121 - Revolutionizing Civil Engineering Through AI with MZ Nasser

Today we are talking about the real AI , one used to investigate complicated datasets with multiple variables , find hidden patterns and classify things in ways that we cannot figure out on our own as humans .

You may remember I had an episode about this already it was episode 28 with MC Nasser , who is one of the leaders of the AI revolution in civil engineering and in fire safety engineering as well , and today I have invited him again , almost perfectly , two years after this previous episode , with a great occasion for this being his book being finally released .

Mc has written a book about artificial intelligence for civil and environmental engineers . It's not a coding book that gives you snippets of code that you have to write on yourself and see . It's a book that will tell you why you need to learn AI and what you can get from that , so a very interesting addition to anyone's bookshelf .

In this episode , we explore why explainable AI is a thing and why we need it , what's causality and how you can figure out some really clever things by understanding how AI came to conclusions and how this can help us understand physics better .

So if you're interested in AI , if you're interested in new technologies , how this is growing , developing and how we will be able to use this in near future , this episode is for you . Let's spin the intro and jump into it . Welcome to the Fire Size Show . My name is Vojci Wigzyński and I will be your host .

Fire Size Show is brought to you in collaboration with OFR Consultants , a multi-award-winning independent consultancy dedicated to addressing fire safety challenges . Ofr is the UK's leading fire risk consultancy .

Its globally established team has developed reputation for preeminent fire engineering expertise , with colleagues working across the world to help protect people , property and planet . Internationally , it works ranges from Antarctic to Atacama Desert in Chile , to a number of projects across Africa .

In 2023 , ofr will grow its team and is keen to hear from industry professionals who want to collaborate on fire safety features this year . Get in touch at OFRConsultantscom . And now back to the world of AI with Emzy Nasser . Hello , great to have you back in the show . Thank you very much for having me .

Thank you , I'm very thankful you took my invite , and two years ago we've talked about AI . Today we're going to talk about AI , but I have a strange feeling . Ai is something else today . When we discussed it , it was a magical black box that can do things you don't understand .

Today is this mysterious chat in your computer that answers your questions and makes up citations . That's what I think the general public thinks about AI nowadays , and both of those are incorrect . So , first of all , how is this AI revolution happening in front of your eyes , affecting your work as AI pioneer ?

I mean , if you look at Twitter , people just share their preprints now much more faster than what joiners or conferences can do . When it's just amazing , I don't know where it's going to be . I can make six months for this year from now .

But also I would love to highlight your recent achievement , which is publishing a real book about this . Not that the previous papers were not good enough , but this one is a real book . It's called Machine Learning for Civil and Environmental Engineers a practical approach to data driven analysis , explainability and co-geality . Published in Wiley .

Congratulations on your book , and what I feel looking through the book is that it's really written with a user in mind , like it seems you are very interested in actually having people use this , and this is not a book that teaches you how to code . It tells you why you should .

I didn't want this to be like a traditional textbook that says you know , this is step one , two , three and four . I want it to be more like a conversation . So you can .

You will see a lot of footnotes everywhere , a lot of highlights , because I'm trying to figure out okay , if I read this as a student , what the questions would be , what would they have in mind ? And maybe , if I can address this directly , like in a footnote , it becomes a little bit easier for them to see what I was trying to say or do .

Basically , we started with a draft and I had a little bit of resistance from my editors because they didn't like the whole footnote approach , but I was able to convince them . It turned out to be nice . My goal from the book is really not to teach you how to code . The point from the book is not teach you how to use algorithms .

The point from the book is to teach you what the idea is , what machine learning is , how to think when you use an algorithm or a machine learning model , and how to . If you look at the results , what do the results mean ?

There's a lot of emphasis on , as you mentioned , explainability , causality , the result of emphasis on saying , if this model gives you a good prediction , you just don't take it for granted .

This is how you actually dissect it and figure out what to do and hopefully it turns out to be like a very good textbook and source for everybody who was interested in this area .

But really writing a book today about how to code when there is a chat GPT that can code and tell you what it's doing alongside it's not as attractive as it used to be to teach people how to code , but to teach people why they should and what are the outcomes they should expect and what should they aim for .

Now , coupled with capability of a layperson to code with the use of chat GPT , this is such a powerful combination . So I think the book has just received an unexpected boost through AI itself that can help the reader to actually implement what is in the book .

Now , when we talked the last time and we were discussing how different models in AI are classified , I came to the conclusion that actually we should classify them by using them for finding results , for obtaining results , and we can use them for discovering stuff , discovering new things , one perhaps more fit for engineering , the other more fit for scientists .

How do you view it two years later , observing the growth in this field ?

When I review a paper like this , I think much more recently I tend to be a little bit more maybe harsh , because we know AI works and we know it works very well to predict this stuff .

So having an extra model that can tell me what's the compressive strength of concrete or what the magnitude of fire would be is something that's good to have , but I don't think it's as noble as it used to be . I'm much more interested now in using AI to figure out stuff that we did not learn before discovering new knowledge .

For instance , if you think about experiment , we can experiment , or a number of experiments . We take the results and then we analyze them . We have an equation .

Ai can do the same thing , but , however , when it comes to AI , it can analyze the parameters or space of features much more effectively and it can accommodate much nonlinearity and hide the missionality pretty easily , and because of that , we might be able to figure out new things that we didn't know before , or at least see things that we haven't seen before , and

I think that's a little bit more insightful , because if we can discover something new or like a new stream of knowledge , you can go back and update equations , textbooks , information , overall very easily .

But applying this on meta analysis of databases , for example , to discover new relations or new variables that are in the play , perhaps new dimensionless numbers this could be very interesting to find .

Combination of parameters like for me , the most obvious is K-R-O-C , the capacity of heat , which defines so much in fire , and it's a combination of three variables or a material and it defines so so much . Perhaps we can use AI to find new K-R-O-Cs in things we were not looking at . Did you try something like that ? Did you have any success ?

Unraveling new ways of looking at data ?

A number of groups have tried , you know , and they are still trying to come up with different ways to measure spawning . To measure spawning , very famous professors and their groups tried that through theoretical analysis , derivations , statistical analysis , numerical simulations .

And if you start to look collectively at what we have , you would see that if we go through the derivation approach , we have to establish very , very hard assumptions and hard idealizations . That sometimes doesn't seem realistic to have because we can't really , for instance , control a lot of things , but we have to assume them .

And numerical analysis by default has limitations because a lot of the software we use are based on the theoretical derivations we have . So statistical analysis can be something interesting to go for . However , to use high-order or nonlinear statistical methods , we have to struggle a little bit with data transformations and all those things .

So it just made a lot of sense to use AI , because when we use AI , we don't have to worry that much on the ever distributions or on transformations or on the nonlinear relationships between parameters or theoretical derivations .

It's basically saying okay , we have now this furnace and we're applying some tests and we take some empirical data and this is what we learn from data . Except with AI it can be a little bit more accurate . And if you think about the problem , as you mentioned earlier , this is really a two-way problem .

For engineers , what we care about mostly is would this concrete a spawn or not ? For scientists , we would like to figure out how would this spawn and why and how can we prevent it .

So to come up with a practical solution , we can just say , with this AI model now kind of a deep spawning very accurately , which we show it can we have a number of algorithms and very , very high accuracy across multiple of data points , thousands of data points , and we show it works .

Now we're trying to figure out why does it happen and if we know why does it happen through AI , what can we do to prevent it ? Versus , you know , if you open a building , it would say USPP , fiber this much , or use less water moisture content this much percentage , and it should be good to go .

In reality , this doesn't seem to work as well , but this is the one thing that we know that can help us innovate the threat of spawning .

And by placing values on those you could eventually reveal a probability at which the spawning may occur or may not , based on the results of large amount of experiments I think it was a few hundred , like 300 issues , correct ? Yeah ? And now , very recently , you've published a paper on 1000 tests that you've investigated in spawning .

I wonder , like how did your view evolve from the first graphical heuristics paper to the new one where you present new observations related to spawning and what you've learned along the way ?

This is a very interesting case study . Let's try to dig even deeper , like how did you approach the first one and what did you change going to the second one ?

We can see in the test it would spout or not , yes or no . And if you collect as many data as we could at that point in time I think we had 300 , maybe 290 something specimens Maybe we can teach AI and you can come up with something very simple , something that doesn't require coding , that doesn't say Python or R .

It's like a page that you can , you know , fill in , connect the dots and gives you an answer . So my student , arash , did that and it was successful . And then we said , okay , well , now this was 2 years ago .

We continued the research and we were able to collect much more data points about 1000 something data points and this is where my other student more picked up the project and , to be honest with you , when we tested the 300 specimen versus the 1000 specimen , the normal grams did not change as much , which tells us that adding more data points is a lot beneficial ,

but somehow the prediction of this phenomenon doesn't change as much if we added more data points , which in a way tells us that maybe to figure out the spawning mechanism , looking at the key features or the key parameters , we might be able to predict if spawning would take place or not .

Now , as we all know , there's a big difference between predicting something if it's in the habit or not versus knowing if something will happen or not . Now , the knowing part is where the causality comes in . The prediction part is just an empirical analysis . It just happens to be carried out using AI versus regression .

We use AI and the predictions from AI exceed all of them , most of them , if not all of the theories . The second portion is with adding more data did not seem to change AI's logic as much , which is extremely important , because one of the things we always , I mean , tackled with reviewers were what if you add more data points ?

What if you add more data points ?

If you use regression , you'd be always confined with regression principles , for instance , linearity hour square is hour square certain things we can't really go away from . If we use a nonlinear method , you can be applying nonlinear method . However , you still be confined with the assumptions used to create this method .

So what we tried to say to do is let's minimize any assumptions regarding this method as much as we can and maybe we can find something new , and let's start with prediction first , if we can predict , and then worry about the mechanisms .

And now we're worrying about the mechanism because we need to figure out like , what are the mechanisms , at least from an AI perspective , that can , if we know , maybe we can go back to the theory and we say , well , this theory has a very good benchmark or has a very good foundation .

However , if we now add them data we got or the information we got from AI basically , on data-driven analysis , this is how we can make it better or this is how we can fill in the gaps that the theory did not treat different because at the end of the day , as we all know , all of these theories were done based on specific type of mixes , specific type of

tests , specific number of specimens . Nobody has enough . Here would be analyzed a thousand specimens or 500 specimens .

You have used it to explore the results of this huge database , asking questions like how does compressive strength affect that ? How much times we've seen spilling ? I'm really astounded in how it can make sense with such a scatter of variables , such a diverse amount of things that are going into and provide meaningful outcomes . I also really like how it's used .

There is this flag needs future research . I love that . Actually , tell me more about this need future research tag for some of the variables .

We also have some endogenous factors , for instance concrete makes , moisture content , existence of PV , fibers and so on . So if we start to apply our domain knowledge on the data we have , then our data theory analysis becomes a little bit in a higher level . It's no wonder just numbers , now we have some physics in it .

Now , if we add a third layer , if we establish now the relationships that we at least know although they might be 100% , complete or full , but some limit of relationships now we're going to grow from domain vision into including physics .

Now , if we have this physics and then we add some relationships from statistical analysis , let's say on causal inference , we grow into a causal analysis and with each layer we discover some new stuff . Now the very basic layer is just predict . For instance , if you think about kids or pets , if you open the fridge , they would expect a treat .

This is association . This is what prediction is in medical analysis . You do something multiple times , you get to know what would happen once you do it again . However , once you start now to link well , now I'm going to open the fridge but I'm going to take water .

They want to expect a treat , because now you have water you don't hold something sweet or something that they would like .

So you start to add a little bit more layers under the features , and the more layers of knowledge you add to the features , the better it becomes overall for analysis , because now the algorithm starts to link things much more cooperatively together .

Has this developed ? Because now you say that going from 300 to 1,000 did not necessarily change the qualitative prediction , though perhaps if there was an outlier in that , it could . We won't know that for sure .

25 to 30 observation per variable seems to be the bare minimum that you would like to at least run a machine learning model that one can confidently say this model now we have some confidence in it . The more data points we have , of course , the more the confidence or the reliability becomes .

However , I would like also to touch on one point of this one , this aspect , and I thank you for bringing it .

I think one of the things that we tend to miss a little bit is our problems in engineering , or , let's say , in fire or spawning , the number of parameters is really confined , but for us , if you think about concrete mixes , we don't really have mixes that have very , very different proportions of materials than the other mixes , because when we build the structures ,

we have to have the mix that can still have steady strength . It still has the other parameters in there . So our space of parameters is not really that very that much and because it doesn't vary as much , we may not need as many data points as we do .

However , if we have a space parameter that varies across all the parameters , let's say width and length or the size of data points , then adding more data points would most likely show us some difference in AI prediction would affect it .

If you look at AIs and medical research , for instance , when the papers analyze , I don't know , 50,000 genes or 100,000 genes of 100,000 people , the number of parameters is so massive because their problem is this massive and hence any change in the parameter and the data points will significantly affect their problem .

In our case , we have a fixed number of parameters and it doesn't vary as much just because we have to have them .

And here , okay , we can talk about the mega Pascal's of the strength and stuff like that . But how about the time delay between the next layer was cost the moisture of the air when it was done ? Was it windy or not , was it sunny or not ?

You know , and you would not normally record these properties of the surrounding world until you understand they can make a difference . And now we start to understand that because there are papers on every single the things I've mentioned and we know that the same in fire experiments .

Let's assume we investigate mass timber experiments , but how many of them would tell you if it was sunny or not today they carried out the experiment , or how often you would consider this being a variable . So I think there's also a trap , you know , of the fact that we recall the things we know that make a difference , but it's not a complete list .

But this also brings to the ability to do sensitivity analysis for different parameters , like you have done for the spalling , and if you see there's a parameter on which the variability is great , or if there's like an unusual distribution , like it can be a peak at the end it can be even result over the whole distribution , you know , it can show you where to

dig further and where to seek more answers and then then perhaps this machine learned fueled approach can help you dive deep .

I mean we just record , for instance , the strength , and you know that when the fire , when the collapse happens or when failure happens , we don't record other things , like you know whether , maybe with the conditions , as much of the For the internal sizing of the furnace , the type of sparrows , or in the middle gas using an umbrella , and all these things are

our parameters that we have to think about . And when we think about this , of course , the more we document , the better our models become , all the better .

But at the same , at the same and at the same point , I don't want to say I'm not trying to do predict you , I'm just trying to say Engineers tend to be a little practical and because of this practicality we have to say , okay , we have to establish now , this parameter does not influence , but we don't know , we don't know how much it does influence in the

first place , and because we just have this hard assumption that it doesn't , and it doesn't what doesn't involve anything . So it may hinder our analysis a little bit and because it may have the right we can't really figure out why but that the good news is , with the parameters we have so far , and I'm talking only on spot .

With the parameters we have so far we showed Very good accuracy and confidence in predicting spot . So perhaps adding more parameters may increase our accuracy a little bit . But the accuracy we have so far is almost 90 percent . So if you think about it , out of 10 samples you can predict like very easily . And predicting 10 out of 10 may seem would be great .

But I don't know if we can confidently say since they saw and this is what . Maybe we have to start to think about other ways of analyzing our models , but it's a very important point .

One thing that is really interesting to me is this coding free versus coding based Approach to machine learning . If you could tell me what do you mean by coding free ? I thought it's always coding .

What this really started from , from thinking about what to do with the AI in civil engineering , like , for instance , a lot of in many , many universities , we don't have coding courses .

We may have a course at the second year or first year that say this is a programming course , and students usually take it the reason freshmen or sophomores but we don't have a follow up courses .

So when I try to do my research and hire students that are familiar with the AI , I usually tend to struggle because I have students that are very , very , very good engineers but they're not really good in coding and if you take them you're going to have to wait until they learn coding . So usually we have a delay to start .

So once I start , we start to look at this problem , we said , okay , well , we don't really have to use AI with coding all the time . That have a high algorithm built in them and they have a very nice user interface and all what you have to do is upload the data .

And when I use , for instance , algorithm X or algorithm Y or algorithm C , the software trust the algorithm for you . Here's your the output . So you're using AI , you don't have to code for it , and it's your job now to analyze data and figure out what to do next . So this opened up the door for learning AI through coding or through coding free software .

My preferences for maybe engineers who are well established senior engineers or they have lived in school is to learn the principles of AI , the big ideas behind AI , and then use coding free software to use AI . For younger generations , learning coding is something beneficial because it's like a new language .

However , as we all know , with chat , jbt , code interpreter , with the data ethics , the reliance on coding , at least for our civil engineers , is more likely to be less and less , because if you can have a tool that you can , say , use XGB both to analyze this data without having to code it yourself , that'd be substantial because you don't have to waste time

and resources learning to code plus using all this coding stuff .

Well , the thing out there is still that you can use non-coding based software , and that's also a brilliant thing . So , in the same way , how we use CFD software today without having to program navier soaks on our computers anymore , because there exists a package that does it for you .

I also know an important thing for you was the explainability in AI , and we've briefly discussed this in the previous episode . Maybe you can tell me again a little more about what do you mean by explainable AI and what's the purpose of having that ? So the goal ?

And the intuition was if I know how the model was able to come up with a prediction , maybe I can uncover the missing relationships that I don't really have now , or the physical relationships that I don't have . So this was a promise and I think the promise continues

to be this way . However , the explainability tools we have so far for instance , line Chab , all these different types of tools what they do is they show you how the model was able to come up with the prediction using your dataset . So it uses your dataset to show you how the model arrived at the prediction .

It doesn't show you why the model arrived at the prediction . For instance , you have five data points and the model predicts , let's say , a number , 50 . It would say , from these five data points , it uses data point 10% , 20% , 30% . So now we're 60% , 20% for the last two points , and this these come up with a 50 value for the prediction .

But knowing this doesn't really tell me anything about the physics of the data or the physics of the problem . It just explains numbers with numbers . Now , this could be beneficial , for instance , because I can go back to the most important features used to come up with the prediction and maybe there is a physical meaning behind it .

However , there is no guarantee that there is actually a physical meaning behind it , and hence explainability continues to be something important . But for us , for dealing with our problems , it may not clear the promise that it initially had , because we still don't know the physics or the missing relationships .

We still don't know the physics behind the problem and most of these explainability methods .

Unfortunately , they still assume the relationship between the features to be some form of a linear relationship , and we know that this is a very hard assumption to have , because relationships between parameters and engineering problems are likely to be a nonlinear than being a linear .

Is it something that can be used to check if the prediction happened within the physics realm ?

However , the problem is we're still using domain judgment to answer this question and we still don't know the relationships between the features . So it's a way to check , but I wouldn't say that this that checking this would imply that the algorithm actually understands the problem .

In real engineering you will have this one mixture you don't have data for and you want to figure out to what extent this is going to spoil or not . So understanding that the model is not only giving you a prediction , but its reasoning is within the limits of the physics , perhaps give you a little more trust towards the prediction itself .

Because if I knew that the model predicts it perfectly , it just assumes the concrete turns into plasma in the meantime . I'm not sure if that's a prediction I would like , even if I know it's 90% on the data set I used to train , 100% , I agree . And the second thing that comes together with explainability the interpretability . What about that ?

Interpretability you take the algorithm and it tells you how not the prediction , but what's the mechanism within the algorithm but arrived at this prediction . So for us the technicality might not be something that's extremely important for us , because for us we just want to know why , for instance , or how come the algorithm come up with this number .

But for people who create algorithms , knowing the mechanism itself also can make a bit different , because , for instance , a neural network doesn't work as well as another algorithm X or Y or Z . So if you can't constrain the mechanism , you'll be able to know why .

On the other hand , if you know the relation , if you know how the value was arrived at from the data , you'll also know how .

It just gives you values and a linear relationship between

the parameters . Cosnary tries to figure out what's the map , what's the DNA of this problem between these parameters .

So , for instance , if I know from causality that parameter one causes two and two causes three and three causes the predictions , then I might as well , instead of having one , two and three on parallel rows to each other , because I only need one , because I only need the one that actually predicts the phenomena , and if I know the relationship between the parameters

, then I can easily come up with a hypothesis I would say okay , now , instead of having five parameters , that gives me a prediction , I know that two of them are outside the parameters and I don't need them that much , but these two cause the other three parameters and now I can confidently use the three parameters to predict something very easily , because I have

the map , so it is kind of like a forced neural network from the start .

Right Now , because multiple parameters could interact and you could end up with multiple maps . It becomes now which one of these maps is more physically possible Fantastic . And from there you come with a hypothesis . You go back and you test and then you come up with a new feeling , for instance , that says okay , well , now I have this map .

Now the other interesting thing is , if you have the map , all what you have to do now is to change one parameter on this map to know if you're going to have a prediction or not . But the difference between changing this map and changing a model , my map has a lot of physics , has a lot of domain knowledge .

Machine learning is based on data and machine learning gives you explainability after the analysis that before the analysis . Okay , so you can't really control machine learning . You can control the causality because at least before doing the analysis , you have a good way of inferring what would happen to this map .

And what you said about testing if parameters are within physical limits that that's what my friend , mihail Malandowski did when he was solving the adiabatic surface temperature equation , because it was in the fourth power and everyone was solving that as an iterative way because it's kind of annoying , it has multiple roots and Mihail just solved for all of the roots

and just he was picking like , okay , this one is on physical because it has a negative energy , this one has negative temperature not really possible and one by one eliminating which actually makes sense in the world of physics , in the physical realm . He was able to figure out okay , wow , this one root is really making sense and it simply works .

So do you have an example where this was used in practice or is something very new ?

And it was something interesting because if you know how things happened and why , maybe we don't have to spend as many resources on experiments , as maybe we can do a less number of experiments . So there are a big causal inference is a very big area in social sciences plus quantum science . There is a lot of professors who do research there .

A lot of the causality stuff comes in medical for instance . If you have a random trial and you have a drug with a vaccine for instance , some experiments can be un-efficant . You can't give people a drug and it might be harmful for them . So you have to use causal inference to figure out how to analyze the data .

And then we discover some maps and then maybe figure out a way to treat other patients . You also , it could be very expensive . You can't have an experiment on 10 million people . It's something for our stuff . We can't test 100 beams in a furnace because it's extremely expensive .

So those limitations will bring causality and much of the research is in social science . How can you control income ? How can you infer education for students with to do and say , low income neighborhoods and the policies ? A lot of these things don't really have an experiment for them and we have to rely on causality to come up with .

And as we were heading till the end of the interview last time when we talked , my big takeaway from the discussion was that if you want to apply AI to a very general problem , you're not going to have a success because it's not like . The amount of data and variables and everything will simply prevent you from obtaining a very general solution .

But if you are capable of narrowing down your issue and then finding an algorithm suitable for that issue , ai generally tends to do a great job . Now I wonder if you have some examples of perhaps failed or unsuccessful applications of AI so we can understand . Everyone is talking about the success .

It's very rarely that someone publishes a failed attempt , so perhaps you can give me behind the curtain pick on what didn't work , so people listening can avoid traps or the simple mistakes that go to

black holes .

He has a very recent paper hopefully it will be accepted where we actually show that if you have multiple algorithms and you take one leading algorithm that everybody's talking about and you apply it to the spotting problem , this algorithm will give you very , very good predictions with physically impossible features .

So , although you have all the features that you want and the algorithm can give you very high predictions , it will still fail to give you something that we can all agree on . That causes this phenomenon .

Other algorithm and this is deep learning , by the way , everybody's talking about deep learning if you take very simple algorithms or like algorithms that are a little bit more modern , they can do a much better job . So , when it comes to a big item from AI is never rely on one algorithm , never rely on one model , never rely on one dataset .

Do your best to use this ability , do your utmost best to use some form of causality and make sure you understand the principles very , very well , because you don't want to just come up with a number at this number , although it's correct , it can't come up from a physically impossible method or way that you can't really figure out .

And if one number can come up like this , there is a potential for the same algorithm , same mechanism , to give you other numbers .

And when it comes to data points , when we use we've tried a few multiple datasets and what we found is if you use 15 parameters or 45 parameters or five parameters , there is a very good chance that the same algorithm will give you more or less the same accuracy , which makes a lot of sense because in medicine they have 50,000 parameters and the models seem to

work . So it's going to almost come down to education . You have to really understand the principles . Don't be afraid to use coding . It would be maybe a little bit easier to use non-coding methods , but the principles will always be , will always be with you . I'm going to end with this point because I think it's important , if you allow me one minute .

Yeah , of course , when we teach finite element in courses , we use what we teach 1D and 2D elements . We don't really teach 3D elements , but when we graduate , we use softwares . Most of them have 3D elements , and we learn 3D elements from the software , not from the course .

Same thing with AI , we can learn the principles in school and then , once we're out there in the field , we'll relay it back to the principles , because these will never change .

I am very sorry for butchering this so many times in the episode and that you had to witness that , but hopefully it was worth it . I think the knowledge brought by MZ is just amazing and the best thing is he's leading by example . The point is that he can do stuff like this as well .

We didn't even explore that much how he is exploring all test results and big databases with the use of AI methods . It's also very intriguing . If you jump into the previous episode , you'll learn why perhaps you should not take all of the data available and why results from 50 years ago may not be that representative for the today's situation .

But overall it is a brilliant way to study the history of fine safety engineering and you know what Many times in long forgotten experiments there still is knowledge to be found . You , as someone who's running science , who's doing projects , who's doing research , who's designing stuff , you have no idea . I have no idea what is in the work I do or what we do .

There are always hidden things that we are unaware of , and there has to come someone who will show and point it to us when it becomes kind of obvious , and this is how the best breakthroughs in science happen . So thank you very much for listening . I hope this episode was interesting and inspiring for you .

There is a book to be read and for the podcast news , it just has passed 100,000 individual episode downloads . That is a big number . Thank you very much for listening to the Fire Science Show and for being here with me .

It means a lot for me to have an audience , to know that on the other side of the microphone there is someone listening and enjoying this . I appreciate you a lot and yeah , let's go up to a million . That's the next girl . Thanks for being here with me today and yeah , that was the billion . See you here again next Wednesday . Cheers , bye .

This was the Fire Science Show . Thank you for listening and see you soon .