[00:00:05]
>> And thanks for the opportunity to talk here what I'm going to talk about is the air base design approach you know if you think about the design approaches I'm going to specifically target the complex systems all the systems which take a long time to design in the design process this can be because of too many complex stages it can be because of too many humans involved in the loop right there is a lot of information exchange but it can also be because of the tools which you use tools which are used to analyze the process and to make decisions me.

[00:00:37]
To the decision making process all those can combine lead to a very time consuming process this kind of problem is very relevant to Siemens because you know we manufacture one of the most complex system in the world which is a large cast iron mines we also manufacture a lot of simulation software which are the design software which goes in the process of creating those kind of you know design so so it's very relevant to us you know if you think about it the largest airline can take anywhere from 6 months to 10 years in the design process depending upon the size depending upon the needs so so it's.

[00:01:10]
You know we talk with the business units and we try to start tackling this problem about 34 years back looking at the new approaches also looking at that Mansmann with the technologies bringing from the hardware side but also in the area of machine learning and artificial intelligence so how can we make our software more intelligence how can we make the design process much simpler Well not much simpler but easy to use by the you know the design engineers who do not know the domain expert in that area so we started this about 3 or 4 years back we started to develop certain frameworks and I'm going to talk about you know that framework kind of a highlight and some use cases which goes along with that so.

[00:01:53]
Before going into that I wanted to show discard one and probably some of you might have seen it from the systems in you know inside it's basically kind of shows the gap between intent and outcome right so if you think about it some of my design maybe basically you know want to explain a swing which he wants and then he explains in a fashion which is shown in the in the corner but then you can see what happens at Barry's stages the information get lost and because of human in the loop also you know what you end up getting is very different from what he wanted and even he.

[00:02:28]
Can sketch to my mate self doesn't know what he wanted he may have wanted something race as simple as that so. The point is that you know there's always a good intention when you start the design process but what you eventually get me something slightly different or very different depending upon how you how do you do various tasks in the design process and how you you know.

[00:02:51]
How do you address them so with the machine learning tools we not only want to do automation but also make intelligent decision process and so we get what we intended for. So the outline of my presentation I'll just give a brief overview of the concept which we are bucking on for last 23 years which is that generally means you're in for design.

[00:03:16]
And then I will talk about some of the core technologies within that framework which goes along towards development of that kind of. Framework for application to different use cases and then I will talk about a couple of use cases from. From 2 different areas where we try to put together the stack knowledge.

[00:03:35]
In order to address certain issues and finally I want to you know just put in a slide various research issues which from my point of view from Siemens point of view which we see which still need to be addressed and that's why you know folks over here. You know it will be good if you guys can see a table.

[00:03:53]
Those topics and try to address those issues so that we can then take those. Methods you develop into commercial software is. OK so before going to talk about you know what whatever your generative engineering agenda engine for design is I just want to you know briefly touch on what is a current design process if you look at the current design process it broadly in 3 Category 3 steps the 1st step is to define and define is that required when defining the requirements defining the problem or from leading the problem in such a way that you can start the process the information the.

[00:04:32]
2nd the state this creates So that's where you start to create solution and that's a problem that's a step which takes a lot of time basically because especially in the complex system a complex parts where you have a lot of design engineers the basically tried to come up with what you call C design so initial design based on their knowledge which they had in the in their mind the tacit knowledge also based on you know what kind of design process they have done in past on the similar product and they tried to utilize that to come up with a new design kind of a variant design and then try to tinker with that to come up with more solutions and the lasted basically then they evaluate all the solutions and if you're lucky you get the right design right away but that's probably never the case what you end up doing is that you know the design which you generated may not be meeting certain requirements or it may not be.

[00:05:24]
Good enough for what you like so you don't basically start we've created this process and you go through this iteration several times now and one thing I'd like to point out here is that this iteration loop is not like you know optimization integration do but it's rather the loop which the human ingenious goes through the whole process if you think about you know a guess that binds which is made of thousands of components each components are being designed or tinkered with when you want to come up with a holistic design and there's a big loop which goes to one loop like that might take several actually months so that's that's that's what happens in the real scenarios.

[00:05:59]
And in then what happens is that either you use up all your time which is which is a loaded a look at it to basically do the design or you've used up all your dollars So what you basically choose this design which kind of almost meets your requirements on me to requirement may not be optimal and you move forward with that and in the end what of happening is that in reality you are focusing on one design and just trying to tinker with it so that it meets all your requirements and other issues that.

[00:06:30]
Many a time in this kind of design process the domain experts are not always talking with each other they day basically do their part and forward the information so the knowledge is siloed it's not there's no one person who knows about everything so in that cases the system of the process becomes very inefficient and usually a lot of iterations are required to achieve this.

[00:06:51]
So what we want to do is basically do we were the digital enterprise hold process by 2 steps one is digital is in the engineer and the physical world knowledge and what does it mean this overall approach if you're going to work as more knowledge based approaches knowledge oriented approach what we want to do is to capture the knowledge as much as possible from wherever it can starting from the human engineers who have knowledge inside their head how do you have how do you got to the Desert Knowledge basically the knowledge of discovering from the engineering data it can be specification sheet or any other kind of you know data regarding the problem and also if you have made similar products in the past and they they have been operating in field can you basically utilize that information to learn more about the problem so that you can you know do a better job of designing this process so so the 1st step is capture and what you really want to achieve is that instead of you know are you basically taking one design in thinking that you want to explode thousands and thousands of design in a very fast manner but you want to discover the right design from those designs so so what we want to do stupid sickly capture the knowledge and then we want to achieve the design discovery instead of tinkering with the design and the way we can achieve that is basically to have this is in a design engine where we use machine learning techniques to basically speed of the process of design generation as well as the analysis of the design itself so which is one of the big bottlenecks So how can we achieve that that you know the designs get generated automatically and then automatically analyzed so that the engineer who is who basically has his main task is more on looking at the come out coming solutions and then be able to better decisions about what.

[00:08:30]
Wants and how the design should look like. OK let's. So so all this work we had right now which we are doing if you're not being a number black hole or anything we have been collaborating both have believe it internally with our business units but also externally with university and other in the silt blackness we're working a lot with the government agencies like the Department of Energy or D.M.V. I on government funded projects to basically you know develop models within this general design frame we're also working with university collaboration especially with Georgia Tech I'm very happy we have a lot of collaboration a collaborative project ongoing right now a lot of Paulino here and also the media Mabus and lease on with whom we have projects we have several use cases so it's not that you know we're trying to design something and apply to one use cases like to actually get close to 26 use cases this is a slightly older slide but these use cases you know across you know for Didn't from turbine design to basically transformer designed to application to building technology to grades and stuff so basically.

[00:09:41]
We developing a lot of these use cases where we can showcase these kind of technologies but we're also developing concepts for example it's very hard to see but we're developing a framework in which you can we can build a pipeline of machine learning technologies without. Too much tinkering so that you can directly apply to a problem and this is kind of framework we want to give it to other design engineers so that they don't need to know exactly about how this of Gotham is going to work or how do I put you in the parameters.

[00:10:09]
For them to study so you know it becomes easier for them to use we also developing a lot of technologies from just a machine learning and. Tools We're also looking at distant memory which what it means is that when we're running these use cases it's generating thousands and thousands and millions of data right how do you actually automate how do you store it in such a fashion so.

[00:10:30]
That you can use it later on so we're working on a concept as to you know the vision story of these results we're working on autonomous element simulation and validation which is basically how do you how do you develop techniques so that the simulation itself that automatic out automatically created is right but I meant it and so you can even run it automatically and then do the validation process also automatically The other aspect we are looking is the most intuitive and interactive system for whatever trade analysis so what it means is that especially if you think from the design engineer point if you you know you do you give him a piece of software and you ask him to trust it but he while he's using it he wants to be part of you know this design process so how can you basically Bill.

[00:11:18]
You know the tools with the engine is also in wall in the design space exploration process and they can directly utilize their knowledge about in the design itself so you know you basically that the design process in a specific direction the design space. We're also building tools for more efficient design space exploration which is basically when you're generating thousands and thousands of solutions what is the best way you can present those to the design engineer you know how can they explore all these different solutions not only looking at several of them but also if they want to go deeper and look at $11.00 design and lot more detail how can you bring up the Properties which is most important to him.

[00:11:58]
And then we're developing these usable components which is basically that a lot of a lot of things if you develop it's we don't want to customize it was specific use case but it has to be a journalist framework so that you know we can. Implement them across these different problems.

[00:12:15]
We're taking all these technologies we develop different maturity level more towards the other 78 and then we want to hand it to our business unit who can then take these these you know the the the various technologies a lot of things and methods and approaches we have develop and implement in commercial software which it seems for a.

[00:12:40]
So this is this is last night on the United design side but this I just wanted to go were a little bit more detail about what we mean so this is you know as I said I just showed earlier that we have 3 step process in the genetic engineering side with this capture explore and discover so in the capture side be working on trying to basically build up these distillers which take the raw data and then convert it into knowledge from audios So whether the raw data is coming from you know engineer or well what it means is that how do you capture the ingenius tacit knowledge you're not going to go to the engineer and ask hundreds of questions and basically take the answer and build a case what do you want to do is to basically build up techniques with with that technique or methods which can monitor the software which the engine is interacting with and then captured that those set of actions in the form of knowledge so for example we did with our next that we built plug ins which can monitor all human action or interactions with the software tool and then automatically convert that into a knowledge form.

[00:13:44]
We want to also have distiller which can look at the engineering data from the past design data and then extract all important features about those past designs those information and then build up our knowledge model out of it the 3rd is basically you know you take if you start to good was a design process the 1st thing you get is a huge requirements document this thing should do this it's not you know why like this and so on and so forth how can it take that document and directly convert into a form of our knowledge or knowledge graph or are in some form in the mathematical equations directly so my object functions are constrained are getting formulated from their design and from the common stock human we also want to look at you know the past 8 about a product which has been in operation and extract information for that so once we have all this knowledge how do we represent it and that's a big key that you want to have a presentation which the machine learning tool then can and use.

[00:14:39]
Those kind of representations so so we want we in most of the use cases where we are following this this process the knowledge base we are using is basically knowledge graphs. And and when we take all this knowledge and then construct this knowledge graph for each of the individual use cases which kind of embodies everything about that problem of that specific use case once you have all this knowledge what we want is that we want to explore design we don't want a human to basically start doing you know creating initial design but we want to have some kind of methods which is start to generate lots and lots of seed to see designs which you know are synthesize these design and then be able to you know run simulation in extremely extremely fast man manner using surrogates to basically quantify the quantified that particular design in terms of certain K.P.H. of certain metrics while while this process is going on if you think about it there's also knowledge and there could be in this analysis process right so how do you extract the knowledge and then bring it back to the Knowledge Graph which you have so and in order to do that once this new knowledge is getting added You still want to use the knowledge in this design process back again so for for that what we need is some kind of explorer which is dynamically adding the knowledge and then utilizing that knowledge and bringing it back in the design process itself so we build several tools to basically you know do that process.

[00:16:06]
You know once once this you know the design process is going on and it's being quantified the design engineer who's basically right there watching the process of getting the output of of this whole effort he you want your genetics so many so many design solutions what you want to eventually provide is the insights about these decide so what which our designs come out of a successful solution what features you can extract and present to the user in what form so that he can take important you know efficient decision so in our process we have this kind of framework but also tools for each one of the yellow box which you see and that's what the applied to different use cases.

[00:16:50]
Any question still I mean I can take questions this way. That I can to. OK so now let me talk about the core technologies one of the core technologies we talk about is the knowledge and the Knowledge Graph which we're using in our case to basically represent all that knowledge.

[00:17:10]
Why not let go of this I mean there are several reasons just the you can read but I want to just highlight the one in black outline so one of the key things that you found out that working with different problems is that if you if you were to presented the knowledge in a knowledge graph form you can easily discover hidden you know contextual and information from this knowledge graph as you add more and more knowledge you know as you have more notes getting generated you will discover a relationship which was not there in one instance but it may be in the other instance but as you add more and more instances about it.

[00:17:47]
These entities are these relationships will automatically emerge which will help you in your design process another big advantage we learn is that it's very easy to extend and merge new knowledge and give you an example so we were using Knowledge Graph to do subtracting when factoring the baby build this whole huge knowledge graph about you know what are the.

[00:18:06]
Resources what kind of capabilities to scan they can have what kind of operations they can do what are the skills they can provide and we basically you know we use that knowledge graph to come up with the most optimal way of manufacturing let's say part right and then become a bit OK what if we have to add the knowledge about it demand of actually right and in that case it was very easy we just added a subgraph it the same similar schema and then make the connections and it was that and all that a lot and didn't need to actually change so then we can do a hybrid manufacturing you can utilize the Knowledge Graph to do you know how to make a part both using additive and subtractive machines and that was that was that was a fun project.

[00:18:48]
Another aspect is that it's ready it's hard to see but what it says is that easy to discover high level information and basically the idea is that you can instead of using knowledge graph as a full scale You can also do graph of crops where the node becomes a subgroup and when you start to do that you will start to see information and pattern at a different scale and learn about about the system in part which you can then utilize you know for your problem solving.

[00:19:16]
So this kind of shows some examples where of the Knowledge Graph in different use cases we have. So Bob generation is simply means a bill of process generation what is the correct sequence or most optimal sequence of of process which you want to follow in order to create a part in the in order to manufacture part so we use knowledge graph to basically embed the knowledge about the manufacturing process and the design and then utilize that and that's advisor again it's very hard to see but it's actually based on a project which you're doing with that with Professor Dimitri Mabus and the goal was that as the user is interacting with the next how how can that knowledge dynamically gets created and then utilize this knowledge to do a prediction of what user wants to do next and if he selects that that's what he wants to do do all those steps.

[00:20:06]
Automatically So this and then we have hybrid affecting as I talked about the goal there was basically that if you have both subtractive energy machines how can you manufacture parts using combination of these machines so so we basically embed all the knowledge into what we call design graph to achieve that in the architecture exploration it's more of a 0 the kind of analysis where you want to explore the architecture architecture level about your system of systems so for example a suspension system can have several components and you want to know what kind of components should be in what combination which leads to a successful suspension so you use graph and reading this approach to basically you know do clustering to see what are the areas in the design space with least the successful combination.

[00:20:54]
Leading to a successful architecture for suspension and last but not the least is the data stored it so we're learning so many use cases we have we have so much data and for each use cases we have different kind of problem and scenarios what we found is that we can store the data stored all the data about the solution and the problem in our knowledge graph such that we can learn about the problem and solution across you know different use cases so you can characterize a problem in a certain way you can characterize the solution in a certain way also and then you can store the solution for example the base of the new Let folks and everything in these Knowledge Graph and then you know if you get a new problem you can find a similar problem there and then you know try to find out what kind of machine learning tools I can use because the similar problem has to have this kind of machine to machine learning to.

[00:21:48]
So this is just the example of of the Knowledge Graph this is a process generation use case and I'll talk more about it so I'm going to skip it. This is just destroys more detail about it that various types of nodes you know what are the design features notes what were the manufacturing feature notes and you know the relation and the process notes and how you know you can know that this process this one affecting feature can be realized by a certain process and all this was embedded in the knowledge.

[00:22:20]
So this is the next advisor knowledge graph where you know each node has information about that is basically the presentation of the state of the of the next software while the user is making an action so I just basically the action made by the user and Nord is basically the state of an X. and X. software which is a design to and each node will also in capsular all the information about what the features and geometry with respect to that state in the next and this actually gets generated completely automatically at the as the user is interacting with the design software.

[00:22:56]
Let me. OK so this one is. Not a scrap it's we've lies for the hybrid manufacturing use case where the goal was that you know given a given a cad part and given a set of resources what kind of subtractive an additive operation I should take to basically come up make this part so we basically beam embed the design into a kind of a design graph where each node is basically the manufacturing feature and there's a surface and then utilize this graph to basically you know come up with the most optimal operation because each of those when affecting feature can be done adequately as well as subjectively and there are certain costs associated with it so we use this kind of crap so basically explore what kind of you know side of sequence of operation which leads to the least amount of cost and the solution for basically manufacturing these kind of parts.

[00:23:53]
So this is basically no different kinds of disabilities we have developed. Converting the raw data into a different kind of knowledge graph this is more on you know the. The components which can be used across a different use cases and they require a minimal amount of in a customization so all of them have a lot of things which are independent of the problem where the new can customize their interfaces such that you know you can change the schema on one of the and and be able to generate different kind of rubs from that the raw data.

[00:24:27]
So just very quickly when we're talking about knowledge and knowledge graph what is the new knowledge in the Knowledge Graph so it can be anything among these are all of them it can be a constraint which is derived from the past 8 of which you for example you may learn that certain action can not be performed in front if another action has preceded it which that's what we find out in manufacturing use cases but also it can be a new relationship between entities that showed earlier for example you find out.

[00:24:57]
You have to do C.M. before drilling that you want to position where your laser pointer where you want to drill and then you can learn that kind of relationship as you as you input a lot of data or take the past data it can be in the form of generation of new entities which basically means a new kind of notes or Also it can be you know a combination of several routes or several several ages which are great together and we're simply if you look at the policy of the Knowledge Graph and and try to find similar to policy in the knowledge graphs.

[00:25:26]
God it can be a new solution itself once you add the new solution in your knowledge graph it brings the new knowledge along with that solutions so this is just an example of you know what it means to to basically have a new knowledge this was the knowledge of each part of the knowledge cover to short earlier and what ends up happening is that this.

[00:25:50]
Edge is not there but as you start to add more data you start to learn that a certain operation is always followed by another operation and which means that in my solution later on if I add this operation then I should also add other operation also and so you can learn these kind of things from the data.

[00:26:10]
OK So the next part of the core technology is a surrogate models. So we have used several methods this is just an example of one method we use the President at work as a forward model in many of the use cases where the goal is that OK you want to you want to have the design parameters but you want to quickly predict what is the K.P.A. are the metrics so that so we applied we applied it to a lot of use cases this is one of the use cases.

[00:26:43]
Which is on the printed circuit board so we are basically trying to build the filters which are and we want to analyze filter in extremely fast manner so what we want to know noise that given a certain design parameters which can be resistance or capacitors in that filter what would be the gain and frequency so.

[00:27:02]
So we basically and this just showed you know that eventually going to develop this sort of get models and apply apply our data to it to basically do this kind of prediction we get a very high degree of accuracy using those reds resonates. It's clear this are more common form of sort of that model which I think you guys already use but one of the interesting thing which we started to use was also the inverse model so the Ford Model is a is a is a very common form of sort of it models but in most models what you want to do is that given a certain output What are the possible inputs so for example if you look at the forward problem the goal is basically to predict why given.

[00:27:45]
Given an X. and in that case any kind of you know deep network will work very fine but if you invert the problem you're been given Y. but you want to predict an X. if you take new left but it will be very hard for a new little project like this problem so we really basically try to look into different techniques to solve this problem here we just show one which is measured in city network but we also use that other techniques like dance to basically do the same job but the idea there is that instead of basically the need to sort of get a model or inverse model for this extremely non-linear map what you want to use is glossy and functions so the glass in so basically you can select this a set of course in function and the goal basically becomes that you want to learn from data the meaning the variance of these different goals in functions such that you can generate basically estimate the input given a certain amount of output.

[00:28:44]
And then we applied in the same problem to the printed circuit board or the filter problem where there was a certain desired gain in frequency and then for those desired in frequency I can estimate what different set of design parameters are designed solution which can which can be generated and you also get the amount of accuracy the amount of what is the probability of this design to me.

[00:29:10]
Desired given frequency department and it also gives you the information about for a given design which parameter it is most sensitive to so it's kind of a local sensitivity information which you directly get from this kind of analysis. Not this is basically you know how do we put all these things together.

[00:29:36]
So you have solved model and inverse model but now you want to use them in a combined fashion to basically do an optimization which is machine learning driven optimization which is a key part to you know design a space exploration process so if you look at the. General level array of traditional way of doing an optimization what you end up doing is that you start with some cold design goal initial design which is you know the what you call a C. design.

[00:30:05]
And then you basically generate lot of lot of C. designs or designs around it and then you basically analyze the. Through some kind of you know I assume relation and simulation analysis tool or some kind of a surrogate model and then you basically have a certain stopping criteria and if you're solving greatest meant you get a final design otherwise you keep on doing this you know.

[00:30:28]
Performing this loop. Till you meet this criteria so the idea is that maybe it can be in Word this process so instead of starting from a cold design what if you start with a design target right so we have an inverse model we can we can start with a certain target it can generate in sort of generating initial design we can generate hard to be sure design which is basically have a high probability of meeting the design requirements and then basically you know you do small parts around those design hard design points to find out you know if I'm gable to get an optimal design if.

[00:31:07]
Since there is also this is using for was an inverse model which are which have some kind of uncertainty in their prediction so you would NEVER it's very hard to say that you got an optimal design but you need to have a certain criteria to make a decision whether you want to stop at a certain point to say this is my most improved design or else you can use that design as a starting point for the traditional optimization of God and that's what we deal is is that if you combine the traditional traditional optimization of the god of him with some kind of machine learning up to my this and you can cut down the design at least the design of my vision in a in a significant manner.

[00:31:51]
And this is just a heuristic which we utilize to basically do the do the exploration which basically in verse and verse model generates the design solutions and then we take the seats and do part the basin in the local design space to do certain evaluation and find the best design.

[00:32:11]
OK Another aspect is also on the dependency and causality analysis So basically one of the things is that when you're doing this design process very early in the stage you want to know for a given design which is meeting your requirement are not meeting your requirement which are the parameters which are most responsible for the design being in a certain state so you want to do some kind of dependency or causality analysis.

[00:32:37]
What we at least implemented but now is somebody should go to lation based approach to to do the influence analysis which is basically in the network given network I can find out which of the parameters in my system are the input output which are most influential So what are the most influential notes in my in my design which which which affects a certain outboard all I can also know which are the notes which are most influenced by that but I would guess.

[00:33:07]
So basically I can do and I something else is like this where we have we want to we're doing a C. of the analysis and you want to do a prediction about sexes of the same sea of dissimulation and in that case we also want to know which but I mean are responsible for this sex is unknown so we can do this kind of analysis to look at certain input parameters and the size basically kind of tells you like how much how much it's a contributing towards a simulation success and that way you get an idea that when you want to change your sea of dissimulation parameters what changes you should make and where you should make those changes.

[00:33:45]
So I like to go to the words some use cases what. We have. Right all these technologies. First used in the turbine design process will be applied that machine learning base optimization approach so this is a dissertation approach where you have a design experts who would start researching design parameters they have some kind of dynamic analysis tool which gives them some engine performance and they use that along with traditional optimization to tool to do the design process and it can be very time consuming and labor intensive.

[00:34:21]
And so what we want to what we did was basically instead of that design process we applied the machine learning base of migration approach which I showed you earlier and we went from requirements to design parameters and we use this forward model and then this model in the loop to basically come up with a new design so either with that design can we say just as that as a design but it can also be done fed into some kind of a traditional optimization approach to find more optimal design.

[00:34:52]
And this this was the actual problem the which we looked into so this is what you see the design pattern with the different K.P.S.. And for the for the forward model this is for the inverse model and this was this is some of the results which we were able to get so we compared our results with heat software this is.

[00:35:12]
Optimization software to disappear after malicious software what you see here is that in the old method Firstly then he lets 50 to ration and and what we were able to get as an efficiency value was 65.98 we applied our machine learning optimization approach. And about 100 you can ration for 30 seconds what we got is still pretty far but then we we took this design part as the input and provided to the software to the he software and did it within 1 point one minutes we were within 11 iteration what we were able to achieve was a much better solution than what he'd got on its own and so that's that's one of the power of you know using this.

[00:35:54]
Machine learning this optimization is that OK but they may not give you the optimal optimal solution but they can give you a very good insight into what can be the initial design point which you start your optimization with. This. Train. Example. So that's a good question so before this one the data we got from us from actual turbine use case.

[00:36:21]
We have about 13000 training said data for that the more the better it is and that and so days of training time also required for that but once you have these kind of models you can use it for different purposes and different optimization issues. OK this is the last use case which is below process generation so when we're doing design process getting exploring the design you're also looking.

[00:36:51]
Design of the production process so what is what is the best best way to manufacture a certain part or what is the best rate what is the most optimal billow process which can be generated to manufacture a certain turbine blades or brains so this is the use case we did of a written apology nation group.

[00:37:13]
What happens is that. You want to design a blade and you want to have a blade with a new feature which is not the kind of mean slot I think. And in reality what happens is that they have a lot of information about the past data like past in past how this blade was created and the debate was created by a certain certain steps number of steps which is a bill of rights that this is an example of the Bill of process but all the steps are there.

[00:37:44]
And then all this is unmanly So it's a human engineer who basically goes and say OK we should do this step and the text is written here and it's a human written text this whole process is completely man. So what we the problem is that OK Can we somehow automate this process can be so if you have a lot of data about how a turbine blade with a certain manufacturing features what was manufactured in the past can we learn from that and be able to automatically generate a bill or process given new sets of requirements.

[00:38:21]
So that was that was a problem OK this is some. So what So in reality you you get certain requirements of the product. And then you basically conduct your your domain human experts our design experts they basically you know interact with themselves as well as their tools to basically create this what we call routing So how do you go out there different process and also you know that these steps are not all in one factory they're across several factory inside a city or across the city so it's a very complex process and it takes a lot of time where you have you know the results what you get it goes to that you have piece just I'm then that outings are created and then they're sent to send to the production site but actually implement what you want to award this this is very kind of time consuming and there's lot of information and inefficient way of doing it and what we want to actually achieve is that somehow be able to you know extract the knowledge out of the past data and centralize this knowledge so that the user is just looking at the outcome solution to make a final check rather than in the process trying to decide what is the best bit of process to basically achieve the goal.

[00:39:36]
So the approach we followed was $33.00 steps approach which 1st as I said was basically capturing the knowledge and capturing the knowledge about how an Indian here in the past has has gone in various steps to basically some of. The steps to manufacture a specific part so that is that that.

[00:39:58]
Aspect than he has done from his own mind from the past expedient so that that did not contain that tacit knowledge of the domain expertise of the human India so how can we capture that once you have captured we want to utilize that to basically automatically then generate the new voting for a new blade or new rains which we are I want to manufacture and then the last stage is that once you've generated these steps you still want to make sure that you know you have to well we did that outcome to make sure that it works in the real world scenario.

[00:40:32]
So to capture the knowledge you know this we had about 65000 sheets like this and this is a very small one but a lot of times the process which is a human and that information had a lot of spelling mistakes they use different words for different words for the same meaning and so there's a lot of.

[00:40:54]
And you want to somehow automatically extract all the knowledge from from from these sheets so what we end up developing was natural language processing based distiller as well as class of different kind of machine learning classifier which basically reads all this information takes the information this information in this column and then tries to classify each word into a product process of a resource and then learn what steps are creating bitch manufacturing feature so in the end all of those 65000 excel sheets what we were what we learned was that OK if you want to create a feature of feature be these are the 10 different ways 100 different ways of manufacturing it so the so you learn the mapping would be in operation manufacturing operation and the interesting feature and all this knowledge basically gets stored in the Knowledge Graph which we have.

[00:41:46]
Once we have all that knowledge then we get a new problem where the goal is basically to have this new feature in this. Blade. The 1st goal is that basically given this problem what is the closest similar problem from the past and how do you basically determine that and that is what we call the very end so we basically were able to generate get a base base design automatically and then generate a variant of that which is the new design so this so this is a process basically you'd use a lot of base based design which you have determined and then select the features that you want and once you add the new feature it what the searchable team does is that it tries to basically take that feature and learn and take the past 8 of how many different ways this feature can be created and then what is the most of them will be of basically adding these new steps in the existing existing process so you end up with something like this where this is the 3 different way of manufacturing the same same divine blade but with new features this is a base design and these are the steps which are automatically learned from the Knowledge Graph and they can be added I that this a different way of doing it is basically just 2 step process to do different operations and adding some of the and you know at different locations so you can generate several of these kinds of solution and the what you want to basically validate them select which one is the best one in this case all the operation comes with the operational time like how much time it takes to basically set up the process and how much time it will take to do the operations and so the goal was simply that what is the least amount of time it takes to basically manufacture the part and then we basically validated this.

[00:43:38]
Denoted solution over a different software called Plan simulation where all these processes can be implemented and then you can automatically implement it and then basically you can see if I want to have an effect 100 parts how much time it will take or in a V. company parts I can generate and then select the best solution.

[00:43:59]
Well OK so that's almost close to my presentation one of the but for the future direction point of view the things which I thought will be you know quite useful to look into still and we are looking into but it would be great for us to look into 2 is that some of them are listed here one is the 1st one is the interactive design of space exploration so given a lot of time what happens is that the human design engineer they have certain idea about this is the kind of design they want but they cannot quantify in mathematical equations but they can makes the make select certain designs how can we make this whole process interactive you know how can you learn that has acknowledged what's called bias inside the human mind basically which is regarding certain design what are the 2nd is design a space characterization and there's a lot of work still there but it's already done but can I just from the data infer very quickly which portion of the design space which are good or bad ought to be contained certain specific features.

[00:44:58]
3rd is basically automatic constraint of design will generation so so so this is basically if you if you got lot of data about the pass both input and output I'm not looking at just a parametric relationship but you can probably do that but also high level knowledge in terms of which can be explained in natural language form when you did have something like that.

[00:45:19]
Automatic formulation of objectives for your for the optimization problem given that climates given let's say past information and how the past problem object functions well can you automatically form the objective function and also you know the constraints for for those that objective function automatic surrogate generation so this is one of the problems which we just started bit Professor Dimitri Mabus But here what goal is that you know a lot of time you you can't expect you know designing is to be a very good knowledge about machine learning tools you want to ease at least have a process to our tools which they can then.

[00:45:59]
Ungendered the sort of good model if you don't do that even never use these kind of tools so the idea is that you know make their life easier so given a set of data how can a generator get model which basically does you know a tuning of hypodermic automatically right so.

[00:46:17]
How do you capture and utilize human bias right so a lot of time there's a bias in the design engineers head and then how do you capture those kind of bias so that you can show proper designs to him design for designers and intend this goes towards the 1st the cartoon which I had was that how can you make sure that what it was the designers intent and I captured that canadian the stand the intent and what do you actually need and then utilize that but I mean trick mapping across scale so when you do design for complex systems you basically have 00 do you want to do the 3 D. right and then you have but I mean does all across all across these different dimensions can you automatically link those those those but I'm into in such a way that if I make change in one of my demo inches I can clearly see the possible changes with a certain probability in my other damage that's a treaty right but in my ship in the T.V. set going to automatically see the change in my requirements so how can we do that.

[00:47:17]
Automatic knowledge capture from requirements document this is the work which we are currently doing also but but it's extremely complex process which is why it's not only N.L.P. but but learning the relationship between the various portions of the document itself which and the features which are described in those documents.

[00:47:37]
Also time evolving knowledge for design version control so a lot of times in these complex things this list lot of. Wasn't some designs are generated when you captured this this knowledge in the Knowledge Graph but it is time evolving right now you have entities and relationship in a static manner what can we do a temporal is.

[00:47:58]
But also. Learning for cross to mean across physics design problems so if you can you use the knowledge which is solve one problem can you utilized and say SR is a very traditional to transfer learning problem but in practicality it's very hard to use and very hard to generalize and then last but not least with a lot of time to choose when I mean given a problem what what kind which machine and go to time should I use and why you know so.

[00:48:27]
I think the dead. Yeah of course. So you know you never group and in general in scope or technology you're always looking for bright talents we hiding right now and we'd be having me a lot more in a yes we should mention here so if you if you are exposed if your knowledge and if you want to work and the research in any of these topic please contact me or live here and.

[00:48:58]
I think that's a yes thank you. Christian. Bodies. Like. For you very. Very good friends of mine and it's OK So this is just one problem we have a lot more used as a limited vine but it's not so different not only just you know that was just added on to me which I wanted to show there's a lot more or.

[00:49:32]
Less it isn T. O.T. but we can we are not we cannot show it so it was like when you. Know no it's not this is. Just so this is this is just what this is basically just one silly one he told in one an earlier state of the divine design process I think what you're deafening to is that if I doing the overall turbine design right right so so this this is going to help in let's say the 1st state so they spend about 3 to 4 months in trying to come up with a one day design or do you know the design of the initial architecture exploration before they start to see if the N.C.A.A. says this is the spec we just had that stays this is not once you have this tool it will at least speed up that process but you have to build 2 like this for all all different places where you are using some you know more complex.

[00:50:25]
In another project which you cannot should result but we are looking at basically the see if we are trying to speed of the C. of the process for let's say you know one of the components I cannot mention but one of the components of the of the compressor which is very time consuming so.

[00:50:42]
I would I'd. Just not belong. Here. You know 1st thank you so much. Thank you. Specially. What do you. Have a lot of. Wish you could get. Caught up. With. This. Week. At work. That's. Why. This. Happened so I was really curious. How do you do. That probably a question exactly to my business in Calexico not here but I know the problem with your talking because.

[00:52:03]
It's a more it's along the lines of software assurance shouldn't of the system if you change a small portion and this is a big problem in the D.O.D. side also these aircraft when you just make one change but you have to go through a 5 month certification process even with the small change in the right so it's a big issue I'm not sure so I think they will they can be several machine learning approaches are at least not machine learning may be problematic approaches which can be applied we have actually a project along the similar line with DARPA on the short it's called Ace showdown compost.

[00:52:39]
Evidence has basically been using evidence when he visited just to do basically a study when you in a complex system if you just in your piece of. The same audio and how does it how does this. Changes for the whole world system so how to do it for deadline I think that I cannot answer that question from the business side but I know that's a big issue in 10.

[00:53:09]
Years or. So open so. Yes some of the some of the softwares which we developed with the government grants. They are kind of open source so once you finish you're basically. In one case at least and. There's a project which we're doing call and the sending the biosphere trying to quantify human bias that that has a good help on that.

[00:53:36]
But most of the use cases which are to internal you know a lot. But of course the project we're doing with tech and Princeton and other folks that they will be opened. This. War less like. You know. This. Is. This one. Right I think the low hanging fruit is basically to speed up the process it is just too much efficiencies because I mean because of human in the loop and because of just the baby are using these tools that we have to 1st solve this process but the idea is that eventually the design should be you know in what he called on pentagon something which is a human will never talk think about I'm not sure if you are like example and D.N.A. use case you may have seen pictures.

[00:54:48]
Yeah yeah exactly so so that was emitted to you know machine learning to it's rather than a human construct conceive a design it's very systematic it's very you know like it has to be symmetric and there's certain things that goes into it and I'm talking like Nadia For example if you're trying to design a shape right now topology optimization and place like that whereas if you go from machine learning techniques they might not even think about that they were just going from Dr manatee point if you what they were created to you have said so the end goal is to achieve what you said but I think right now we're just trying to also speed up the whole process which is very inefficient right now.

[00:55:29]
Making that point. Because. We also very much. Like. Yes or a pretty particular. You bought. It with. You it was all. For. This was so easy to sort of this create this. This was the exact question I thought it all right is that. It's like those. In that you mention.

[00:56:27]
Here or. There. You know what. You're thinking about. What you know you deal just to do with that answer. Well I. Was part of the research right we're going to find out about. You know your requirements Garcon it becomes very interesting your burns document nails. To go the people I deal with really do just take the knowledge I don't think you can.

[00:57:00]
Compute or to know all that is within that. That's really I mean we get to a very. Very urgent sales in the process of much more kind of a standard Well you know this is a we came up with your stuff here is becoming as though it was in engineering so it is interesting there actually I like to see what is it that you can actually get out of it what is really freaky you know I mean bring.

[00:57:31]
I wish I had a better answer than just I don't. Draw this room for.