Oct. 23, 2024

174 - My predictions for AI in fire

174 - My predictions for AI in fire

AI is changing the world. But can artificial intelligence truly revolutionize fire safety engineering? In this episode I took out my crystal ball, and tried to find answers in what aspects of fire engineering we could truly see a revolutionary impact of AI, and where it is more a disturbing gadget with no real application... 

Overall, working in the space for many years, and having talked with many luminaries of the use of AI, I would say my expectations are toned down a lot. I am still excited, but I've also learnt that really good AI applications require exceptional knowledge and resources, often at a level not accessible in fire science. While the building blocks are there, we may be short in the data, processing power or knowledge to truly apply it. At the same time, as AI became a buzzword for novelty, a lot of people pursue very simple AI applications with extraordinary claims - something that kind of triggers and annoys me...

In this episode I try to give my opinion about the use of AI in:

  • supporting fire modelling
  • writing and reporting (engineering and in science)
  • data processing
  • predicting fire phenomena
  • studying fire phenomena
  • material properties
  • fire testing and legislation

Each of those I've ranked on an arbitrary scale of 
1 - b-s. 
2 - not too excited
3 - a good use of AI
4 - very exciting use of AI
5 - paradigm shifting use of AI

Which got the highest rank? Listen to the episode to find by yourself :)

In the episode I refer a lot to the previous episodes of the Fire Science Show in which AI technology was discussed. You can listen to those in the AI/ML section of the podcast, find them all here: https://www.firescienceshow.com/category/ai-ml/

Big thanks to the inspirational guests and my close friends that helped me clear my mind on AI, especially to MZ Naser, Jakub Bielawski, Mike Spearpoint, Danny Hopkin, Matt Bonner, Xinyan Huang, Michael Kinsey and Mike Stromgren! 

----
The Fire Science Show is produced by the Fire Science Media in collaboration with OFR Consultants. Thank you to the podcast sponsor for their continuous support towards our mission.

Chapters

00:00 - Future of AI in Fire Safety

15:48 - AI Applications in Fire Safety Engineering

26:17 - Limitations and Potential of Fire Prediction

34:45 - AI Applications in Fire Testing

48:19 - Future of AI in Design Engineering

Transcript
WEBVTT

00:00:00.681 --> 00:00:02.266
Hello everybody, welcome to the Fire Science Show.

00:00:02.266 --> 00:00:20.993
Today's episode is going to be a little easier than usual and instead of presenting you fire science facts or an interview with an eminent guest, I'm going to sit here and present some of my opinions, opinions about the use of AI, artificial intelligence and machine learning technologies in the world of fire science.

00:00:20.993 --> 00:00:26.492
And what do I think about their future use in our discipline.

00:00:26.492 --> 00:00:27.841
How prevalent will they be?

00:00:27.841 --> 00:00:31.371
Will they make big change in some aspects of fire safety engineering?

00:00:31.371 --> 00:00:41.459
That's what I've been thinking about a lot recently and I felt it's about time that I share those opinions with you and I hope that it's going to trigger some discussions.

00:00:41.459 --> 00:00:43.103
I wonder what are your opinions?

00:00:43.103 --> 00:00:46.612
So reason to do an opinion analysis episode.

00:00:46.612 --> 00:00:48.828
It's an inspiration from other podcasts that I listen.

00:00:48.828 --> 00:00:53.427
I've heard that being done in some other projects and I've kind of enjoyed those episodes.

00:00:53.427 --> 00:00:59.831
So I wondered if the Fire Science Show audience will also enjoy an episode of that kind in this podcast.

00:00:59.831 --> 00:01:00.924
I hope you will.

00:01:00.924 --> 00:01:08.908
And the topic, the theme of the episode the artificial intelligence is kind of inspired by the recent Nobel Prizes.

00:01:08.908 --> 00:01:18.403
As you may know, the Nobel Prize in physics was given for the development of neural networks, the fundamental thing of machine learning and AI technology itself.

00:01:18.403 --> 00:01:29.052
This award was given in physics to John Hopfield and Geoffrey Hinton, and those gentlemen have enabled us this massive, massive revolution that's changing the world as we see it today.

00:01:29.052 --> 00:01:41.451
Also, there was another Nobel Prize in chemistry given for the understanding the structure of proteins, and that was done through computational modeling and also through a massive, massive use of machine learning.

00:01:41.451 --> 00:01:46.019
That helped us understand the most complicated structures of proteins.

00:01:46.019 --> 00:01:52.370
That fueled an insane development in biology and in medicine.

00:01:52.370 --> 00:02:06.210
Think about it we are at the level where we are using machine learning, ai and computational methods to create new medicines, to create new pathways to solve medical problems in our organism.

00:02:06.210 --> 00:02:12.492
That's how confident we can be with those solutions in some aspects of the modern world and science.

00:02:12.492 --> 00:02:15.789
And the question is are we at the same level in fire science?

00:02:15.789 --> 00:02:17.606
Are we at the same level at civil engineering?

00:02:17.606 --> 00:02:40.733
Another reason for doing this episode I was recently invited by Warsaw Technical University that's the biggest technical university in Poland to give an inaugural lecture for civil engineering department and in that lecture I was sharing with the first year students my predictions about the use of AI and how AI will change their discipline as they finish their school in some years.

00:02:40.733 --> 00:02:44.003
So I've been predicting a lot recently.

00:02:44.003 --> 00:02:48.146
I'm not sure if I feel super comfortable with that, but I guess it's a fun thing to do.

00:02:48.146 --> 00:02:49.612
It's a fun exercise.

00:02:49.612 --> 00:02:56.300
It pushes people to think a little bit, it pushes people outside of the box and I think it's just an interesting exercise.

00:02:56.881 --> 00:03:02.473
In this episode we're going to narrow things down to some areas that I've came up with.

00:03:02.473 --> 00:03:08.163
That's going to be use of AI in supporting fire engineers in their modeling tasks in fire modeling.

00:03:08.163 --> 00:03:12.112
Use of AI in supporting fire experts in their everyday job.

00:03:12.112 --> 00:03:14.844
So how can AI support a fire expert?

00:03:14.844 --> 00:03:22.108
And that will transition into use of all kinds of large language models, such as chatbots in fire engineering.

00:03:22.108 --> 00:03:27.800
We're going to talk about use of AI in working with data that we have in fire science.

00:03:27.800 --> 00:03:33.573
We're going to talk about the use of AI to predict fire phenomena or explain fire phenomena.

00:03:33.573 --> 00:03:51.731
We're going to touch a bit of fire testing and then we'll perhaps move into some more crazy uses of AI, as incorporating AI in our codes and frameworks and how could the world look like when or where we would rely on it very heavily in fire engineering.

00:03:51.731 --> 00:03:52.693
A lot to go through.

00:03:52.693 --> 00:04:04.233
I'm going to give my opinion on a scale of one to five for each of them and provide you an explanation, one being something I find pretty useless and I don't see much potential in it.

00:04:04.233 --> 00:04:10.793
Five, something that I believe could be a groundbreaking, paradigm-shifting change in fire safety engineering.

00:04:10.793 --> 00:04:16.552
I wonder if you agree or disagree with me, but anyway, let's get going with that.

00:04:21.440 --> 00:04:22.987
Welcome to the Firesize Show.

00:04:22.987 --> 00:04:26.511
My name is Wojciech Wigrzyński and I will be your host.

00:04:26.511 --> 00:04:46.011
This podcast is brought to you in collaboration with OFR Consultants.

00:04:46.011 --> 00:04:48.949
Ofr is the UK's leading fire risk consultancy.

00:04:48.949 --> 00:04:59.793
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.

00:04:59.793 --> 00:05:15.591
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.

00:05:15.591 --> 00:05:27.264
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.

00:05:27.264 --> 00:05:38.312
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.

00:05:38.312 --> 00:05:41.223
Get in touch at ofrconsultantscom.

00:05:42.086 --> 00:05:50.307
Okay, so about the time to clean my crystal ball and look into depths of that to see what the future of AI and fire safety engineering looks like.

00:05:50.307 --> 00:06:01.749
I'm kidding, it's just some of my opinions and predictions, though I must say the theme of AI and machine learning has appeared very early in the fire science shows in the first episodes.

00:06:01.749 --> 00:06:08.730
I think it was episode four where I had Matt Bonner and we already talked about use of AI in the field of fires and facades.

00:06:08.730 --> 00:06:23.569
Then I had episode with Sina Huang, also within the first 10 episodes, where we've talked about AI as a supportive tool for firefighters, and we've recently revisited that episode with Sina as well to see three years later what were the real implementations from their project.

00:06:23.569 --> 00:06:32.043
See three years later what were the real implementations from their project Very early, somewhere around episode 25,.

00:06:32.043 --> 00:06:36.579
I've had MZino, Sir, and we've talked about how to learn artificial intelligence and how to apply that, and we've also revisited that episode some episodes ago.

00:06:37.360 --> 00:06:44.906
Later on, I've been here with you when the chatbot revolution happened and we had an episode with Mike Kinsey about how to use chatbots in fire engineering and how to perhaps make a fire.

00:06:44.906 --> 00:06:49.634
Chatbot revolution happened and we had an episode with Mike Kinsey about how to use chatbots in fire engineering and how to perhaps make a fire chatbot.

00:06:49.634 --> 00:06:58.201
And not long later I had an episode with Mike Strongren where we were brainstorming some of the future technologies in fire safety engineering.

00:06:58.201 --> 00:07:09.209
So the theme of AI, machine learning and new technological developments has been in the Fire Science Show present for a long time, since the very beginning, and it's not going to change.

00:07:09.209 --> 00:07:22.742
And you know, having all those conversations with those people who are much better than me in the world of AI, who are the real experts, it gives me some confidence in being able to give my opinions on how the future may look like.

00:07:22.742 --> 00:07:24.608
So I've promised you some topics.

00:07:27.620 --> 00:07:37.290
Let's perhaps go outside of the list I've presented and just start with a very general question Can we predict an AI, a general AI model that could just do the entire fire safety engineering job?

00:07:37.290 --> 00:07:40.548
And I would say in the next five to 10 years?

00:07:40.548 --> 00:07:41.812
No, not really.

00:07:41.812 --> 00:07:44.048
We'll always be necessary.

00:07:44.048 --> 00:07:47.103
The fire safety engineers will always be necessary at the job.

00:07:47.103 --> 00:07:59.000
It's just too complicated, too much interaction with other stakeholders, too much risk in that, too much uncertainties involved with the use of AI as a general tool.

00:07:59.480 --> 00:08:01.944
So if you ask me a question, will AI replace fire safety engineers?

00:08:01.944 --> 00:08:02.245
I would say no.

00:08:02.245 --> 00:08:03.685
If you ask me a question, will AI replace fire safety engineers?

00:08:03.685 --> 00:08:17.321
I would say no, at least not in any foreseeable future, because in 20, 30 years, who knows how big the breakthroughs will be In machine learning?

00:08:17.321 --> 00:08:18.764
Perhaps we'll all be replaced by some sort of artificial intelligence.

00:08:18.764 --> 00:08:28.644
But I would say at the current level of the complexity of the fire engineers' work, it's highly unlikely that the entire field could be replaced by some sort of general AI.

00:08:28.644 --> 00:08:33.644
I think there are other fields that would be replaced first, such as podcasting, for example.

00:08:33.644 --> 00:08:42.087
I'm now being challenged by my friend, grunder Jumas, who's creating AI-based podcasts for his newsletter Burning Matters.

00:08:42.087 --> 00:08:45.514
Actually, you should subscribe to that one if you have not.

00:08:45.514 --> 00:08:48.529
It's a great read on Monday morning.

00:08:48.529 --> 00:08:58.145
But Grunde has recently been using technology called Notebook LM that pretty much takes the text that you supply it with and it creates a podcast episode.

00:08:58.145 --> 00:08:58.948
This is crazy.

00:08:58.948 --> 00:09:03.948
I'm gonna get out of job soon if this technology improves at this pace.

00:09:03.948 --> 00:09:13.282
So yeah, in terms of replacing fire safety engineers, I think we'll replace fire podcasters earlier and if that happens, you'll know.

00:09:13.282 --> 00:09:14.024
You will know.

00:09:14.585 --> 00:09:17.809
Now, going into specific topics that I've mentioned.

00:09:17.809 --> 00:09:29.729
So first, use of ai as something that supports fire safety engineer in their CFD calculations, and I would split that into three groups.

00:09:29.729 --> 00:09:41.971
One is use of AI in the preparations of CFD use of AI in helping with the calculations, and use of the AI in post-processing CFD results.

00:09:41.971 --> 00:09:50.950
In terms of preparing your CFD projects, I think in next five to ten years there's not going to be a big breakthrough in that.

00:09:50.950 --> 00:09:52.360
It's not going to be that helpful.

00:09:52.360 --> 00:09:54.044
I would rank it two out of five.

00:09:54.787 --> 00:10:05.768
I'm not sure if we'll see that much use in that, perhaps in specific applications like splitting your models into meshes or just helping you do some mundane repetitive tasks.

00:10:05.768 --> 00:10:07.304
Then, yes, you could use that.

00:10:07.304 --> 00:10:13.625
You can use chatbot to write FDS code but to build a complex model of a building for your analysis.

00:10:13.625 --> 00:10:26.434
That's just so much responsibility, so much work and you need so much interaction with other stakeholders in the process that a simple machine learning algorithm will not be able to do it at the level we would need.

00:10:26.434 --> 00:10:40.417
Perhaps later in the future, if we got really really good BIM integration, if we got really powerful processing technologies, standardized design, that would just streamline the process.

00:10:40.417 --> 00:10:50.797
Perhaps then we could use some AI-based technologies to prep our models quicker, but I think it's pretty far from now For supporting in calculations.

00:10:50.797 --> 00:10:54.886
I'm actually quite positive in that, so I would give it 4 out of 5.

00:10:54.886 --> 00:10:58.624
The technologies already exist and they're getting better and better.

00:10:59.225 --> 00:11:34.386
What I mean about speeding CFD is that you could make a hybrid in which you use your computational fluid dynamics to solve the region of interest in your model like the fire plume, like the jet fan locations, like the inlets, outlets of your model, so the places where most interesting physical phenomena take place that define the outcome of your simulation and then you use machine learning algorithms to solve the other parts of your model, the ones in which the interesting physics is not happening, and you can speed up the simulations tremendously.

00:11:34.386 --> 00:11:44.090
I've seen applications like that and it's just brilliant how much quicker you can solve your CFD while not losing the accuracy of your solution.

00:11:44.090 --> 00:11:48.746
But that's only for the parts of the model where the interesting physics is not happening.

00:11:48.746 --> 00:11:56.410
If you're asking me about just replacing CFD with AI as a black box tool, I know there are applications made like that.

00:11:56.410 --> 00:11:59.400
I would give it somewhere like one, two out of five.

00:11:59.400 --> 00:12:01.423
It's just too complicated.

00:12:01.423 --> 00:12:14.291
It's very unlikely that we could actually teach machine learning models on big enough data sets to give us a generalized solver for smoke and fire in buildings.

00:12:14.291 --> 00:12:22.456
It's just very unlikely, maybe for some repetitive geometries, but the more repetitive geometry is, the less reliant you are on the CFD prediction anyway.

00:12:22.456 --> 00:12:25.437
So I'm not sure that's going to be that useful.

00:12:25.437 --> 00:12:42.970
I would not trust such a tool because simply the complexities coming out from architecture or details of your building were going to be the drivers for the smoke flow and the outcomes of your simulation, and I don't think we could teach AI to a level where it would capture all those intricacies.

00:12:43.460 --> 00:12:47.427
Now the last part, post-processing, and I'm quite optimistic in this.

00:12:47.427 --> 00:12:49.206
So I would give it three out of five.

00:12:49.206 --> 00:12:56.006
I think it's going to be quite useful in our everyday job, but probably not a breakthrough.

00:12:56.006 --> 00:12:59.009
It's just going to help us streamline some processes.

00:12:59.009 --> 00:13:16.028
So I can see tools being created that will help you post-process your simulation results, that will help you go through those simulation results, turn them into a report that you can just prepare it quicker and supply it to your client in a more clear way.

00:13:16.028 --> 00:13:24.610
Perhaps you could also use chatbots to check the clarity of your reporting, to check the clarity of your communications.

00:13:24.610 --> 00:13:34.769
I think it's just going to be useful, but not something I would consider a massive breakthrough, just something that makes our lives a little bit easier as fire safety engineers.

00:13:34.769 --> 00:13:55.947
It's also something that I'm talking about recently in conferences about the post-processing of simulation results and how, with the growing power of computers and this AI support at CFD solving how we're getting access to faster and faster CFD and this support will be necessary to process the amounts of data that we will be creating in future CFD analysis.

00:13:55.947 --> 00:14:02.866
So overall, ai for CFD modeling and just modeling in FIERS I would give it overall three out of 5.

00:14:02.866 --> 00:14:05.393
I think it's going to be a good companion.

00:14:05.393 --> 00:14:16.450
Nothing that changes the game, but also quite positive that it's going to have a good, positive outcome on the work of Firesafe Engineer in this space.

00:14:17.539 --> 00:14:31.927
Next on my list AI for the design work, and here, similar as in preparation of CFD models, the incorporation of AI directly into engineering framework is just difficult.

00:14:31.927 --> 00:14:41.152
It's too much interaction with other people, it's too much soft knowledge that a generalized AI model could handle.

00:14:41.152 --> 00:14:52.143
If your work is very specific, you work on a specific type of projects, specific type of code you could probably get something out of it.

00:14:52.143 --> 00:14:58.884
You could perhaps teach a neural network to work in the specific projects that you're working with, but the more general you get, the less useful AI will be.

00:14:58.884 --> 00:15:11.312
I saw some applications like optimizing the design of sprinklers, for example, or perhaps calculating the evacuation distances and optimizing some footprints of the buildings for evacuation.

00:15:11.312 --> 00:15:18.620
There's a lot of happening in generative design, where you can build entire layouts with AI in split second.

00:15:18.620 --> 00:15:33.673
I think those technologies are very interesting and possibly will be used in the future, but I think in the next five to 10 years it's not going to be something that's going to be a big breakthrough in our discipline.

00:15:33.673 --> 00:15:42.806
I mean, it's an interesting gadget and perhaps could find some use in some specific aspects of the design, but I don't think it's going to have a big impact overall.

00:15:42.806 --> 00:15:48.315
Our job is just simply too complicated for the artificial intelligence at its level.

00:15:48.879 --> 00:15:50.005
Another thing is chatbots.

00:15:50.005 --> 00:15:54.150
So here I'm also not that excited for the future.

00:15:54.150 --> 00:15:58.220
The more I work with the chatbot, the less excited I am actually.

00:15:58.220 --> 00:16:21.145
I mean it's quite impressive in what it can do in terms of assisting me in programming and assisting in some tasks, but at the same time, many times the outcomes are just purely stupid and it's very difficult to actually tell this stupid answer from the smart answer when you don't have enough knowledge to distinguish between two of them Because they are as confident they look the same.

00:16:21.145 --> 00:16:29.746
You have to have good knowledge about what you're asking about to be able to tell if the answer is correct or not.

00:16:29.746 --> 00:16:42.567
So in a way, it's helping me more with automating the stuff that I would be doing, rather than replacing me at the stuff that I'm doing, and I view this technology only through this lens.

00:16:42.567 --> 00:16:51.291
So if I had to rank it on a scale of one to five, I would give it a solid two, maybe three in some specific spaces.

00:16:51.291 --> 00:16:53.928
I don't think it's going to be that much a breakthrough.

00:16:54.399 --> 00:17:02.530
If someone made their own chatbot like we discussed that in episode with Mike Kinsey a very specific code that has a very specific knowledge.

00:17:02.530 --> 00:17:04.451
It could be a good support tool, but still just a support knowledge.

00:17:04.451 --> 00:17:13.523
It could be a good support tool, but still just a support, not something that could replace or do job for a fire engineer, and I think this is a big trap.

00:17:13.523 --> 00:17:25.969
I think a lot of people will try to use chatbots to do engineering design for them because it looks good, it sounds real, and those people will get into a lot of trouble, lot of trouble.

00:17:25.969 --> 00:17:26.490
Don't do that.

00:17:26.490 --> 00:17:29.547
Don't use chatbot in your engineering design.

00:17:29.547 --> 00:17:40.262
Don't use large language models unless you build them yourselves and you're very, very familiar with the input that was used to teach them and you can evaluate the output of them.

00:17:40.262 --> 00:17:43.469
Only in that I would rely on such a technology.

00:17:44.369 --> 00:17:56.606
One more aspect, a side topic of that I think it's quite interesting to use AI in teaching and in onboarding professionals in your companies, and you know CPD quizzes.

00:17:56.606 --> 00:18:11.413
I think there are many supporting roles that we could use AI to just be sharper file safety engineers, you know, test our abilities to some extent, cross-check our work, perhaps use the chatbots to make it clearer and just communicate better.

00:18:11.413 --> 00:18:20.705
In that case, yes, this is a good technology, but still just a supporting tool at most, not something that will change our work tremendously.

00:18:20.705 --> 00:18:41.545
So, overall use of large language models and general AI in helping fire engineering in the fire engineers everyday work I would give it two out of five, and since this is a fire science show project, I'm also going to give a prediction for fellow fire scientists, and here I'm less excited than for fire engineering.

00:18:42.066 --> 00:18:51.799
I really think we should not be using AI in our research work in terms of analyzing the data, preparing the research outcomes.

00:18:51.799 --> 00:19:10.022
I still think you can use AI for some of your processing work to clean out the data, which we'll be talking in a few seconds but to rely on AI on on, for example, writing a scientific paper, this is just horrible.

00:19:10.022 --> 00:19:11.306
This is a misconduct.

00:19:11.306 --> 00:19:12.288
You should not do that.

00:19:12.288 --> 00:19:14.843
I see people do that and you can immediately tell.

00:19:14.843 --> 00:19:22.284
I mean, the moment we will be not able to tell a real researcher work from a chat, gpt generated work.

00:19:22.284 --> 00:19:30.359
That's that not going to be funny and I understand that technology is improving, but you can tell it's creating a lot of bullshit.

00:19:30.359 --> 00:19:35.893
It's just not able to process and give reasonable outcomes.

00:19:35.893 --> 00:19:44.018
It's basically it's learned on what's already been known and science is about finding stuff that we do not know.

00:19:44.018 --> 00:19:50.854
I think it's a horrible application to just rely on AI in creating research.

00:19:50.854 --> 00:19:52.626
I think it's a serious misconduct.

00:19:52.626 --> 00:20:01.153
Yes, you can use AI to check your paper for clarity, to correct the English grammar, to rephrase things.

00:20:01.153 --> 00:20:02.987
I think that's a very nice use of AI.

00:20:02.987 --> 00:20:10.299
You can use AI to help understand if there's more to the topic that you've covered.

00:20:11.182 --> 00:20:16.553
There are some AI-based tools that help you conduct literature review.

00:20:16.553 --> 00:20:18.787
That's actually quite interesting.

00:20:18.787 --> 00:20:29.576
We are using a tool called Research Rabbit and it is very interesting and very helpful in going through literature, but those are just very basic supportive tools.

00:20:29.576 --> 00:20:33.746
The science must be done by the scientist, not the AI itself.

00:20:33.746 --> 00:20:40.063
So the use of AI in research, I can give it the most two out of five.

00:20:40.063 --> 00:20:45.862
And if you think about using AI to write the research, I would give it one out of five.

00:20:45.862 --> 00:20:47.286
It's a horrible, horrible use.

00:20:47.286 --> 00:20:51.960
However, that's for, let's say, five-year horizon.

00:20:51.960 --> 00:21:04.955
If we think long-term, it actually could be that when the networks get better, we perhaps could use AI in science to change the way how science is done or science is communicated.

00:21:05.519 --> 00:21:10.853
We're in a space where there are way too many research papers published worldwide.

00:21:10.853 --> 00:21:22.173
It's hard to keep track with what's being published, with what's being done Definitely too much of that and perhaps AI could help us replace some parts of those papers.

00:21:22.173 --> 00:21:28.560
If you think about the paper, the method section is just a description of the method that has been used.

00:21:28.560 --> 00:21:30.130
You could replace that section with just telling the of the method that has been used.

00:21:30.130 --> 00:21:43.692
You could replace that section with just telling the novelty, what's new being used in your model, and then you could imagine having a link in the paper that would just get you into a chatbox sort of technology that would explain what has been done in the paper.

00:21:43.692 --> 00:21:52.872
I had this interesting discussion with some friends Mike Spearpoint, lucas Arnold and Arnaud Trouvet about this in the ESFS conference and it was quite interesting.

00:21:52.872 --> 00:22:11.130
Like, if you think about papers and what goes into a paper, sometimes we put too much stuff into there, and if information can be processed at blink of an eye, perhaps instead of the researchers spending their time writing some repetitive parts of the papers, they can just focus more on the science.

00:22:11.130 --> 00:22:23.011
Perhaps AI could be used to make those papers more concise, more clear, more focused on the problem they're solving and thus becoming better communication tools.

00:22:23.011 --> 00:22:30.662
If that's a new use of AI, I would support that, but I'm not sure if the current technology is at the level that I would support that, but I'm not sure if the current technology is at the level that would allow for that.

00:22:30.662 --> 00:22:33.387
So I'm mildly optimistic for a future.

00:22:33.387 --> 00:22:35.952
So far, maximum two out of five.

00:22:35.952 --> 00:22:39.386
Perhaps in future it will improve.

00:22:39.386 --> 00:22:44.201
Anyway, if you're using AI to write papers for you, that's like zero out of five.

00:22:44.663 --> 00:22:48.528
The next part is using AI in data simulation.

00:22:48.528 --> 00:22:52.115
So some people say that data is the new gold.

00:22:52.115 --> 00:22:54.207
I disagree with that.

00:22:54.207 --> 00:22:58.799
I mean, if you look at the gold, it has some technical functionality.

00:22:58.799 --> 00:23:01.209
It's a great conductor for electricity and heat.

00:23:01.209 --> 00:23:12.962
You can process it into extremely thin foils of the material and that has various uses extremely thin foils of the material and that has various uses.

00:23:12.962 --> 00:23:20.002
But you know, the value of gold comes from the market speculation, not from its inherent technical value, and I don't think data is something speculative.

00:23:20.002 --> 00:23:21.626
I think we have real value in data.

00:23:21.626 --> 00:23:31.267
For me, the data is the new coal that could fuel the new industrial revolution in the world, and it is already doing that, and with that I completely agree.

00:23:31.768 --> 00:23:47.950
Having data in the modern world to give you better predictions, better outputs in engineering as well, is fundamental, and following on that, we can use AI to help us work with the data.

00:23:47.950 --> 00:24:05.786
First of all, that's an interesting dynamic there, because if you want to do some really good machine learning or you want to use the most out of the data you have, the clearer the structure of the data you have, the better organized your data is, the more points you have, the better outcomes you will have.

00:24:05.786 --> 00:24:09.385
But then again, to collect the data from multiple stakeholders.

00:24:09.385 --> 00:24:37.853
For example, if you want to do analysis across countries from different statistical regimes that were analyzed like, think about the EU Firestat project that we've also discussed in the podcast If you want to gather data sets that big coming from this many spaces, it's actually quite challenging to work with that data and here you can use the AI to work through the data, organize, structure it and actually help you work with it better.

00:24:37.853 --> 00:24:56.011
I think that's a really good application of AI because you have quite good control over the process and you can validate the outcomes from the process with some well-established methods that we know for machine learning, and it's something that can be very powerful.

00:24:56.440 --> 00:24:59.351
I can refer you to an interesting paper some music.

00:24:59.351 --> 00:25:03.767
It was published in Fire Technology Journal by a Polish scientist, marcin Mironczuk.

00:25:03.767 --> 00:25:07.859
He did analyze the reporting from firefighters.

00:25:07.859 --> 00:25:22.711
So in Poland after every fire incident the firefighters would issue a report and Marcin did process those reports with language model many years ago, long before chatbots, and he was able to extract a lot of interesting data from those reports.

00:25:22.711 --> 00:25:36.093
So he was taking a random pieces of text that firefighters filled on site and turned those into databases where you could browse through how many fires happened, what size of the fire, what did ignite, and so on and so forth.

00:25:36.093 --> 00:25:47.605
So using AI to create really good and valuable database from a very chaotic dataset, that's a really really powerful use of AI.

00:25:47.605 --> 00:25:58.664
And if we can find users like that in fire safety engineering and fire science and I'm very sure we will, because data is something we always struggle with we always need more data.

00:25:58.664 --> 00:26:01.209
We always need data to feed our decisions.

00:26:01.209 --> 00:26:04.509
So I would give it three, maybe four, out of five.

00:26:04.509 --> 00:26:14.534
I think this is a really really good application and I hope more people start working on incorporating AI into working with data in fire engineering.

00:26:14.534 --> 00:26:16.969
I think there's a great future in this application.

00:26:17.880 --> 00:26:25.770
The next on my list perhaps this one is going to be a little controversial, but let's talk about fire phenomena and predicting fire phenomena.

00:26:25.770 --> 00:26:40.872
So, looking at the models that we currently have having myself play a little bit with some algorithms to predict phenomena that we see in our laboratory, to work with the data we have in our laboratory I would give it one out of five.

00:26:40.872 --> 00:26:59.584
On one hand, yes, those are powerful tools that, if you teach them correctly, they can quite well capture the fire phenomena or the outcomes of fire phenomena that are happening within this very specific range of variables that you've used to teach your machine learning algorithm.

00:26:59.584 --> 00:27:03.894
But these algorithms cannot predict outside of that.

00:27:03.894 --> 00:27:38.334
They are absolutely unable to predict anything beyond the space at which you have taught them and this is extremely challenging of proteins, with machine learning having access to, I think, 100,000 of three-dimensional structures of proteins and then being able to extrapolate that into 100 billion types of protein.

00:27:38.334 --> 00:27:40.827
I mean, they've certainly achieved that.

00:27:40.827 --> 00:27:51.568
They certainly did achieve the predictive capability of AI in a field that perhaps is 100 times more complicated than fire science.

00:27:51.568 --> 00:28:06.634
So I'm not saying it's physically impossible that you capture extremely complicated physical phenomena with AI, but I think it's out of the reach of the fire community at this point in time.

00:28:07.180 --> 00:28:09.484
The people who did that they work for Google.

00:28:09.484 --> 00:28:15.182
They have one of the best machine learning teams, if not the best machine learning team in the world.

00:28:15.182 --> 00:28:26.750
They have unlimited access to resources and people and they were working on a problem that's multi-billion dollar in impact.

00:28:26.750 --> 00:28:33.268
If you can predict the structure of protein in impact, like if you can predict the structure of protein, you can save billions in your resurgent development.

00:28:33.268 --> 00:28:56.980
So the scale at which they have been working is something completely different than the scale that most fire scientists are at and I think fire phenomena I think it was Hoyt Hotel who said that the fire and life are two most complicated problems to solve in the world and we certainly do not appreciate how complicated, how multivariable the fire phenomena are.

00:28:57.462 --> 00:29:08.752
We're talking about hundreds of chemical reactions happening at the same time, creating different energy and product outcomes in every millisecond of the combustion.

00:29:08.752 --> 00:29:21.126
We're talking about turbulent mixing at all of the scales, from the level of turbulent diffusion into the flame up to entrainment phenomena and flow around structures.

00:29:21.126 --> 00:29:34.084
We're talking about flame spread across complicated three-dimensional materials with their thermomechanical properties that change with temperature and time.

00:29:34.084 --> 00:29:41.585
We're talking about humans involved in the process and all the interaction between the human and the fire.

00:29:41.585 --> 00:29:46.094
It's really the world of fire is so, so, so complicated.

00:29:47.181 --> 00:30:03.311
We have some success in simplifying it enough, but to capture the real thing as you have in the laboratory, you can put the world's best fire scientists to an experiment in a fire lab and I can guarantee that some of them will still be surprised by the outcomes.

00:30:03.311 --> 00:30:04.605
We get that every day.

00:30:04.605 --> 00:30:06.646
That's kind of the beauty of our job.

00:30:06.646 --> 00:30:19.329
And now you take those limited outcomes that you have, limited data points that we have from fire experiments and you're trying to teach an algorithm, you try to give predictive capabilities to that algorithm.

00:30:19.329 --> 00:30:21.080
It's simply not going to happen.

00:30:21.080 --> 00:30:28.261
My trust to such applications is very minimal and in the literature we see a lot of people try.

00:30:28.804 --> 00:30:33.528
Okay, there are some spaces in which people actually try and quite succeed with that.

00:30:33.528 --> 00:30:36.117
For example, prediction of spalling.

00:30:36.117 --> 00:30:45.722
That's an interesting one where the ai could get very good results compared to our ability to model spalling, which we do not have today.

00:30:45.722 --> 00:30:52.388
Predicting flashoashover, predicting backdrafts research like that has been done and shown good success.

00:30:52.388 --> 00:30:59.125
Predicting the size of fire based on the images of the fires very interesting technology.

00:30:59.125 --> 00:31:01.648
We've just talked about that with Xin Huang in the podcast.

00:31:01.648 --> 00:31:13.525
So there are some applications in which it works, but it's very narrow and they are only able to feed you within the space of variables that they've been taught about.

00:31:14.248 --> 00:31:23.073
And the most exciting thing about science is where you discover something that does not fit to the model, where you are outside of the current scope of knowledge.

00:31:23.073 --> 00:31:25.224
These are the exciting parts of science.

00:31:25.224 --> 00:31:38.989
These are the parts that make you want to pursue a career as a scientist, and AI, by definition, cannot work in those areas unless you're Google, unless you have really unlimited resources, huge team and ability to go there.

00:31:38.989 --> 00:31:44.247
Perhaps we need a Google Fire of some sort that would help us solve the fire phenomena.

00:31:44.247 --> 00:31:57.913
But until that happens, my prediction but until that happens, my prediction for the future of AI as a predictor of fire phenomena is very limited.

00:31:57.913 --> 00:31:59.240
I don't think this technology will be very good at that.

00:31:59.240 --> 00:32:05.512
It can shine in some very narrow, very specific applications, but again, not at the groundbreaking level.

00:32:05.512 --> 00:32:09.428
It's not going to give us the discoveries on its own.

00:32:09.428 --> 00:32:18.042
So my prediction here one, maybe two at the most out of five.

00:32:18.083 --> 00:32:22.577
Though there is another way and in this way I'm actually quite excited about it and that's the causal AI, that's the explainable AI.

00:32:22.577 --> 00:32:26.848
I had Amzina Sir in the podcast talk about that a bit episodes ago.

00:32:26.848 --> 00:32:28.172
I'll link it in the show notes.

00:32:28.172 --> 00:32:31.821
Talked about that a bit episodes ago.

00:32:31.821 --> 00:32:33.265
I'll link it in the show notes.

00:32:33.265 --> 00:32:38.273
So this technology, what we're doing with it now is that you're using AI to predict some phenomena like spalling.

00:32:38.273 --> 00:32:40.183
Actually, that's a good example.

00:32:40.183 --> 00:32:49.297
I had MC Nasser deliver a workshop for my institute and he's shown us exactly how they're using explainable AI to understand spoiling.

00:32:49.297 --> 00:32:50.861
I must say that it was brilliant.

00:32:50.861 --> 00:32:54.809
I'll try to get him on the podcast to talk about that in detail one day.

00:32:55.652 --> 00:33:13.623
But anyway, the point is is that you're not only using AI to predict an outcome, to predict if spoiling appears or not in this specific combination of temperature and other properties of your material, of temperature and other properties of your material.

00:33:13.623 --> 00:33:21.905
It's about investigating why AI believes it will happen or not, indicating which variables are most influential, exploring the space, trying to really understand the physical phenomena.

00:33:21.905 --> 00:33:29.901
And what's beautiful about AI is that it works on multiple variables and combinations of those variables that a human would not be able to process.

00:33:29.901 --> 00:33:32.826
And, yes, it does give an amazing outcomes.

00:33:32.826 --> 00:33:59.867
It really shows us some things that we would not see as a human, and I think in this application, you can consider it giving the scientist another pair of eyes, another pair of eyes through which the engineer can view the problem in a different way, in which the scientist can study the thing in a different way, which perhaps leads to new conclusions, new insights and new models to be developed.

00:34:00.509 --> 00:34:05.724
And I think the use of causal and explainable AI in fire science itself.

00:34:05.724 --> 00:34:07.489
That has a bright future.

00:34:07.489 --> 00:34:30.949
That's a use of AI that I would support, not to use to predict fire phenomena, but use AI predictions to understand which aspects of the predicted phenomena are important and which are less important, so you can better channel your energy resources into researching those parts which have the biggest chance to be most impactful on the outcomes.

00:34:30.949 --> 00:34:34.791
And in this case, well, I am very happy with that.

00:34:34.791 --> 00:34:36.403
Three, maybe even four.

00:34:36.403 --> 00:34:42.619
Perhaps this is going to be, in five to 10 years, a technology that will give us new breakthroughs in fire science.

00:34:42.619 --> 00:34:44.668
I would really really love to see that.

00:34:45.619 --> 00:34:56.340
And one more use that I'm actually quite optimistic about and you can classify this either as a science or engineering is use of machine learning classification algorithms to look at the data.

00:34:56.340 --> 00:35:07.952
So the classification algorithm is a sort of AI method that takes a bunch of data points and tries to group them based on very specific variables.

00:35:07.952 --> 00:35:12.481
That looks into that, and what it does is basically what we do as humans.

00:35:12.481 --> 00:35:26.030
When you drop a scatter plot of your data and you start seeing that, okay, some of the data is on this end of the plot, another data is on another end of the plot, you can figure out why some data points are here, why some other points are there.

00:35:26.030 --> 00:35:34.701
The limitation of humans is that we usually do this at two-dimensional plots, so there's a very limited amount of variables that we can look at at the same time.

00:35:34.701 --> 00:35:37.748
It's very difficult to work with three, four, five variables.

00:35:37.748 --> 00:35:40.623
Actually, above five, it becomes almost impossible.

00:35:40.623 --> 00:35:41.425
We have to group them.

00:35:42.327 --> 00:35:51.050
And for machine learning algorithm it's doing the exact same thing, so it basically plots the data points in a scatter plot and then figures out the relations between them.

00:35:51.050 --> 00:35:58.451
It just can do it in whatever amount of dimensions it wants, so it can look at 30 different variables at the same time.

00:35:58.451 --> 00:36:05.911
It can look at combinations of variables, extremely complicated combinations of variables that we would never be able to process as humans.

00:36:05.911 --> 00:36:19.351
It of course, requires enough data and enough quality of data to support this type of analysis, so it's not something that's going to give you answers directly, but a tool that can help you process your data to find those answers.

00:36:19.351 --> 00:36:30.541
It gives you another dimension of looking on the data that you've collected, on the data that you have, and allows you to really really, really figure out something new from the existing data that you have.

00:36:30.541 --> 00:36:40.510
And this makes it a very, very exciting application, especially that you can also look at the data that was gathered tens or dozens of years ago.

00:36:40.510 --> 00:36:49.025
You can combine all the research done since now on a specific fire phenomenon and look at it from a way no one has ever looked at it before.

00:36:49.025 --> 00:36:53.304
I think this is a very, very exciting application of AI.

00:36:53.304 --> 00:37:01.813
It does not give you a direct answer, but it guides you to an answer that you would otherwise not find so very, very interesting.

00:37:01.894 --> 00:37:06.445
I'm actually very optimistic about the use of classification models for fire science.

00:37:06.445 --> 00:37:16.213
So clearly, three and a half, maybe even a four for that, because I see it's something that in next five to 10 years will yield breakthroughs in fire science for sure.

00:37:16.213 --> 00:37:17.943
So we're almost there.

00:37:17.943 --> 00:37:42.847
Now let's move to the world of fire testing, something that could impact my own institute, itb, and we understood that many years ago I think it was 2020, when we've started the first research project into use of AI in predicting outcomes of fire tests and, in general, scoping the possible use, future use, of AI in fire testing.

00:37:42.847 --> 00:38:01.728
And there are two things that come out of this, and let's split it into just using AI as a predictor of fire test, as a replacement of a fire test, and the second part would be use of AI in some sort of digital twinning or expanding the test results.

00:38:01.728 --> 00:38:05.635
So, first, replacing fire tests with AI.

00:38:06.097 --> 00:38:14.014
So you could imagine a world in which you're about to test a load-bearing wall, let's say, and you have a machine learning algorithm.

00:38:14.014 --> 00:38:23.371
You provide the machine learning algorithm with the properties of your wall and the way, how you're going to build it, where are you going to build it, what's the load on the wall, etc.

00:38:23.371 --> 00:38:27.664
You press enter and you receive a rating of 65 minutes.

00:38:27.664 --> 00:38:41.927
Actually, if you just took hundreds of fire tests that you've done in the laboratory and you've created a neural network that was taught on this data, you could probably do an exercise like this already today.

00:38:41.927 --> 00:38:44.052
Now, will it give you a rating?

00:38:44.052 --> 00:38:45.161
Will it give you an output?

00:38:45.161 --> 00:38:48.108
Yes, it will, but what's the trust to that output?

00:38:48.108 --> 00:38:51.362
In my opinion, the trust to that is zero.

00:38:51.362 --> 00:38:55.371
I don't trust it at all, and that comes from my experience in fire testing.

00:38:55.371 --> 00:38:57.344
That comes from my experience in fire laboratory.

00:38:57.344 --> 00:39:12.565
We had an episode with Piotr Tarkowski about the fire laboratory and fire testing and fire resistance and you can refer to that to see how complex it is, and I don't think a simple machine learning algorithm would be able to predict outcomes of a FHIR test.

00:39:13.126 --> 00:39:16.132
There are just too many nuances that go into FHIR testing.

00:39:16.132 --> 00:39:32.485
There are too many tiny details that make a huge difference in the outcomes of your FHIR test to have a generalized neural network predictor for the outcomes of the test, at least a predictor at a level in which you could just simply assign test result based on such prediction.

00:39:32.485 --> 00:39:34.126
Of course you could.

00:39:34.126 --> 00:39:41.608
As a manufacturer, you could probably develop your own neural networks that would help you in product research and development In this case.

00:39:41.608 --> 00:39:46.492
That's cool, that's probably quite useful and perhaps you could achieve good results.

00:39:46.492 --> 00:39:59.931
But we are nowhere there to be able to truly predict the fire properties of materials and assemblies with just machine learning at the level in which we could really trust that development.

00:40:00.010 --> 00:40:03.125
It's a kind of similar to fire science and prediction of fire phenomena.

00:40:03.125 --> 00:40:12.233
It is just way way too complicated and we do not have enough data points to make sure that the neural network has seen everything to predict everything.

00:40:12.233 --> 00:40:21.697
And as soon as you go outside of the space in which you've done your teaching to the neural network, you stop having reliable outcomes of that neural network.

00:40:21.697 --> 00:40:27.815
Now you know the fire test outcomes in terms of fire resistance, for example, is one thing.

00:40:27.815 --> 00:40:34.434
Another is the things that happen during the test, for example the temperature rise and we've done some AI.

00:40:34.434 --> 00:40:50.036
That was done together with the Imperial College London, with Matt Bonner, who was a PhD student back then and now he's an engineer at Trigun in the UK, and also together with Professor Gil Moraine, under his supervision, and together with Matt we've burned a lot of facades.

00:40:50.036 --> 00:40:56.838
We've done this huge database on the test outcomes of facades with external methods.

00:40:56.838 --> 00:41:02.710
It was called Kresniknik and it's a very interesting paper out there in building an environment.

00:41:03.070 --> 00:41:09.231
A follow-up to that was to use AI, to teach AI on the big data set, and we really had a big data set.

00:41:09.231 --> 00:41:15.331
We had hundreds of facades, tests to teach the machine learning algorithm on them and we did that.

00:41:15.331 --> 00:41:27.755
We did some predictions with machine learning on what the outcomes of the test would be in terms of peak, excess temperature and the time at which this temperature is exceeded, what those parameters mean.

00:41:27.755 --> 00:41:33.777
I have to refer you to the paper or perhaps to episode four of the podcast where I've talked about that with Matt Bonner.

00:41:33.777 --> 00:41:46.746
Anyway, we achieved quite good outcomes of the machine learning, so it was able to predict in a not very bad manner the temperatures and other properties that we were assessing.

00:41:47.306 --> 00:41:49.391
However, it was not significantly better.

00:41:49.391 --> 00:41:51.536
Actually, it had some spaces.

00:41:51.536 --> 00:42:01.467
It was worse than simple regression models and for me, the lesson was that machine learning perhaps is not always the tool you would like to use.

00:42:01.467 --> 00:42:11.798
There are tools that we know, there are tools that we can apply widely, that have perhaps better control, better trust to them, and that can give similar outcomes.

00:42:11.798 --> 00:42:26.108
So, yes, you could possibly do that, but if you have such a vast, such a good data set and you're only interested in certain outcomes of those tests, you can achieve it easier with more simple methods.

00:42:26.108 --> 00:42:37.681
And if you want to predict a complete outcome of a FHIR test at a level where you can really really trust that assessment, it's not going to happen, at least not in five to 10 years.

00:42:37.942 --> 00:42:46.217
Perhaps in 2030, where many other branches of engineering are replaced by world of AI.

00:42:46.217 --> 00:42:48.588
Perhaps then yes, I don't know, maybe.

00:42:48.588 --> 00:42:52.838
But yeah, as I said, the FAR podcasting will be replaced first.

00:42:52.838 --> 00:42:58.518
So when that happens, you know FAR testing is somewhere next on the list to be replaced.

00:42:58.518 --> 00:43:00.632
I hope it's not going to happen very quickly.

00:43:00.632 --> 00:43:02.331
I'm sure it's not going to happen very quickly.

00:43:02.331 --> 00:43:08.577
It's also about responsibility, about, you know, about having a person that can go to jail if something goes wrong.

00:43:08.577 --> 00:43:12.516
I mean, that's, in a way, stupid but kind of impactful.

00:43:12.516 --> 00:43:33.951
This responsibility aspect of our engineering is important and just because of that, many of the applications that we talked about in this episode will not happen because the trust is not there and AI cannot be responsible for its action, and people who do not trust AI will not want to be responsible for the AI outcomes either.

00:43:33.951 --> 00:43:40.106
So, yeah, that's a massive wall out there that prohibits the use of AI in fire engineering.

00:43:40.186 --> 00:43:48.733
So, for fire testing, if we talked about predicting the complete outcomes, the fire resistance of an assembly, I would give it one out of five.

00:43:48.733 --> 00:43:50.918
It's not going to happen anywhere soon.

00:43:50.918 --> 00:44:02.309
To use it as a support tool to investigate the outcomes of fire tests in a better manner, like with the explainable AI used in science, yeah, that could work.

00:44:02.309 --> 00:44:04.253
I would give that three out of five.

00:44:04.253 --> 00:44:27.309
I think it's quite good application If you're about to design a large research project and you would introduce AI in some sort of digital twinning of the experiments, have great data sets, have great control over how data is created, which points you investigate and how you process the outputs and you apply causal explainable AI to a research project.

00:44:27.309 --> 00:44:29.856
I would say that could be even four out of five.

00:44:29.856 --> 00:44:40.679
I think it could be a really exciting project, but you would have to have good resources for that and have really good control over what you supply to teach the neural network there.

00:44:40.679 --> 00:44:48.317
So, yeah, potential use, but again, in a generalized outcome I'm not very positive.

00:44:48.317 --> 00:44:49.518
It's going to make a big impact.

00:44:49.518 --> 00:44:54.634
One out of five and finally, some future technologies.

00:44:54.675 --> 00:45:09.052
So, as you've noticed, I said I'm going to use the scale one to five and I've not given a single five yet because there's not many things that I see AI being a true paradigm shifting tool in far engineering.

00:45:09.052 --> 00:45:23.092
My enthusiasm has went down a lot as the more I work with those tools, the more I learn about them, the more I understand limitations, the more I understand challenges related to use of AI.

00:45:23.092 --> 00:45:42.851
You see, it's an extremely powerful tool, but to use it very well, to use it to the highest of the potential you can get from AI, to use it at a Nobel Prize level, you have to involve Nobel Prize resources in that work.

00:45:42.851 --> 00:45:45.396
No, we simply don't have that in fire science.

00:45:45.396 --> 00:45:53.369
If we had, we could apply those resources not just in AI but in other aspects and possibly achieve more breakthroughs out there.

00:45:53.369 --> 00:46:09.847
So the technology itself is brilliant, but I just don't see fire science being made big or rich enough to really tap into the most powerful capabilities of AI and machine learning revolution out there.

00:46:09.847 --> 00:46:22.387
It's simply not there and most of the uses that we see every day are rather simple applications of AI to very limited data sets, in a very limited way, which give interesting results.

00:46:22.387 --> 00:46:24.693
Yes, they give very interesting results.

00:46:24.693 --> 00:46:37.570
They give us very interesting insights to our data, but the ability to generalize that into more complex space is not there yet, not at the level of involvement and resources that we have.

00:46:37.570 --> 00:46:56.012
However, if I was to brainstorm what would happen if we had unlimited resources, unlimited capabilities, would there be a true paradigm shifting use of AI in fire engineering and in civil engineering in general?

00:46:56.012 --> 00:47:01.969
I would say I see such an use and that could completely redefine the way how we design buildings today.

00:47:02.530 --> 00:47:07.809
That goes somewhere in line of the golden thread that Dame Judith Hackett has mentioned.

00:47:07.809 --> 00:47:12.347
It goes along the things that we've discussed with Mike Strongren in his podcast episode.

00:47:12.347 --> 00:47:16.855
It's about use of AI to assure compliance with the law.

00:47:16.855 --> 00:47:25.577
To explain it in simple words, if you think about the design process today, what we do is, as designers, we propose a solution.

00:47:25.577 --> 00:47:39.338
Then that solution goes to a third party, is evaluated, then authority having jurisdiction approves the solutions, puts the stamps on the solution, on the paperwork, and from that point the solution is approved and can be built.

00:47:39.338 --> 00:47:51.273
Now the process is tedious, the process takes a lot of time and once you have your design stamped, it's very difficult to change it because the stand means that it has been approved.

00:47:51.273 --> 00:47:53.152
You have to build what has been approved.

00:47:53.152 --> 00:48:00.755
If you want to change something, you have to go through the approval process again and again, and this means additional resources, additional costs.

00:48:00.755 --> 00:48:05.824
If you want to change something, you have to go through the approval process again and again, and this means additional resources, additional costs.

00:48:05.824 --> 00:48:18.940
It means more complex process and sometimes people would actually come up with better solution for some aspect of their building but would not implement it because the cost related to changing the design would outweigh the gains from the new solution at all.

00:48:19.625 --> 00:48:36.313
And here I think if we had a general AI that could be used to assure compliance, like you would have, ai that could be looking at your project and checking if the project is compliant or not, ai that you would trust.

00:48:36.313 --> 00:48:43.670
It would have to be extremely powerful AI that would be able to understand the design in vast detail.

00:48:43.670 --> 00:48:51.996
But if we had that eventually and it could run continuously over your project, it could just assure compliance in real time.

00:48:51.996 --> 00:49:02.780
Once you send out something to be built, okay, you cannot change that, but until it's built, you can do those changes and you can immediately see if the change is approved or not.

00:49:02.780 --> 00:49:18.318
If we had that, you could cut off months of the design process because the changes could be implemented later in the project, because some technical details could be chosen later on, and we could completely change the way we design buildings.

00:49:18.844 --> 00:49:25.668
In the modern world it takes many years to design a building and many decisions that we take early on are completely unnecessary.

00:49:25.668 --> 00:49:31.534
We're like, really, we choose some technologies five years before they are being built in our building.

00:49:31.534 --> 00:49:40.538
In that five-year time, maybe there's a better technology already available at the moment you're building it, but you cannot deploy it because it was not approved in your design.

00:49:40.538 --> 00:50:06.166
So if we had a way to drop the verification of our projects and compliance on some super powerful AI that could truly change the way how we design buildings, that would be a paradigm shifting thing, and I think this actually could happen quicker than a breakthrough of a scale five in fire science and predicting fire phenomena.

00:50:06.166 --> 00:50:09.836
I think fire phenomena would be way, way too complex to achieve such a AI.

00:50:09.836 --> 00:50:12.291
But in design I think we could eventually get there.

00:50:12.291 --> 00:50:13.494
10 to 15 years.

00:50:13.494 --> 00:50:22.798
We could get to a point where you would have a third party checker for your project that could just run everything against the code and just tell you if you're compliant or not.

00:50:22.798 --> 00:50:25.373
That goes into, of course, authorities.

00:50:25.373 --> 00:50:33.617
You have to have machine-readable codes, you have to have some tools that will help the AI cross-check your projects.

00:50:33.617 --> 00:50:45.478
But I think we could get there and I think this could be an application of AI that I could value at level five really, but yeah, not something that will happen the next day.

00:50:45.945 --> 00:50:57.134
So anyway, guys, these were my predictions for the near and later future of the use of artificial intelligence technologies, and I wonder what's your opinion.

00:50:57.134 --> 00:51:01.128
So I'm not that excited or optimistic about it.

00:51:01.128 --> 00:51:04.376
Perhaps in most of the aspects I was rather pessimistic.

00:51:04.376 --> 00:51:09.954
As I said, the more we work with it, the more we use it, the less excited we are.

00:51:09.954 --> 00:51:13.447
And, yes, you can do amazing stuff with that.

00:51:13.447 --> 00:51:22.213
But to do amazing stuff you need extraordinary effort and extraordinary abilities of the programmer to use it properly.

00:51:22.213 --> 00:51:29.994
It's not something that is easy, it's not something that you can drop extremely complicated problem on the chatbot and it will solve it.

00:51:29.994 --> 00:51:38.465
Actually, one more trigger for this episode In the Imperial Hayes Lab we have this chat and one of the friends, nick, who was also in the podcast.

00:51:38.826 --> 00:51:51.278
He dropped an example that someone has been researching the age of facades of the buildings and the person was showing images of facades to a chatbot, asking them how old is this facade?

00:51:51.278 --> 00:51:53.853
It was for, let's say, medieval buildings.

00:51:53.853 --> 00:52:08.971
Anyway, this was so wrong in so many levels, so painful to see this application, because the person got some output from the chatbot but there's absolutely no way to quantify that output, how correct or true it is.

00:52:08.971 --> 00:52:14.688
The use the chatbot was designed for is a language model, not a facade aging model.

00:52:14.688 --> 00:52:28.068
If you wanted to do this correctly, you could do a visual library, hundreds of facades, teach the neural network how to distinguish between older and younger facades, and perhaps you could achieve that.

00:52:28.068 --> 00:52:35.641
You could achieve a visual inspector of facades, but it would take you a lot of work and that person has taken a massive shortcut.

00:52:35.641 --> 00:52:37.905
Instead of programming their neural network, instead of teaching it on the correct data set, that person has taken a massive shortcut.

00:52:37.905 --> 00:52:47.945
Instead of programming their neural network, instead of teaching it on the correct data set, that person simply used chatbot to generate output and, yes, it worked, in terms of that person got some answers.

00:52:48.648 --> 00:53:00.768
And that's the risk when you ask a question you will always get an answer, but the trust to that answer, the trust to that output, is absolutely minimal, and this is a brilliant example of the challenges.

00:53:00.768 --> 00:53:15.855
Like we can get output, we can get answers from our AI supportive technologies, but the trust to them is limited and unless you know very well how it worked.

00:53:15.855 --> 00:53:22.338
Hence the power of explainable and causal AI, because we understand more about how it got to the answer.

00:53:22.338 --> 00:53:24.179
We just don't get the answer itself.

00:53:24.179 --> 00:53:33.822
If you don't know how it got to an answer, the trust to that answer is very limited, unless you know the answer on your own and you can cross-check AI.

00:53:33.822 --> 00:53:39.211
But the question is do you really need AI for that?

00:53:39.231 --> 00:53:48.608
So, in real science, in predicting fire phenomena, in replacing fire testing, in doing more complicated things, I'm not that positive.

00:53:48.608 --> 00:53:50.331
We'll get there in a few years.

00:53:50.331 --> 00:54:00.092
In replacing some mundane tasks, in helping us in our everyday work, in supporting us in what we are doing anyway and just helping us to do it faster.

00:54:00.092 --> 00:54:15.172
Yeah, the AI technology is there to support us, and a fire engineer who is in five to 10 years, a fire engineer who is fluent in using AI in supporting their work will be way more effective than a fire engineer who is not.

00:54:15.172 --> 00:54:16.630
So I highly recommend using it.

00:54:16.985 --> 00:54:18.085
But, dear friends of the Fire Science Show, it's not so.

00:54:18.085 --> 00:54:18.284
I highly recommend using it.

00:54:18.284 --> 00:54:23.838
But, dear friends of the fire science show, it's not gonna replace fire safety engineering anywhere.

00:54:23.838 --> 00:54:28.992
Soon it might replace fire podcasting and I'm terrified by that that.

00:54:28.992 --> 00:54:43.974
I hope, when the day comes and there are amazing ai generated fire podcasts, you will still pick the genuine thing and I will be there for you delivering more fire science every Wednesday.

00:54:43.974 --> 00:54:45.498
That's our agreement.

00:54:45.498 --> 00:54:48.186
So thanks for being here with me today.

00:54:48.186 --> 00:54:53.137
Thanks for listening about my rambling on AI technologies in fire science.

00:54:53.137 --> 00:54:59.724
If you're anywhere in New Zealand in Wellington or Christ Christchurch there's opportunity to meet in person.

00:54:59.724 --> 00:55:12.784
I'm this week at FireNZ conference in Wellington and next week, until Wednesday, I believe, I'm in Christchurch and I'm happy to meet as many of the Fire Science Show audience as possible.

00:55:12.784 --> 00:55:14.804
And yeah, that's it for today.

00:55:14.804 --> 00:55:16.646
See you here next Wednesday.

00:55:16.646 --> 00:55:17.246
Cheers, bye.