¶ AI in Fire Safety Engineering
Hello everybody , welcome to the Fire Science Show . Today we're going to talk AI in fire safety engineering and I am very excited because we're not going to hypothesize what the use of AI will look like . We're going to talk about experiences in doing that and that's something very unique .
With my guest Professor Sinian Huang from Hong Kong Polytechnic University , we've hypothesized how it could look like and talked about his early experiences three years ago in episode seven .
Actually , that was one of the first episodes of the Fire Science Show and today , three years fast forward , we can talk about a lot more experiences that Simeon and his group has gained over those years . They're on the forefront of implementing AI in various kinds of fire science . And we're not talking chatbot AI .
We're talking neural networks and using it to predict fire behaviors , to predict fire phenomena , to measure fire and to help in engineering design .
It all started with smart firefighting , so the theme of episode seven and the discussion back then was how we can use AI to assist firefighters and , if you're curious how this ended , it's at the same time interesting , to some extent disappointing and , at the same time , exciting .
Some pathways did not lead anywhere , but some pathways lead to extremely exciting places and we're also going to talk about that later in the episode , and if you want to learn about that , well , first intro and then let's go with the AI in fire safety engineering . Welcome to the Fireiresize Show . My name is Wojciech Wigrzyński and I will be your host .
This podcast is brought to you in collaboration with OFR Consultants . Ofr is the UK's leading fire risk consultancy . Its globally established team has developed a reputation for preeminent fire engineering expertise , with colleagues working across the world to help protect people , property and environment .
Established in the UK in 2016 as a startup business of two highly experienced fire engineering consultants , the business has grown phenomenally in just seven years , with offices across the country in seven locations , from Edinburgh to Bath , and now employing more than 100 professionals .
Colleagues are on a mission to continually explore the challenges that fire creates for clients and society , applying the best research , experience and diligence for effective , tailored fire safety solutions . In 2024 , ofr will grow its team once more and is always keen to hear from industry professionals who would like to collaborate on fire safety futures this year .
Get in touch at OFRConsultantscom . Hello everybody , welcome to the Fire Science Show . I am here today with Professor Sinan Huang from Hong Kong Polytechnic University . Hey , sinan Hi Bozsik , good to see you again . Very happy to have you back in the podcast . You were one of the first 10 episodes . I can call you the OG of the Fire Science Show .
Yeah , I'm F-7 . Yeah , nice 007 , licensed to do AI in fire safety engineering . Good , good , and we're going to continue the discussion that stopped in episode seven . Gosh , that's a long time ago , but in that episode we've discussed smart firefighting and different ways of using AI to assist firefighters .
I remember you were very happy back then by a large grant that your unit was given on this topic . I know a lot of papers came out of your office . So now , fast forward . Three years have passed . Let's see where we are on smart firefighting today .
So if you can summarize , what's the biggest change between 21 and 24 in terms of using AI in assisting firefighters ?
Yeah , I think the last time when we first talked about the smart firefighting back 2021 , and we just get that grant and I just recruit a few students doing this project . Also , I'm new to AI . Right now I think I'm still new to AI .
I haven't got a chance to really run AI as all the students doing the hard works , but I think I know more about the AI application and how it works and what problem AI can solve to help us , no matter it's to support firefighters or support 5G news . So I think I know more about the tool .
The last time we talked , you also mentioned that the use of AI comes from patches of code that are implementing real packages . You don't have to be an AI scientist in implementing AI . Now , working with this for three years together with your students , was it very hard to enter the world of AI ? I mean , three years have passed and you've shown some success .
I wonder if that can inspire others to try out AI . Was it very hard to implement a lot of challenges ?
I think it's very simple to use the AI algorithm . So most of my students who have most of them have never used AI in their research before , or maybe they haven't done any research before , so when they start to do the AI-related research , I think it takes less than one month to be able to run some simple AI program or reproduce some previous paper .
So I think running the AI algorithm itself is also a challenge . The most challenging thing is to identify the problem that is worth solving by AI , and that requires a lot of knowledge about the file as well as the capacity of AI .
And how about choosing the correct algorithm ? Because with my understanding of AI , I understand that there are supervised and unsupervised models . You have neural networks , but you have also classification tools . There is so much like when you start digging . There are so many choices .
Have you figured out the way how to assign correct algorithm or model to a correct problem ? Or is it the other way you know an algorithm and you find a problem for it ?
Yes , that's also something we are learning during the process of this project , and overall feeling is I still think the algorithm is not so important . First , it can be solved . Most of the algorithm problem people are facing can be solved by increasing the size of the database or the number of useful data .
Sometimes you have a very large database but most of the data could be not valuable , so not really helping training AI . But if you have a good database , even if it's small , it can basically solve most of the training problems . That's my feeling . It can basically solve most of the training problems . That's my feeling In some aspect .
For example , we are doing a lot of file simulation . We are predicting the file development showing the smoke movement . So if you want to use AI to generate very nice or very real CFD file images , then the algorithm is quite important and from my experience , diffusion model is definitely the best to generate very detailed flow motions , smoke motions .
But the problem is the training of the diffusion model is very , very long time and even when you use it for prediction , the rendering time of these images the AI prediction images also very long time . And even when you use it for prediction , the rendering time of these images the AI prediction images also vary in long time .
So unless you really want to achieve that detailed structures , usually you don't need a diffusion model . In my experience , some models like GAN model or GL model they all solve pretty good problems and , for example , if you just want to predict the ASET , you just want to know when the smoke layer will drop to two meter high .
Then you don't need to know the detailed flow structure of the smoke , you just need to know when the smoke layer touch the critical line . So in that sense there's really no need of using advanced AI model .
It's really interesting because your group was able to use AI in a way like many of us would use CFD . I see so much resemblance , by the way , how you are using AI with how many engineers would be using CFD .
Also , you know , in the way that we don't really comprehend CFD that well as engineers like , and you don't need to because there are already made packages that you apply . You also have to understand the problem definition and set the boundary conditions . Like you said here , you choose the appropriate model or appropriate tools for the problem you solve Now in CFD .
I would say one of the reasons it is such a popular tool in fire safety engineering is because of the realism of output . You get those really beautifully looking plots that look like fire . People use CFD because they get those lovely images and everything .
You just said that you can use AI very quickly to just get the ACID value of two meters and yes , I agree you can do that . But the question is the perceived value of the tool . If , if you lose the beautiful images generated by the diffusion model , will that tool be believed that it truly is two meters ?
You know , without the layer of the graphics , it's hard to convince someone that the result really is the two meters , and it's all a matter of trusting the tool which , let's be honest , the trust to AI generated results is , in general , quite low , I would say .
Yes , that's the reason .
A lot of FDS simulations can cheat the public because they look real , and I'm sure the software companies or SmokeView or other rendering tools , they have tried a lot to make these images look real and so far what we can do is first , of course , we can prove that even a rough smoke layer without all these small eddies , they basically have the same smoke height
as the one with detailed eddies . And that's one way . The other way is we can also use AI tools to generate these eddies to make it more real If it's required to convince people .
Okay , it's just a matter of cost and time , right ? I am disturbed by the amount of similarities with AI and CFD as a tool , and I wonder how the future will look like . Will we be using one or another , or both ?
Yeah , if you print that in paper , probably you cannot tell the difference whether it's AI generated or it's CFD generated .
It's disturbing because I know at least cell biology and genetics . There were hundreds of studies that were retracted with very serious accusations that you know the genome sequencing is like a line with some lines in it , so it's very easy to fire up Adobe Photoshop and just cheat a part of the image .
Serious scientists would face serious accusations that the images in there these are falsified .
They cannot blame that for AI . That's actually a real human cheating .
Yeah , I know , I know , I know , but I'm stressed because the ease of this tool and the ease will increase in the future . And what if the AI is wrong ? What if it gives us incorrect output because someone did choose the train set as a too small one ? How sensitive actually was it to data ? You said the biggest challenges were with getting data correct .
Was it truly one of the biggest challenges in your project to get the data and how much of the data you actually needed to get reasonable outcomes ?
Depending on the problems you're trying to solve . For example , recently we are trying to use AI to forecast smoke flow in very complex shape atrium . In that sense , we need a lot of simulation CFD simulation to form the database . But even so , I don't think that database is that big . We only have a few hundred case with complex shapes .
We have another few thousand case with relatively simple shape . So I don't think that's large enough . Because essentially what we want to do is I mean , nowadays so many consulting companies , they are running CFD simulations for different buildings , different structures , but all these data are not fully used . After they finish the project it's in the hard drive .
Nobody is really using it . But in fact , all this data can be trained for improving the AI capacity . But if we have this database to train AI , that will be amazing . We have a very large database and we are not asking extra input , we're just using what's already there .
But of course , the pre-processing for these data will take a lot of time because every company , every engineer , they have their own habit of making the model to run the simulation . So that's also what we see is that in fact , a lot of time is spent on pre-processing the database rather than creating database itself .
¶ AI in Fire Engineering
If you wanted to use this data like the way you said , I think there would be so many human variables with what you just said . Different people would do it differently . I wonder if it's even possible to quantify all the choices that people do Like . There must be doses of choices Like what design fire did you put ? How do you place it ? Was the suit healed ?
Was the heat of combustion ? Was the makeup moisture ?
That's actually the good thing , because everyone chooses different parameters . That just makes the database become richer . Ah , okay , not just a few settings . So everyone has different settings , so the database becomes very good , very large ?
And how about training or integrating experimental data ? Because I know that you use CFD and you have a good reason for it , which I hope you will reveal . But how about using experimental data ? Let's start with CFD . Why do you train more on CFD than on experiments , from what I understand ?
Yes . So first of all , we still use some of the experiment data . I would say , before we do a large amount of CFD simulation , we always calibrate the model with the experiment data . So in that sense I consider we already include some of the experiment into the database , because some parameters used for the boundary condition may come from the experiment .
I wouldn't say it's 100% numerical input . It also has a lot of input from the experiment . The problem is , even for the experiment it's very difficult to quantify the result . For example , we all have limited sensors . For example , even if you have some couple of trees , you have a few points .
Even you have cameras , you have only a limited view of the smoke motion . Compared to the CFD simulations , the data you get from the experiment is extremely limited and have a large uncertainty . For example , the fire you used in the experiment may not be so well controlled .
If you are burning , for example , a wood crib , who knows what kind of large fire , how large is the fire it is , and there's a smoke . I mean , every wood burns different smokes . It's difficult to quantify that . So in that sense I feel it's quite difficult to directly use experimental data .
And I think , most importantly , for file engineering design is you only consider certain representative scenarios rather than the so-called real scenario . There's no such thing as a real scenario . Even the same building have the same furniture burning for a hundred times , every time it's completely different . So in that sense , we cannot forget .
Doing design is just following certain framework and test some possible file scenario to give certain confidence . We are not trying to simulate a potential real file .
I would summarize what you've just said is that it's hard to capture all the uncertainties in the learning process , like , if you learn based on CFD , you , if you input it one megawatt , you're certain you've inputted one megawatt .
And here , if you had a crib that was supposed to give you a megawatt one day it was very hot and dry and the lab was well ventilated , you had 1.1 . Other day you've done a repeat . It was a moist after rain . You had 0.95 , right , and yet you put something into training .
You tell the trained model it was one megawatt there's an uncertainty in the input that you've not accounted for .
So there are a few aspects . First , from AI training point of view , the experiment data have their own format . So you may have some temperature sensors done one experiment , you have some other group of sensors done in different locations in a different experiment , and these data formats don't match with each other , so it's quite difficult to train them .
If you run the CFD simulation , you can collect the data in a consistent way . Then it's much easier for AI to train them . That's one aspect , but I think the second aspect which is most important is when we do the current state-of-the-art practice , we never ask the guys who run CFD . We don't question if their model represents the real file or not .
We just assume okay , your simulation is reasonable and correct . So in that sense we only need to compare with the CFD simulation . We don't have to compare with the real experiment , because this is a design practice .
So , basically , assuming that the CFD is the state of the art tool , you basically create a tool that is at the same confidence level as the CFD . Yes , okay , that makes sense . Out of all the AI implementations you've done , let's pick one and go deeper . How about the fire prediction ? I love the fire prediction , so .
So your group has built those tools that are able to predict the size of the fire based on images , I believe , and I I found it really interesting , especially that there are videos online that showcase the real , uh , real-time capability of this prediction software , and it's just magical .
Are the videos fake or it really works like that , like real-time , showing the hit-release rate ?
Yes , we're actually having a latest paper will come out very soon , that we have an online link Everyone can upload their video and we were exporting the real-time hit rate .
That's good , so I confirm , this is really amazing
¶ AI for Fire Size Prediction
. So tell me , what was the big idea behind starting this and how did it go ? Actually , what was the point of doing this study ?
So I have to say the idea come from when I was teaching the fire dynamics class . So there's one session I have to teach the students what's the definition of the fire heat release rate and then I have to go to the only two methods that we can measure the fire heat release rate .
One is to measure the mass loss rate of the fuel and the other is oxygen calorimetry . You measure the oxygen depletion based on the smoke measurement and eventually the students questioning okay , both methods can only be used in labs . Can any method can help us to measure the real fire . So I think that's a really a good question .
Some students I don't really remember the name of the student , but someone asked me about that so I think that's something we have to think about , because if , of course , you can put a big hood above a house or above a burning car , but everything is done in the lab , you cannot put a big hood where you have a fire incident and you put it about there and
measure the heat rate . So none of the methods that we have so far can actually measure the power of a real fire . And that just inspired me to think about AI method and I think during that time we have a lot of advancement .
For example , using the mobile phone , we can use a facial ID to unlock our phone and we have a lot of facial recognition everywhere . In China you can use a face ID to pay , actually Okay , so the image is really powerful . That's what I feel . And doing experiments and all these fire experiments , we have a very rough view .
So if the fire size is an area , the volume is larger , of course it's more powerful . So I think there is a certain correlation between the size of the fire as well as its power , correlation between the size of the fire as well as its power .
And if you really look into the details of the fundamental flame sheet and that definitely makes sense because flame is essentially like a coating , it only has a small sheet and all the reactions happen in that sheet .
So if you can get to the area of that sheet , definitely you are able to quantify the file release , heat release rate , so , and that area of the sheet is proportioned to a certain degree to the image can captured by the camera .
So that's the original idea , but of course we know it's very challenging to actually train the database Then we are just super lucky . I would say super lucky because NIST had such a wonderful database .
So they are burning all different kinds of things I think they have more than 2,000 different material things burning in the lab and as they record all the heat release rate by oxygen calorimetry , they also have all the images , videos , you can just download from the website . So that's just amazing and you can use this database to train .
I'll quickly plug in . I had an interview with Matt Bundy . That's episode 110 of the Fire Science Show where we've just discussed what you've just said the NIST fire calorimetry database . So Matt told us all the tricks they have for recording cameras , automated storage , processing .
Like a lot of effort goes into building a database like that and we are very grateful to NIST for developing and maintaining this database . So you had images from , or videos from , the database . You had the heat release rate plots , what goes from these points to having a trained model that can predict fire size .
But because I assume it's not a simple like press enter , here's the images and you robot learn .
But it must be quite a process so basic for all the ai model , you need to identify the input and output and you pair them in the training . So for this specific case , the input is the image of that moment and the output is the heat release rate measured by the oxygen calorimetry . So you pair them . So every second you have a pair .
If you burn something for 20 minutes let's just take one point per second you have 1,200 pairs of heat release rate as well as images . Then you put all of them into the training database and now you have 1,000 different fire burning and each test you have 1,200 images . So together you have a million of data pairs for the training .
So together you have a million of data pairs for the training and that results in a very amazing trained AI model that can basically give you a hit-release rate if you input any image Okay , but the image is a collection of pixels .
It doesn't reflect the real world . If you put a matchstick right next to the camera , it's going to appear huge on the image . So how did you solve the dimensionality of the fires at NIST ? Or were the cameras conveniently positioned ? Always the same way ?
Yeah , first we removed those cases that clearly the camera is putting a different location and later on we added some additional data from our lab to measure the fire from different distance and use that to calibrate the fire image . So we rescale the image to be the same as our database .
So in the database all the images are rescaled under a certain scale and for any practical applications . I think there are three methods you can approach to solve the distance or the scale problem . First , you can use the reference lens .
For example , if you see some fire is burning in a car passenger car and you know roughly how long the passenger car , how tall it is , you can use that as a reference scale to help you to scale that fire image . Use that as a reference scale to help you to scale that fire image . That's one thing . The other is you can use a bimolecular camera .
You have basically two cameras that can measure the distance between the camera and the fire and that can also give you a reference scale . And as a third option we provide is if you put that in UAV and the UAV can measure the height between the ground and a UAV , that actually give you a reference lens as well .
So , depending on the application , you can always find a good reference lens to help you solve the distance issue .
So , so this is probably the challenging part of modeling . And do you see any ? Okay , because , fueling the curiosity , you see any practical outputs . How helpful do you believe this could be more to scientists , more to firefighters ?
I think there are potential applications . For example , initially we make this system trying to help the firefighters . For example , if one fire happens and someone may be passing by the fire and they can just use their phone to record .
In fact a lot of people when they see the fire they will start to record , they will do streaming in the cloud server so people can see it .
So if the firefighters can get access of this data in the fire engine on their way to the fire engine , they can know the initial fire development , they can understand what's the original condition of the fire or what's the original fire source . I think that helps the firefighting operation to a certain
¶ Advancing Fire Safety With AI
degree . I think it also can help the fire scientists to do some research , although I haven't really used it yet , but I think in some cases not everyone has a nice oxygen calorimetry in the lab . So in a random past we burn something , but sometimes it's difficult to quantify their heat rates .
And if we just record the fire videos and using our algorithm you can have a rough idea about the evolution of the heat release rate . And that is much cheaper compared to install a big hood to measure the heat release rate .
Even though it would be uncertain and some rough estimate . I think also for some field experiments it would be uncertain and some rough estimate , I think also for some field experiments . It's often a difficult variable to get a hold of , even approximate . What was the heat release rate during an experiment ?
Or you could perhaps have a footage from a real world fire incident . You know how far the smoke penetrated the building , you know where it ended up . You want to do some forensic analysis and all you have is a video from the scene . Perhaps that can give you a data point you know , to match your simulation with the incident itself .
I think there could be a lot of places where this could be used . I also know you were trying to figure out the sizes of fires based on temperature measurements . I recall a paper related to tunnels . So how did that differentiate from the images ?
So inside the tunnel . Essentially it's very challenging to get a full image of the tunnel fire . For example , your CCTV camera is installed a certain distance away from the fire source so of course if you are lucky you can capture the initial development of the fire .
But most of the tunnel fire are in very confined space so it generates very heavy smoke so your camera doesn't work quite very quickly . Instead you always have the thermocouple , sometimes the optical fibers , to measure the temperature , so linear heat detector . So we can assume you always have good temperature measurement in the tunnel .
Then with this temperature sensor you are able to predict the evolution of the fire , essentially the fire size . I'm not talking about the location of the fire , because everyone knows where the location it is . It's just near the sensor which has the highest temperature . This is also related to some question .
A lot of people ask what kind of problem in fire is worth to be solved by fire . Because I see a lot of papers . They use temperature sensor inside a tunnel to predict the location of the fire .
A lot of paper on this topic which I feel is just meaningless because we all know the fire is just right next to the thermocouples which has the highest temperature measurement , even if you don't use any AI method .
You know it's near there and if you use the AI method to improve the position , maybe one meter or half meter more accurate , it's meaningless , but still a lot of people are doing that . So I think AI is a good tool , but you have to use it to solve the right problem , not abuse it .
There's a lot of questions you don't , a lot of problems you don't need AI to solve .
I think that's an important statement . I also see that in the world of general science , engineering sciences and AI , that people are applying AI for trivial problems and claim that it's a new solution , whereas a really good solution existed and there was everything fine with that . Ai doesn't bring you any novelty .
Sometimes it would be put into papers or research just because it's , you know , innovative . It gives you this novelty factor and for many years already I've in my reviews and I see that from other reviewers that the tendency is that just using AI in your paper is not yet enough to claim the novelty . You know that's not enough .
You have to solve something important with the use of it , has to bring the new value . It has to bring a new how to say it gain over the existing methods . Where do you see those gains ? Like you do a lot of AI , where do you see the real benefits of employing those methods ? Let's say , the finding the fire in tunnel is a bad application .
Where do you see the good applications ?
Some problem I think it is worth using AI to solve is transient fire problem , so things related to time or complex space . So because all the conventional fire dynamics textbook or the lectures we are given is using analytical solutions , but the analytical solutions can only solve low-dimensional problems .
I have looked at all the classical equations we have in our community , for example Cargill's correlation , arbert's correlation or the ignition temperature , flame spread rate . All these classical equations at most have two-dimensional , one-dimensional in time , one dimensional in space , or two space dimensions . There's no equations can include more than three .
If it include more than three , the equation looks extremely ugly . It has a lot of coefficient trying to fit it into a certain special cases . In fact , that's a lot of people are doing . They do a certain test and are trying to modify these classical correlations to make it explain some special cases , but all of them just make that equation so ugly .
In fact , when I think more about it , it's a fundamental problem . Yes , these kind of polynomial expressions are not good to describe a high-dimensional problem . It's just not possible . For example , we solved the CFD simulation . It's a high-dimensional problem . It's just not possible . For example , we solved the CFD simulation .
It's essentially four-dimensional , at least four-dimensional . So you have time one-dimensional , time three-dimensional scale . You cannot write the equation of a temperature distribution , even you know the solution , because the solution is in numerical . You cannot use a polynomial to write down the temperature distribution . You can fit it .
You can fit it in a super ugly way because the polynomial is not designed to describe high dimensional problems . So instead , how can we describe the high dimensional problem ? We actually use a neural network . The neural network is essentially a kind of expression .
It's in parallel to the polynomial just to describe a high dimensional problem and with that tools we are able to consider additional dimension , for example transient problem . When I first working on the AI , I start with a tunnel because I feel a tunnel is relatively simple One dimension right , one dimension yes . And I searched all the literature .
I tried to find the data . If the data helps me to do the prediction of the tunnel fire , essentially no , because there is no transient data in the literature . No one will plot the time evolution of the fire . They always pick up the steady state point . They just ignore all the transient point .
Because I already have so many datas , even the steady state point is already a lot to plot . Same even you're trying to plot something , you can only include low dimensional information . So for high dimensional information , you cannot plot it . So ignore the time dimension is an automatic choice .
But if you don't have the time information , how can you predict , how can you forecast the time evolution of the file ? So that's basically . After searching all the literature on the tunnel file , I found none of this data is useful for the AI .
That's the reason I started to run the CFD simulation to create my own database to include the time evolution data for the training . Otherwise it's just using steady state . You cannot predict a varying fire , right ?
I think you're very right in here , and especially that you would always , or very often , consider fire safety as a factor of time . A timeline approach is a very fundamental to fire safety . You know ACID also . That's the thing that engineers would mostly do .
I think you're quite correct in your assessment that this is a space where those new approaches can actually shine and where the existing approaches are perhaps unsufficient .
¶ Challenges in Firefighter AI Communication
One more thing I wanted to ask you , because we've started this talking about firefighters and you know it was smart firefighting as the theme of your research . I wonder how that went , the connection of these new tools and the firefighting , because from our discussions , there's much less firefighter context in what we just discussed than I would have expected .
I assume it was challenging . Yeah , it is super challenging .
Initially , when we started this project , we went to the firefighting training centers and we worked with firefighters to do some tests together .
It all started very well , but later on , when we are trying to do something or give some advice about the firefighting operation , then that conversation went to the dead end , because it seems like we don't know much about what firefighters do and the firefighters think we are not giving them anything useful in their daily practice . So it's quite a challenge .
In fact , I feel a lot of fire scientists . They don't really understand the firefighters' work or they don't talk to the firefighters that much , and it seems like there are two different communities and it's difficult to break the boundaries .
And once , after one or two years , I realized that the problem may be lies in the data , because we don't have the data sensors in the buildings . So even our algorithm is very powerful . If we don't have the real time data feeding these AI algorithms , there's no way you can do fire forecasting . So that's the first challenge we found .
And then I look at the building . Oh , we have so many smoke detectors . Sometimes we have temperature sensors . Can we get these sensor data ? And that's another challenge we met . Basically , it's so difficult to get these sensor data , for example the smoke detectors . They only give you yes and no , so zero and one information .
Even you can get these information , you cannot extract the data from sensor directly . They all go to a control panel and the control panel data can only access by the software developed by that company . Every company has their own barriers to preventing other companies to accessing the data .
From the firefighting point of view , from the data point of view , that's a disaster . Basically you cannot get any data . So in the end we just install our own sensor and create our own system . That unfortunately cannot be accessed by other groups either . So there's some challenges .
I think , and thank you for sharing that very openly , but I think you hit the nail on the head with the issues communicating our societies and there's definitely a gap between the fire science as understood by us and the fire science as understood by fire scientists , who are fighter fighters and they also do their own fire science , you know , and that is not
necessarily the exactly same science , even though we're talking about the exact same phenomenon of fire in buildings . Indeed , it is something that I find very intriguing to nail down why . Why is that ? Why ? Why cannot we find a common language ? Why cannot we find the solutions ? And I also see that we don't have the background and experience of firefighters .
We don't really understand how firefighting works and for firefighters , they perhaps don't see our view on the understanding of fires . I've noticed they speak different language , they focus on different phenomena , they pay attention to other things that we would pay attention to , like we would pay a lot of attention to heat release rate .
Your algorithm is one example of how much effort we can put into knowing the heat release rate , because , from my perspective , that's the most important variable I can learn . I wonder what value would a heat release rate have to firefighter .
Like if a firefighter knew there's like seven and a half megawatts in front of me , what true value would it give to them ? I know there are approaches that allow you to quantify the amount of water you need to extinguish based on approximation of heat release rate , but that would be it . I don't know any bigger achievements .
If you've shown them those tools , what was the reaction ? Yeah , it's cool and that's it . Or they've seen a practical use for that .
It's useless . I have to say it's essentially useless to firefighters . Sometimes the students ask me okay , so my AI predicts the temperature is 10 degree lower than the experimental measurement . So what's the problem ? If you tell the firefighter the smoke temperature is 240 degrees Celsius and the real temperature 250 , what's the difference ?
There's no difference to them . They only need to know whether it's safe to enter or it's not safe to enter . That's the information they need . They don't need numerical values . The value is meaningless to them .
And they also have their own diagnostics . There's a parallel world that's developing . At least in Poland , they would use thermal cameras a lot , right ? So perhaps they are less reliant on external sources of knowledge and more reliant on what they can hold in their hand than they use , right ?
Yes , so the IR camera is definitely a good example . Showing exact number is meaningless because we all know during the fire test if you're using some images to measure the smoke temperature , you will get the wrong data . But it doesn't matter . Well , the firefighters they only need to know which part is hot . See some images to measure the smoke temperature .
You will get the wrong data , but it doesn't matter For the firefighters . They only need to know which part is hot , which part is relatively cooler , and that information is already sufficient for them to judge where to go , whether they can evacuate safely .
And that kind of information is that kind of inaccurate information are very useful and actually are what the firefighters need . And sometimes we need to translate this information . No matter it's coming from the raw data or it's generated by the AI . We need to translate that into the language or the information .
The firefighters actually need it , because they are already overloaded in the operation . They are receiving massive information . They don't want you to give them a number how much megawatt . They don't need that . They need something simple that can help them to make decisions quickly .
That's the reason they like IR cameras , although the IR camera does not measure the correct temperature .
I think it's also a matter of trust and having power over something . If you hold an IR camera in your hands , you have the power to do the measurement yourself , interpret it yourself and take decisions yourself . When you're reliant on an external system or a person , you're taking information from the other person and that person may be wrong .
It's about consequences of being mistaken . I think people are much more accepting that they can be wrong and accept the consequences of that choice , rather than accepting the fact that someone else can be wrong and having to live with the consequences of that person being wrong .
In fact , a lot of our effort is now focusing on the digital twin technology . Okay , so the digital twin essentially is used to compress information . So , for example , if you want to have a digital twin of the building fire , you project that in a big screen that helps the fire commanders to make decisions . That is basically a demonstration .
You collect hundreds , thousands of data points every second but you transport that numerical information into colorful images . Simply some dangerous area , you put a red color , some safe area , you put a ring there and you compress this massive information into something vivid and something animated can be easily understood by the file commander .
So that's actually a lot of our effort is doing . So we try to use AI to compress information , compress that into something human can easily understand .
You also can gather data from different interfaces and compress it into one value so you can take observations from outside , from inside , from sensors and process . Where a person's standing is limited by the location where they are , they can only see what they see and not see what's in the back .
With my limited knowledge of firefighting process , I know the importance of assessing the full perimeter of a building . For example , they do this 360 check around the building to not miss the information coming from behind the building . You had UAV that would do the 360 while people are entering inside . They perhaps do the action quicker .
But again , we're brainstorming things that perhaps could be useful for them and we need to learn from firefighters what they want .
And after working on this project for almost five years , I feel like AI is not really good to work with humans or firefighters . Ai is perfectly working with robots . Okay , because the robot can just adopt the AI algorithm . They collect all the information and they process automatically with their powerful CPU GPU . But the human brain actually are quite limited .
We cannot handle so much information at the same time . We can only accept a few key information to do judgment , otherwise we overload it . So in that sense , I feel all the digital training , all the AI , all these things , they're best to work with robotic firefighter rather than the real firefighter .
That's definitely a trajectory for a future .
¶ Advancing AI in Fire Safety
I think it was like 10 years ago when I've read the piece that was covering development of humanoidal robots and the first use that they've predicted was something carrying loads , like , you know , a companion to carry whether we're talking about military setting and you know , carrying stuff with soldiers , but just as a carrier .
And the second thing that they foresaw was firefighting , and I found it very interesting and I remember that time a friend from Italy , Fabio Limo Ponciani . He was developing interesting firefighting robots .
I see that as potential future direction , though , knowing how devastating fire conditions are to my quite robust scientific apparatus that I put inside the fires and how fragile robots and drones are , I'm not sure if that's the place where I want to have my robots .
If our goal is to protect the firefighters , then letting the robots to do the job can solve everything .
So , sinian , wrapping it up , what are the next steps ? What should we talk about in three years ? What do you expect to happen in the use of ai , fire safety engineering and smart firefighting ?
I , I think it's . Currently we are trying to do a few things . First we are still exploring the ai driven fire safety design . So my students is further developing the database of the building .
So every time we have a very complex shape of the atrium or warehouse we are able to give prediction of ASAT very quickly and we hope that tools can be implemented into some software , some mature commercial software . They cannot generate the final result . They can give you a very quick early identification of what kind of case it runs .
If I recall correctly , the first instances of those AI tools for HVAC were available online . Can I link them in the show notes ?
Yes , please yes we have . It's just open . Everyone can play with that . It's only for a cubic box , it cannot cover a complex shape .
And the other thing is input using AI to do drawing , and our recent tests are showing that the design made by AI can have a lower number of speaker heads , so it can actually save the material you need , compared to some engineers' design , because we're trying to ask a few engineers you do this design , then we let AI to do the design , then we compare which
one is doing better . So hopefully we can achieve certain automation in this area , because I heard some engineers say , okay , they have to pinpoint the sprinkler head for this drawing , so it's kind of like a boring task . So right now we can use AI to just automatically generate that layout . So that's from design point of view . I think it can be .
A lot of things can be improved in three or five years . That's exciting , yeah .
And the other area I'm working on right now is essentially the robotic firefighting , because I feel this is essentially the final goal of the AI , because the this is essentially the final goal of the AI , because the robotic is essentially use AI to do all the measurement , to do the decision-making and essentially stop the fire .
If you use the AI to process information and pass that information to firefighters , then the efficiency is really low because we don't know if the firefighters understand what AI provides and there is also certain moral risk whether they should trust the AI information or not . But if AI is coupled with robots , the UAVs , then you don't have to worry about that .
They just pair so perfectly and they can do the job based on the information they get from the building or maybe from the wireless .
So I guess in the future we do robotic firefighting instead of smart firefighting .
It's still very challenging because you see , nowadays what AI replacing is not the conventionally hard labor work . They are not replacing firefighters . They are basically replacing researchers like us . They can generate papers , they can generate reviews , these kind of things humans used to think this is so unique , only humans can do . But right now AI can do that .
But instead , if you're carrying something , lifting something , these kind of things are very difficult to be replaced by robots , because our humans , we have evolved for millions of years . Just look at our hands . It's so elegant . There's no , basically no robots can achieve the function of the hand , but our brain only developed for a few thousand years .
So they are actually kind of replaced by AI now . But for the other part of the human human as a machine is actually very good it's very difficult to be replaced by robots . I think firefighters don't have to worry about being replaced by robots Anytime soon . It's not that easy . Not that easy , okay , thank you .
Thank you , scientists have to be worried about that .
Scientists , yes , and we are very well aware of the competition from AI-powered colleagues , who I'm not sure have the same level of morality as there should be for being a scientist . Sinan , thank you so much .
It was a pleasure talking with you again in the Fire Science Show and , as always , you have the most exciting topics , so I hope the audience enjoyed this as well . Thank you very much for the invitation and that's it . It's quite odd that something can be at the same time exciting and disappointing , right ?
I mean the interface between fire science and firefighting . The fire scientists people like me who learn about compartment fires , discover physics and so on , and the firefighting fire scientists who do the stuff that's important for a fire brigade .
We don't have a common language and even if you have the best intentions , it is real challenging to find the connections , and I know that there are people who want those connections . I know that there are people who want those connections . I know that there are people looking for that .
There are people working really hard , like I am , about building the bridges and closing the gap between us , because it makes no sense that there is a gap right . But it truly is a challenge and it needs good communication . I've said it so many times in the podcast Good communication is absolutely necessary to close the gap .
Institutions like FSRI are necessary to close the gap . Institutions like FSRI are necessary to close this gap . Open-minded firefighters who listen to Fire Science Show are needed to close this gap . Open-minded fire scientists who first ask firefighters and then start their research are needed to close this gap .
And I hope together we will be able to close this gap because it's really needed in the world . And from the exciting things wow , ai-powered autonomous firefighting robots where do I sign in ? That's a crazy interesting idea , like a science fiction level idea . I love it .
And now we can also deploy them in space or in Mars habitats to make it even more exciting . The next thing use of AI to support fire safety engineering . Like truly , I know the guys from HK Poly . I had my students stay with them for many months . We work together with HK Poly and the way how those guys use AI like really is crazy .
Like it's really like applicating CFD in fire safety engineering looked like 15 , 20 years ago . There's basically no difference . You take a package , you implement that . It gives you results . It's next level . And looking at those guys , I know how the future of fire safety engineering will look like , and I hope this discussion gave you an insight to that as well .
So that will be it for the AI experiences with AI in fire safety engineering episode . I hope you've enjoyed those experiences . If we find some novelty in the research of Sinian and I'm sure we will I'll bring him back again to the podcast .
Sinian is the top one scientist of my generation and I admire him so much so I would love to have him more in the show and I hope you would like that For the next week . Another interesting topic in the fire science show . So make sure you tune in on Wednesdays .
And I have to tell you that my season two of Uncovered Witness another four episodes in the other podcast , uncovered Witness , fire Science Revelations is also live and we're covering human behavior in fires in that series . I think it's also very exciting a little different format than the fire science show , narrated , scripted , more in-depth .
Look into one specific part of fire science , this case human behavior in fires . I am sure that if you like fire science show and if you like the Fundamentals of Fire Science series that we present in here Uncovered Witnesses , fire Fundamentals on Steroids and you'll love it . Thank you for being here with me . See you here next Wednesday .
Bye , this was the Fire Science Show . Thank you for listening and see you soon .