All right. Good afternoon everyone. Thank you for staying so late. I think we are in the last block of the conference. Really exciting so far. Today I’m going to present you how you can put mechanical engineers or designers in the driving seat for AI adoption within your organizations by basically making it easy intuitive for them to use datadriven tools. And I’m going to do that by presenting a use case we did with our colleagues from TMAC.

0:31 But yeah before we start I actually want to ask you a few questions. How many of you have already applied any machine learning methods for design optimization? Okay. Not the majority of the the audience but some people. And how many of you can confidently say if I ask you right now you come to me for an AI project and I ask you how much design data do you have or simulation data do you have how many of you can confidently answer the question or tell me okay I can look it up in a system and I’ll tell you right away no one okay all right yeah I was expecting at least some people have an overview about their data all right so couple of words about myself and the company my name is Asparanov.

1:20 I’m one of the co-founders of Keyword and currently I’m the product lead at the company. I’ve been working in the field of data science for the past nine years but my roots are deeply in mechanical engineering design and testing. I’ve worked a lot also with industrial designers. And yeah in fact actually most of our company come from manufacturing, aerospace, automotive background and industrial design and we truly understand the physical systems that our customers are working with because we also still utilize them and we have worked with them heavily in the past.

1:56 But we also understand how complex and challenging the path toward successful application of AI u in the industry in the engineering industry can be and that’s why we created keywords we created keywords basically to empower mechanical engineers and designers to make more datadriven decisions to make it a bit easier for them to do data analysis data processing and to build u it can be 3D deep learning models like what we saw from physics X.

2:29 It can be just machine learning models on table data whatever you want. All right. Okay. Starting from where we left off again we are a team of all of our team comes from mechanical engineering manufacturing industry and we built keywords with the idea to make it simple for you for designers and engineers to utilize datadriven tools. And yeah for the past decade maybe two decades the industry has done something really valuable and that is the whole industry engineering industry has digitalized and what that means is that companies nowadays are sitting on tons and tons of simulation measurement experimental data sets.

3:10 Which is great but the problem is that companies are barely tapping into that data to make something useful to optimize their design iterations to derive any big data insights. And yeah companies very often come to us and the first thing one of the first things we ask them is how much data do you have for that particular use case and it usually takes them a few days sometimes even weeks until they can give us an answer and in a majority of the cases the answer is also not correct.

3:42 They either have they think they have a lot of data but turns out only 10% is good data or they think they only have few data points but turns out some other team somewhere has many more data points. So yeah, why is that? Why do you think is that? First, data, engineering data comes in many different shapes and forms, many different formats. It’s usually scattered across multiple teams.

4:07 And the reality is that many of these teams do not have the capacity really to gather this data, process it, and make something useful out of it. So AI and machine learning is increasingly becoming a necessity in the engineering and design field but the path to it is still very risky complex and AI remains underutilized and one of the main problems to the AI adoption is that the engineers and the designers lack control and lack the tool actually make something useful out of this data and this is where basically keyword steps in.

4:41 We want to empower engineering and design teams to make more datadriven decisions by integrating directly into your design and simulation workflows by providing you AIdriven analytics u processing and very intuitive machine learning and deep learning capabilities via no code and we want to empower your engineers to do that themselves. Now for the past five years since we have been around we’ve done various projects and here here is one key thing that we learned in order to have a successful AI project you need to have a indepth domain understanding about the engineering domain about the physics you need to understand exactly what is going on there in physical terms but of course if you want to build an AI model you also need to you also need to understand how to build AI models you need to understand how to build architectures you need to understand how to process data, how to provide a scalable infrastructure.

5:38 Then once you train the model, you need to know how to deploy it for other people to use it. And let me tell you something else that we validated and seen over and over again. Engineers are not data scientists and data scientists are also not engineers. But it is far easier to educate and teach engineers how to use datadriven tools than to teach data scientists how to do engineering.

6:01 And this is why at keywords we believe that tools must meet engineers where they are and engineers need to be empowered to operate and drive the AI adoption within their companies. Now to ground this into an example today I’ll show you a use case that we did with our partners at TRAC. We first met Try a year ago in the US at a conference where they have presented a very interesting paper about their first baby steps applying machine learning and deep learning for design optimization.

6:35 And the results were promising. Everything looked great but it was hardly scalable beyond their Jupyter notebook Python script. And that same day we also presented our paper on how to derisk AI adoption across organizations. They happen to be in the audience and they saw their challenges directly addressed. So they reached out to us and we shared the common goal how can we make AI and deep learning accessible for mechanical engineers and designers who do not necessarily know how to scale this and build sustainable models.

7:05 So yeah, we decided to replicate their use case and they essentially gave us 50 of the data points that they used for their own model that they developed and what essentially took them more than three and a half months to complete were able to replicate within our platform within one and a half days from start from the road data to final train model that delivers almost the same performance.

7:30 Now you see the data this was a pressbench assembly data with three parameters varied they provide 49 simulations coming from solid work simulation they provided them in the form of p files and what we essentially did is we use one of our graph neural network models within the platform to teach it to predict the stress fields and the displacement fields based on on the input geometry. Okay.

7:57 So before we dive deeper into the process, I just want to give you an overview of the whole steps that we’re going to go through today. The presentation is more about the workflow rather than the actual results but yeah you can see on the far left side step zero just collect the data. You need to find where the data is and as soon as you find it and give it to the platform the platform can extract everything that’s contained in this row CE files can be in any format can be from seammens can be abacus doesn’t really matter and then once the data is extracted it’s converted into a structural structural shape that can then be analyzed it can be processed and clean and can be transformed to be parsed to any of the AI models that you want to train within the platform.

8:39 After you train the model, deploying an actual model in production to multiple users is just three clicks away. You can really create instantly multiple apps and just share them to the people that you want to utilize. And there you have your results. So let’s take a deeper look at the first step. So this is the first step where we extract the raw data. As I said, we worked with Solid Works simulation files in the MP format.

9:05 So using our client extraction tool we process this into a structured data model. And here you can see a representation of how one data point looks like. You can see it contains the original project name software that it was originating from even the file size. It contains boundary conditions scatter fields vector fields and you can all visualize also the underlying mesh within the platform. Now the key here is the speed of execution.

9:34 Like all in all it took us four minutes to set up the extraction and then the total extracted data was done in 40 minutes. That’s for 50 data points. And the other key here is that you’re bringing all of these raw unstructured data on a unified data format. Unified data model meaning that you can combine this FEA structural data with other data sets from abacus or whatever other software that you’re using in order to build a larger database.

10:04 Okay. So now if we have our data extracted and put in a structured shape then we can proceed to analys to analysis and this is where you as a data scientist or as a engineer who is empowered to use data tools. If you want to train your machine learning model you will start first by analyzing correlations between the parameters that you want to use for your model.

10:27 You start analyzing distributions see if you have uniform distribution everywhere. You don’t want to have any outliers because they can spoil basically the performance of your model. And what we essentially did is we saw that there was one simulation file that was lying way out of the distribution for the maximum stress. And we took a deeper look and turns out that simulation was not executed properly, did not merge.

10:50 Also the file size was very small. And yeah, we essentially went through many steps. If you want to show to see more details, come to me after a presentation. I can show the platform in real real life based on this case. And eventually we had our cleaning data. All in all, it took us approximately 1 hour to complete. Now that we have our cleaning data, we were ready to process it and put it into shape that we can feed to our U governor network model.

11:20 And to do this, again, everything is graphical. No code. The only thing you need to select here is what kind of operations you want to apply to your data. And you start first by selecting since we have a 3D case here. We need to select which regions exactly of the original geometry we want to use to feed to our AI model to teach it to learn the stress fields or displacement on it.

11:40 So that’s what we did. We selected volume regions. Then we applied an operator to convert it to a graph. Then we applied some clipping and at the end we had the full pipeline. All in all processing took approximately 15 minutes to complete with a setup. Of course we are experienced we know how to do it but it took us 5 minutes. Okay. So once we have the clean data we are ready to train the AI model within the platform.

And as I said here we’re training a graph neuronet network model. And for this model, we selected the geometry as an input together with the three design parameters that were varied. We had that information. Some cases you don’t have that information, so you don’t need to necessarily input it. And we also selected some boundary conditions as as constraints. Then for the output for the target for what our model has to learn, we selected stress fields and displacement fields and our model was ready to go.

12:37 We spin on we spin the training and this is by far the most timeconuming process took approximately four hours per iteration. And eventually after two iterations we had our model train on the next day and once the model is trained using it and deploying it production is again just three clicks away. So we have these tiny applications within the platform that you can select to build and then you can deploy to many users within your organizations where the only thing they can see you can restrict them.

13:10 They can see certain models developed for them. And here the only thing they need to select is the model they want to use from a list of available models. And then they need to select basically what kind of new design they want to test on or new parameters they want to evaluate on running the tests as we saw also with physics X is instantly it’s super fast.

13:33 No need to talk about it. It’s a datadriven method. And that’s one of the advantages of it. So you can put it into an optimization loop and so on. Now the focus for me today is not to show a lot of the results but I’ll show you just what this model is capable of achieving. You can see on the left side the ground truth for the stress fields and on the right side you can see I hope you can see clearly actually.

14:01 You can see the predicted values on the test set that was not used for training and we can see that visually they u yeah align quite quite well. Now again, I doubt that this model is very generalizable. It’s only 50 data points. If you want something that generalized over more cases, you need more data. But that’s not the focus now for the for the demo. Now I just want to summarize again the whole process from start the beginning.

14:26 U if you are going to approach this by yourself and you want to build such deep learning model pipeline by yourself that would take you a few months to complete. This is the experience that our partners at TRACK had. It took them it took them two three weeks just to set up the pipelines. Then couple of days for analysis, another couple of days for processing. For the model train, they had to dig deep and learn all about architectures and so on.

14:54 And then once you have the model trained, then you have another bottleneck where you need to take care of all the prediction inference infrastructure, how to deploy to your colleagues and so on. With the platform that I just showed you. Data extraction setup done in minutes. Data analysis, AI powered. Yeah, everything is interactive. You can sync across multiple visualization charts if you want to. Data processing again setup takes couple of minutes to complete.

15:23 Model training setup again couple of minutes, but then execution can take a little bit longer. And once you have the model ready, you can deploy it directly to your engineers. And that’s what Tramik did. They instantly deployed the model to four or five engineers from their company for utilization. All in all, it took us one and a half days to complete. Now I want to end my presentation just with couple of key points that are important for us.

15:47 AI will definitely have a significant impact on the whole industry on engineering and design industry. However, the path to it, it’s not about having the the newest and fanciest AI models. It’s about having the best data. And organizations need to start putting effort to organize their data. Now, they need to have a good data strategy. They need to ensure that they manage and collect and organize their data.

16:18 Even if you’re not building models now, you might be building in the future, but even if you’re not building any models, just having an overview of all your data assets can deliver you a lot of insights. And yeah, another thing that I want to emphasize again is that engineers are not data scientists and data scientists are not engineers, but it is far far easier to equip engineers with data tools than to teach data scientists engineering.

16:43 And that’s why the tools must meet the people that are closest to your production and design where they are. That’s it. Thank you. To learn more about the CDFAM computational design symposium series, to see the archives of previous presentations, and to learn about future events, visit CDFAM.com.