Behnam Nouri is one of the AI experts at Siemens Energy. He talks to Peter Seeberg about the simulation of gas turbines and explains how he and his team were able to reduce the calculation time for simulations using machine learning.
Thanks for listening. We welcome suggestions for topics, criticism and a few stars on Apple, Spotify and Co.
We thank our partner HANNOVER MESSE
https://www.hannovermesse.de/de/
Job offering: Heilbronn is looking for a professor
https://cdn.hs-heilbronn.de/ce3c0933328f570c/d9e7413e6a15/232-P-IT.pdf
Sora Video https://www.youtube.com/watch?v=fG3IE9dkyKY
Stefan Suwelacks Posts https://www.linkedin.com/feed/update/urn:li:activity:7164215963301703680?utm_source=share&utm_medium=member_desktop
Benedict Evans https://www.ben-evans.com/presentations
Behnam Nouri https://www.linkedin.com/in/behnam-nouri-33381063/
this podcast is presented by Hanover Messa your leading event for industrial [Music] AI hello everybody and welcome to a new episode of our industrial AI podcast my name is Robert Bieber and we have a very weird situation because normally Peter and myself record this episode but Peter is traveling with an unreliable internet connection in Asia so now we could have gone and clone his voice instead Peter asked me to share some topics with you first a reminder from last week’s podcast from chiao Pang Lee he suggested that while the US and China have been leading in geni development by building Foundation models the opportunity for European companies maybe in creating value back by applying foundation models and we know since the weekend we have a new Foundation model it’s called sora and Sora and Industrial use cases how do they go together very very fine and Stefan zilak has explained it wonderfully and I have a quote for you so Stefan wrote Sora is a glimpse into the future of engineering okay the result of the text to video tool from op AI are not only amazing to look at but they reveal an incredible understanding of physics and that is important I don’t know if you know the video by Sora with this ships in the coffee cup and what is fascinating for me and for Peter and for Stefan maybe also the flags are waving in the wind and the ships moving along with the waves light bouncing of the water surface very very rich istic in order to render the scene the model has to First implicitly construct 3D models from the text input whoa compute the physics of the scene render the scene and now what comes together these are three core task in most engineering workflows free core task and Stefan is absolutely right gen is not only for marketing it’s also for industrial engineering workflows wow and one comment by myself if you want to know more about the technology of Sora I link to a YouTube video in the show notes so Peter is in Asia I was in Austria to plan our event AI in the Alps thanks to hanov fessa for being our event partner and our guests this year are Professor Dr Yannis branter together with Dr Sebastian Lena they will hold a workshop on simulation and AI our guests are trustifi they will talk about certification and our guest is Professor Dr Marco Huber from frover who will discuss AI use cases and one spoiler we will have an episode on simulation and AI with Johannes in the next weeks my news for today I have a reading suggestion a very interesting article for all the decision makers in the industrial sector by Benedict Evans It’s called Ai and everything else more in the show notes so that was it I hope I got all of Peter’s topics correctly we will now move on the main part and hopefully Peter will have an internet connection by the next time so Peter all the best to Asia enjoy your holidays and see you hear you soon bye-bye hi there welcome to a new episode of the industrial AI podcast my name is Peter seberg and I’m your host and my guest today is benam Nuri he’s the AIML tools expert for knowledge-based engineering and platform design with seens energy benam and I are going to talk about design process acceleration by means of integrating machine learning hello benam hello Peter thank you thank you for inviting me you’re welcome looking forward to what it is you have to share with us let’s start with you introducing yourself to our industrial AI podcast listeners please hello my name is bami I’m working now since 15 years on turbine optimization I started to work with Ros Royce on a project called virtual turbine where we virtualized the full turbine the first time and then I switched 10 years ago to zman’s energy where I was responsible to building up big process chains big multidisciplinary optimization process chains of high efficient cobine blades and veins and since five years I’m more and more interested in integrating machine learning into our process chain because we saw there’s a limit of computational time and what we really require to get a robot design so we integrated piece by piece machine learning tools into our process chain to get the designs faster and more robust and here we collaborated to Zan energy with Zan’s digital industry together to create some products which I will talk to okay good yeah you and I we met already at the first AI INX conference that Robert and I have started it was actually the AI in what we call the Monas yeah that was in word spoke right and I do recall also I kind of from the very beginning that that we met I kind of linked you always with turbines and that’s what I’m going to ask you maybe to give our listeners a small introduction into turbines you just mentioned rolls Wars the first so the first thing I think oh turbine that’s has probably has something to do with with these can I call them engines I don’t know but you will tell us that you know make sure that the plane will fly I think later on we’re going to talk about gas turbines so I think there’s different types of turbines what is a turbine and why will we later on be talking about them yeah a turbine is in short like a machine which converts mechanical energy into some sort of rotating energy and then you can generate electricity which we do at seen energy or Ros we just generate trust and push the airplane forward and wind turbine it’s working the same so we use the nature the wind to rotate a big shaft and the generator and generate electricity and the same applies for coal power plants so there we have steam turbines right we’re not going to go into the details of the politics behind it but you already mentioned so I think we concentrate maybe later on on gas but the same could be done you say and maybe you’re going to tell us a little bit more about your company as well and then we’ll hear that wind or coal and there was one that I heard about and I just want to share it because I think it’s rather interesting you should confirm if that is actually correct understanding of me but there is what I would have called 100 year old project I would have called it a water turbine but I think the official English name would be more like Hydro electric power station called valan that’s south of Munich and that started exactly January 26 1924 so what you then call that and I do recall that I visited there maybe 10 years ago and what I was seeing there so when where the water falls down that would be a turbine as well right yeah yeah that’s a water turbine so you use the force of water and height difference to rotate a shift and power generator in the end the the first power generation in the US and Germany were all driven by water turbines in the beginning and then they started to use cold okay yeah and now we we want to try to move away from coal I guess towards wind and probably also still gas is what we’re talking about okay very good yeah I recall that 100 years ago and the guy who who proposed it the engineer and then the the responsible people or I think what they said is like in town in the city of Munich not all the streets were having lights people were not having electricity necessarily so they said why are you going to build this why are you going to build this and now one 100 years later they’re very happy to have already Alternative Energy turbines so tell us a little bit more about seens energy I think it’s separate from seens but still connected related and maybe you’re going to share with us what parts of the energy Market that you play in yeah so zens and zens energy we were one company so I was hired in zans are 10 years ago and in 2020 I decided to divide this into two companies so zens I they still hold like the trains and the digital industry where we collaborate closely together and zement energy we have all the transmission business so smart City business uh we have electrolyzers to convert electricity and water to hydrogen and we have gas turbines and V turbines which our our main business and we we really care about also the energy transition and here Gand play also a big role because they are reacting very quickly so with this renewable energy Revolution and like countries which where you have a lot of sun and H the people are going home the sun goes down and suddenly you need a lot of energy at evening to pour up all the air conditioning and here gas turbines play a vital role so in all the countries so even Saudi Arabia Texas or California they all need gas even if they are blessed with a lot of sun and wind and yeah also Germany the gas toine demand needs to be high because we know we have periods of up to two weeks where there’s virtually no wind no sun and then you need a resource which is reliable and you can power gines in two different ways so one is the combined Cy where where you add the steam turbine at the end there you can reach up to 65% efficiency with our best 9,000 HL turbines and single cycle where you just power the gas turbine and you can react very fast you can just turn it on for one two hour and then turn it off again okay thank you very much for that just that you know some of you listeners will have known that but some and including myself maybe have not been that aware I think I’m going to ask you one final question on what should I I mean these we’re talk if we’re talking a gas turbine it’s a huge thing what what is the size what should I think of when I see pictures yeah so the size of a gasan could be like a burlin double decker bus it’s 12 13 M long a diameter I think of up to 3 4 M and the nice thing about gastan is they have a lot of power in in a very concentrated space so so one gas turbine can generate 300 400 megawatt of electricity and this you can build in one building compare this with wind turbines where the biggest on show wind turbines have 6 megawatt and have a length of 200 M so you need a lot of them with a capacity factor of 20% that’s a typical capacity Factor here in Germany okay that’s like a factor I don’t know Factor 80 or something like that yeah so so if you want to replace like 300 Mega gine fully with wiin yeah you have to multiply it by five and then divide it by six and then you know the numbers of wind turbines you need good thank you for that today we’re talking about uh the acceleration of the design process I assume that we’re probably going to use the gas turbine as an example and your claim that I’ve read is that integrating machine learning into the design process has the potential to increase the Computing time from Days down to minutes maybe let’s start that you explain us how today or how yesterday your design process worked and then maybe you’re going to tell us how you’re changing it yes so we introduced Gest Prov with an efficiency of 60% in the year of 2000 and since then we increase efficiency up to 5% and every time you increase efficiency by 1% % you have to get 100° hotter turbine entry temperature the thing is we haven’t found any metal our competitor or not any metal which just can hold 500° C without melting so we need a lot of very complex geometry and cooling Technologies coating and then you need a lot of simulations to verify that all the disciplines you need to calculate to verify that turbine blade fulfills it is increasing exponentially like in the year 2000 with its hundreds of calculations and now we have to do 100 thousands of calculation in all disciplines like CFT FAA conjugated transfer tell us tell us tell us one or two of these acronyms what they stand for FEA for example yeah FAA is fin element anales which is used for structural mechanics that we have very time consuming elastoplastic calculation which takes on our best computer cluster up to 70 hours and CHD is a conjugate heat transfer where you use cftd to solve the energy equation with the fluid and the solid to get very accurate temperatures and here you have to be very very accurate so you need a very fine mesh and which is very time consuming and CP intensive okay so what is the tools you you you use for exactly accelerating you say you know where maybe in the past you would needed days you can do to days in minutes are they internal tools is that products you only use yourself are they tools that are designed by maybe semens themselves and they are available on the market or what are they exactly yeah so so every guest have I develop we have a combination of commercial tools and in-house tools so you have inhouse tools which built the geometry of the terine blade and then um commercial tools like ANM our course they just help us to do the simulation so they’re very computationally intensive and then in the last 10 years we introduced process chain so before it was always one designer is doing the aerodynamics the second is doing the thermodynamics the third the structural Integrity test but now we we combine this more into a process chain so we can calculate all disciplines at once which also takes a lot of time yeah and then here we introduced our first machine learning models to surate some of this disciplines piece by piece okay what type of machine learning are we talking about can you not go too detail but at least at the level of are we talking yeah what kind of machine learning is it reinforcement learning are we talking generative what what is it that you’re doing that you introduce that you I mean simulation I think you talk about the simulation process simulation has been done long time even when I started computer design 40 years ago but what is it that machine learning is uh is changing so to say yeah so even 20 years ago there were some simple surate very basic linear or polinomial surrogates which you can replace basic scolar relationships piece by piece but now we we explored so we had a we had a competition here at zens was called Innovation F and then we we applied for it we pitched in front of the CEO and then we won small budget to to explore a lot of Technologies and there we tried all kinds of machine learning models which were suitable like random Forest CNN and so basian neuron networks and and we just tested which of them gives us the best prediction was the least amount of training STA and can you share with us which one was the winner or is that internal information you would rather want to keep internal yeah published some papers about this oh you did okay yeah yeah I mean lot of good models where basan and the Goan process they were both quite good and then what was also very very important is that these machine learning models could predict their own certainties so you could ask the model what efficiency will this turbine blade have and I can say Okay 90% but I am very confident or I am not confident and if it’s not confident we say okay then we need more trainings M so we need more simulations to make them more more confident and then ice think is about this model is say yeah you need to trade data creade simulation data for weeks train for minutes and then the predictions is in seconds milliseconds and suddenly you could go from thousand members to 100,000 members and then you can you can upgrade your the probability or the like yeah the probability of the model that it says itself so and then at a certain point you’re going to be happy with I don’t know at 90% or where where’s that’s it’s a it’s a very general question it’s always asking like yeah yes as you say we’re we’re working with with with probability machine learning probability Basin models go models at what point is then there the cut off where you say okay now now we’re not going to put more effort into because we’re happy with x% yeah this this is case dependent this is really case dependent sometimes we are happy with 98% sometimes a very intensive calculation we say we are happy with is 90% confidence it’s really case dependent and then really also what is the optimization criteria now interesting that you mentioned Bose the basian uh I recall Bas is it the what what was his name uh far well I forget the ter now the um is was a kind of priest Bas or you know G you know they are they are all I guess we call them part of the machine learning but they are approaches that are not you know not from the last couple of years right I mean we’ve been doing day for many years now but and that is that is the question on one hand you call them out specifically and I understand you are they help you so they help you move forward then now the other I’m going to combine this the question with what about you know newer approaches like generative design whatever the base of those approaches is is that something that you looking at as well I mean does generative have a role to play in again accelerating the simulation design process yeah exactly so these are our newest research projects where we look into generative design or also on 3D surrogates so surrogate model where you can just give any geometry and then it can give you all the predictions you need but this is really the newest type of technology and then here we see also all the small startups which are building the Sur Gates so they all have different models they all have different approaches so so nobody has found now this is the right way so this is still in a think explorative phase so where where every company is testing what is best also we are testing it with some PhD students they just use different models which have been used for complet different applications like image recognitions or some biological like medicine applications and then we used it on simulation data sometimes they were very successful sometimes not but this is something we now have to explore and I really believe in 3 4 years we will be in a stage where with some small piece of training sta you can build like a generative AI model which can do all all the predictions which takes weeks in seconds minutes now when you talk about predictions I’m going to ask you a question that turns it almost the other way around it’s like you just mentioned I think healthcare so I’m not sure if we then talking reinforcement learning I think we talk it we call it still generative but it’s like so designing the new proteins right and I know exactly I know what the outcome needs to be so I have a number of variables uh whatever that is in healthcare I have no clue but you know there’s a b c and d and they need to have specific values because that’s the new medicine I’m looking for now my question of course then is going to be you know we heard from you what it is that you need to do is improve efficiency from 60 to 65 so you still have 35% ahead of you for the next couple of years yeah that is a car no limit so 100% so there’s a thermodynamic limit yeah I understand that but nevertheless so the question then is do approaches exist and because that would in my understanding typically be generative to say you know I need an efficiency of let’s say let’s be realistic you know 68% and you tell me how the design looks like is so it feels like the other way around does that exist does that make sense yeah yeah so we have optimization criteria yeah so you could say okay let’s push it to I don’t know 67 68% efficiency and then see what is the condition into the ter so you can push the temperature quite easily the simulation up and then you have to verify that all this the other disciplines hold and this is the the difficult part so you have your metal geometry you have your Coatings your coolings and then you have to do all the simulations and then verify all the disciplines so you have aerom mechanical discipline plop plastic discipline and all the disciplines needs to be into our review criteria because our customer want to run the turb for years not replace it after every year run and you want don’t want to have it to be the size of two buses it should still be the size of one bus or or smaller I guess yeah yeah even even more compact would be nice okay maybe if you can you can walk us through a number of applications so it I believe it’s one step more detailed tell us a little bit about you know what exactly it is that you do in the different approaches maybe how you do it we just talked about different approaches is that something you can do yeah so so now we we have this collaboration zens energy and znti there’s a tool called Heats and heats you you build workflows so you can couple your geometry building tool with your CFT tool with your FAA tool and do simulations there and here we developed together a tool called Heats AI accelerator and heats AI accelerator Works in this way that you run your process chain on the classical way so you do simulations which takes a lot of time and after 100 member uh the tool automatically checks builds a surrogate model and says can I replace the simulation with a given certainty so it always checks so it’s doing a decid so inside the process chain so maybe the member 1001 is not good enough but after a certain time remember 200 300 400 you see that the AI replaces more and more simulation so maybe in the beginning 10% and later 30 50 80% and then you get a very good boost and you accelerate your optimization by a big factor this is one tool where we put it in yeah let’s let’s stay there a bit because it sounds magic that’s why I’m quiet it sounds a bit like and it’s probably not but even if it’s not you can then correct me it sounds a bit like you know I started looking into chess and I’ve probably mentioned that when what was that Alpha go and then Alpha zero I believe but that was more about the algorithm only knowing the rules so of s so the rules would be you know the rules of whatever Thermo uh design uh the details of your engineering but then it it runs Millions it does millions of games and at the end it’s it feels a it felt to me a little bit like you know it becomes more certain and certain and then at some point in time it starts winning and then it’s um You Can’t Win anymore from it maybe that’s a different approach from what you with it’s a different approach yes it’s more like you have a a surface and all this surface you have pressure temperature and other like results and then you you train the artificial intelligence what happens if the turbo plate is a little bit lean pressure such side or it’s sicker sinner yeah it changes the shape so we we are changing the shape and we giving the AI that happens if we do the TR blade sicker remember that and then piece by piece it learns from it and then after a while it says okay now I’m sure I can predict all the geometry optimization you’re doing with the surrogate and that’s the the thing and it’s not like chess you you can just train one surrogate and it it builds the turbine blade for any engine you have to retrain every time so every every engine type 50 HZ 60 HZ bigger smaller and then also the different turbine components you have to retrain model very good you are going to share a second third application with us yeah so the second application which is really important for us is also the the robot design aspect so the the thing is with optimization you can get a very good nominal design so you can design it on Peak Performance Peak lifing and so on and then it could happen that you are on the response surface on the tip of the Eiffel tow so any shaking any variation on conditions so you run the engine in kuate or in Iceland is a different condition or from casting you always have some variations in W seconds and so on so you have to test your your best designs on all the variations and there are a lot of variations so you have boundary conditions geometric so you have to test millions of millions of variations and again here you could do it with simulations but this could take years of calculating or all trainers surrogate and then the surrogate can give you some sort of uh robust design response okay and this is one of the examples where you say this then can happen you know within a I don’t know minutes but a very big factor faster than you could do it in the past exactly yeah several orders of magnitude faster if you yeah train the model with the right data then you can get this robust design answer okay I think there’s one one more on creep is that right that you can talk about yeah so so creep is a aaso plastic calculation it takes a long time because you have to recalculate every time step with FAA so you have to do the calculations thousands of times share with us what is creep because we we did this before there was a misunderstanding I think between what I call drift at that time and I thought we were talking about a running machine kind of slowly moving towards giving data that are different and the model then starts reacting but creep is something different right yeah so so if you have a piece of metal and you eat it m it gets larger elastically and then you cool it and then it goes down oh okay that’s the creep but if you heat it too much and pull on it with a huge Force which really on the G then you get into the elastic plastic deformation all right yeah and then we we have to calculate this inside our design limits okay that sounds almost like how you produce uh metal objects on not yeah when the steel is when this when the steel is fluid or not you can you can put it in the form that you want and then you cool it off I guess or t g is very very hot yeah right right right so cre is that is then something you can basically say you do not want to happen or can you say it will happen will happen but it just stay inside the limit so we can verify for customers okay and what do you do then to calculate creep now with the new machine learning based approaches yeah so we again the same thing so we do FAA calculations and we save the data in a specific format and then now we are testing several models to predict this time intensive calculations and then we can do temperature variations geometry variations and then the the simulation which took several days on 500 CPUs can be done in yeah seconds that’s good yeah that’s really good yeah yeah so then other other people can use the same CPUs uh in the time that you don’t need it yeah exactly yeah so there’s always a competition between all the sure CES compostion aerodynamic everyone wants CPU yeah access to yes on one hand you know we hear you know everybody wanting to buy the what is it h100 Nvidia or whatever it doesn’t matter what processor and more and more and more we all need to design chips Etc but at the same time if you say you know we’ve been using specific approaches and we can now I don’t know what the number is I just say is it 10% or 1% of the CPU time to do even better things then that’s that’s almost the development on the in towards the other direction right yeah the thing is in the past you need 10,000 CPU hours to design a 60% turbine blade but now you need Millions so so the number of simulation that what you need to create a high efficient product goes exponentially up yeah like the CPS all right so bottom line we still need more yeah this Nvidia a100 v00 they get very important for model training so this this highly efficient gpus we we did benchmarks so on the CPU took us weeks to train and on the GPU hours so that’s why the Nvidia Shares are going up so much exactly you need this GPU power for modate training and of other companies trying to get into that market I’m again reading Patrick Ginger from Intel but that’s another topic we’re not going to get into today we’re actually going to come to a close here maybe you can share with us what is the what is the status of you know the topic of today applying machine learning to the industrial simulation design process if you look around the world you gave already as you are aware where your turbines as an example are going to be placed in around the world you talked about California Saudi Arabia just as as examples so where are you where is your competition we don’t need to know names but how is applying machine learning to the design process what is your view around the world and how is it going to change what how do you see the machine learning into the design process change over the next you know the complete way of doing things over the next 5 to 10 years yeah so so in our industry I think it’s in the proof of concept phase so we using it for some very challenging designs and putting it piece to piece to all designs and I think all our competitors are in the same phas so it’s getting piece by piece integrated into our design process some Industries like the Formula 1 car design they a little bit further so what I heard is say completely optimize on the surrogate so simulation is just there to feed the surrogate model with more confidence and then the optimization is completely done on the surate and this is I think the way it will go either you make parts for cars for turbines for Aero engines or airplanes data will get more and more important so you need to save your simulation data and feed piece by piece your surrogate model until your surrogate model can build every piece of your product with a good confident and then you can design faster and better and that’s I think will be the the change so in the past it was always you do simulations to to design to verify that your cut geometry is in the right shape and can be manufactured and I think in the the future in five years and 10 years we will all feed a big generative AI model and then um this generative AI model can be run by experts to do very very challenging designs watch F let’s stay as a final on on this very topic I want to understand so is this gen model and is that you know some kind of foundation model is that an uh industrial Foundation model that we have talked about in the past for the last year you know everybody’s talking about it I haven’t seen it yet or do you see that uh while it may be a foundation model as such that could then be generally available to everybody at the same time you’re going to fine-tune it as one option I don’t know with your specific seamen energy and you said Formula One whoever and everybody’s going to use their is it then simulation or whatever engineering Knowledge from the last in your case 170 years or whatever this seens is what what is the the approach there of both the foundation and the specific knowledge yeah I think it will be a combination of models it could be a big foundation Model A large language model which would help like a complete new engineer to guide him how to design so it really would answer questions and guide him what steps to make but I think for all the predictions and so on it would be really a combination of several VAR variable alter encoders GN MLP models which together do the best predictions with the data they have right so still you as Sean energy would be able to to produce your gas turbine moving towards I don’t know what your goal is we just heard 60 65 70% and your competition they would use their own engineering Knowledge from who as long as maybe some other ones will use new ways of doing things but it could be on a common Foundation model then the question of course is who’s going to produce such a foundation model who has the interest in doing that is that you know does it need to come from industry like you know from seens and together with competition as far as that is allowed or who would who would do that yeah that that’s a big question yeah okay will we have a lot of models from a lot of startups so will someone invent a model which can understand all simulation and Engineering data let’s see who is going to come up with such a model it feels like the more interviews we do around generative AI that it always comes up let’s see who going to who’s going to be the first and and who’s going to be the second third if we’re going to have this one big or if we’re going to see in the course of the Year many different uh in this in our case interest specifically industrial fun Foundation models benam thank you very much thank you Peter the listeners if you want to get into touch with bam you can do that best on LinkedIn bam b e h n m NRI n o u r i will’ll put the is LinkedIn profile in the podcast notes as well otherwise if you the listeners have any questions comments please send me a short email Peter aiot dode uh I’m very happy that you stayed with us so far looking forward to have you with us again and bham thank you very much and have a nice day thank you goodbye [Music]