further from some ground truth timeline absolutely and so what we're really trying to understand is the climate aspect so how do these forces affect regional patterns in the earth system through time so essentially can the neural network recognize these patterns that are attributable to the different forcings so essentially the neural network is learning this time evolving nature without giving it any information temporarily to understand the patterns like for instance one area in the north atlantic is actually cooling due to climate change while the rest of the globe warms so the neural network actually can identify this region this anomalous cooling over the north atlantic to learn how it evolves in time essentially predicting the year and i think our group has used this idea of predicting the year for a couple different research applications and i think going back to the initial you know stage of setting up this problem we were skeptical that it was going to even work that the neural network could how could it possibly learn you know this temporal information by just being given this static map and so i think what's come out of all of this work is that the layer wise relevance propagation is revealing these regions that temporarily change and the neural network is learning how these features correspond to each other and change through time you've described some of these neural networks as being simple how simple are they can you describe the architecture of it for these neural networks we have two hidden layers essentially 20 nodes in each hidden layer and then this output layer is essentially predicting the year so we've tried you know different architectures adding more complexity and it still seems that about this complexity gives us the most accurate prediction for predicting the year of the maps and in thinking through the results and where this can go obviously it answers a lot of scientific questions are there any policy questions we should learn about naturally there's a lot that probably needs to be changed on that front but is you mentioned some of these hot spots or critical zones do we really need to ban aerosols in certain locations yeah it's an interesting point so one of these challenges as i mentioned for running climate models is this idea of these different forcings and how do we attribute changes in different regions to each forcing and i think result that came out of this work and one we've known in climate science is that aerosols are actually really difficult to simulate they're really important for the climate system so i think what comes out of this work from a climate perspective is that we really need to check how our climate models are simulating aerosols are they realistic with the real world and how might it change going forward in the future for instance one thing we expect is that more and more aerosols will actually be leaned up going forward in the future as more and more countries you know add policies to reduce pollution and so the question is how are these changes in aerosols in the future going to affect overall climate change due to carbon dioxide and i think these regions identified by this explainability method highlight some of the areas that could be particularly sensitive to changes in aerosols and that we really need to check our climate models and see how well they're simulating this important component do you have a sense of the degree to which the explainable aspects of this work how critical does it hinge on that and getting things published i think that was definitely the critical part of this work finding these regions and then comparing them for the different forcings i think for me at least in machine learning and how i feel about it with climate science is that we need really these explainability methods or interpretability methods to really understand what's going on in these machine learning you know black boxes per se i think at least in our field you know there's still some skepticism about these types of methods but i think adding these visualization tools is so critical to getting more and more people thinking about trying other data science methods to understand important questions like climate change well i think a healthy amount of skepticism's good i assume that when einstein first published physicists had to be initially skeptical they've all gotten on board now where is the climatology community on the adoption of ml uh hype cycle i i think it's growing exponentially one thing i like to do every morning with my coffee is sort of go through the climate science journals and and see what's being published and you know what comes on archive or early online releases and i think even just me in the last year it's just been an incredible growth of applying machine learning methods to earth system earth climate problems and i think people are coming in from all different perspectives to me i i've set up really simple problems but used explainability methods to hopefully learn some new science and other people are using it for prediction problems trying to understand you know weather forecasts and then other people are using machine learning for parameterization so how do we improve these climate models that i've been talking about you know how do we improve their simulation of aerosols and can we use machine learning methods to really reduce some of all of this computational demand to run super computers and again you know there's more and more growth in machine learning and in fact recently one of our journals just came out with a new journal devoted to ai methods for weather and climate which actually just came out a few weeks ago and i think at least this society it's called the american meteorological society i really think by showing their adaption um of this new journal shows how much growth is really undergoing for machine learning in our field and to what degree is it critical or maybe not so critical that someone with strong machine learning chops learned some fundamentals in climate science can i just come in brute force bring linear algebra to the table and help or do i really need to know some of the underlying models i'm always a proponent of team science and transdisciplinary research i think it's so beneficial to work with domain scientists and data scientists coming together to understand these problems i would say you know in my field we are data rich we have data coming from climate models we have data coming from satellites coming from weather stations on earth and all of these data sets come in different formats they have different uncertainties different grids and i think someone who's outside the field it's challenging to understand especially the pre-processing stage of our data and i think it's so beneficial to approach problems like climate change from this transdisciplinary idea you know working together with domain scientists and data science to answer the questions i think is really critical to understanding the data problems we have and potential solutions what are some of the open challenges in that regard is it compute or access to data or the volume what really is a limiting factor i think all of the above but i i would say we have so much data right now and just almost not enough people to really analyze it our climate models produce you know petabyte after petabyte of data and we just really can't using particularly traditional methods on our desktop computer we just can't analyze all of this important data so i think adding in you know work with data scientists and using you know non-traditional methods like machine learning i think is really a way we're going to be able to really look at all of this data and think about it from different perspectives well climate's something that's very much on my mind the more i learn about it the more concerned i get and i guess studying it for scientific reasons alone is is more than enough to do good research but i'm curious if you have any thoughts on the impact factor your work can have if there's any direct correlates to how we make a positive impact in this area my work and how i think about it is that i i hope you know i see climate science as moving away that we're working with more and more people outside of our small niche field you know as you mentioned you know climate is a huge problem and i think more and more i'm hoping that climate scientists work with people in other fields to think about our data and reach more impactful results so i guess thinking about my current work as a postdoc i see it in hope that i can show you know we can use new tools to really understand the climate system going you know in the past present and future and by working with data scientists we can really revolutionize our field by understanding all of this data we have is this paper uh kind of the culmination of a line of research or is this an ongoing effort for now i've kind of moved in a slightly different direction still applying you know machine learning to climate problems i think what i you know envisioned this paper really is showing how we can use explainability methods to add credibility for machine learning and i think by using these single forcing large ensembles was kind of a unique data set to really play around and show the value of something like layer-wise relevance propagation and i hope it sparks interest you know in applying these types of methods to future machine learning problems and i think there's an endless amount of problems we can think of setting up with machine learning and i hope this paper shows that by adding these explainability methods we can really learn new science the layer wise regression technique you applied is visually stunning i encourage listeners to go look at the paper even like me can't perfectly read the data without certain background it's obvious that you can tell what the model is relying on its specific features of the globe so it works really well for this problem i'm wondering if there are any other techniques you've explored along the lines of explainable ai personally i have not because i'm still relatively new to machine learning again only a year into it and for me at least this method is great because all of the data i work with always come in something like a map and you know what's awesome about this layer wise relevance propagation is we get a heat map of a physical map in return and that allows me to really compare my data so it's been great for a lot of the problems but i'll mention you know it's not a perfect method it is subjective to how we interpret i get one of these layer-wise relevance maps even for wrong predictions so sometimes it actually could be useful for understanding why the neural network was wrong but i'll give an example from a current project i'm working with is essentially i use lots of different climate models and i input a map of temperature into the neural network and then the neural network has to tell me which climate model produced that map so we run many climate models because they're not all perfect in climate scientists climate science we then compare how well they do and i was hoping then to input observations into this neural network and then the output would be which climate model was most similar to the real world observations and i would then use the layer wise relevance propagation to look at the regions that were most similar but it turned out the interpretation of this layer wise relevance was actually the opposite of my intuition instead of the neural network using features that were most similar between the real world and the climate model it actually used features that were strikingly different so i will say you know with these explainability methods they're certainly not perfect and there's definitely subjectivity that goes into them and i'm hoping you know moving forward in this sort of data science and climate science sphere to really explore different methods and how they compare against each other well i know in a lot of statistics literature you always report a p-value even though there's plenty of discussion about why that's not always the right thing there's something to be reported in common cases do you envision techniques like yours becoming a standard in the published literature i think it's going to become standard to include explainability methods for machine learning going forward in climate science i already can see it being used more and more for a lot of us you know who are not in the data science world directly a lot of these methods are new and even though they've existed in the computer science world you know for up to a decade or more so i think you know as we discussed earlier working between domain and data scientists will allow these explainability methods to be more and more common in climate science and i envision in you know in only a few years more and more of these methods to be commonplace with any type of research publication that goes into my world of climate science to have some sort of method to show how the machine learning model is making a prediction well zach is there anywhere people can follow you online i am on twitter at zlabe and so in addition to research and thinking about these types of problems i also am very interested in using data visualization to communicate climate change so i'm very active on twitter and showing different types of visualization methods to talk about climate change we'll have a link to that in the show notes zach thank you so much for taking the time to come on data skeptic great thank you for having me that concludes another installment of data skeptic time series thanks to our guest today zac labe thanks to our sponsors bare bones and estrada myself claudia armbruster as associate producer vanessa bly guest coordinator and our host kyle police [Music] you