From Objective to Optimized Design Web App Generation Powered by Stochos and GenAI

Hello, this is Ertan Taskin from Ozen Engineering, and in this video, I'll be demonstrating an application from objective to optimized design, web app generation powered by Stochos and GenAI. This work is a continuation of the two previously published blogs and videos. I will demonstrate those here.

You can go ahead and check the Ozen website resources and blog. There are two references that you will find. The first one is the probabilistic inference by Bayesian optimization application on the 2D benchmark case.

We have a simple shower head geometry where we have an inlet region and outlet region, and outlet inlet conditions specified here, as well as some outflow region specified on the bottom portion of the geometry, and six openings, circular openings, for which the diameters of them are specified as parameters.

So, I wanted to modify the parameters and get the optimum design to get the highest pressure drop, as well as the uniform pressure distribution in the outflow region.

So, in the first application, we used probabilistic inference Bayesian optimization in the optiSLang and performed the analysis to get the most appropriate designs.

In the later application, then we continued to utilize the data generated while preparing this work, and with some additional codes, to create an app, a web app, to predict the geometries of the conditions that are already studied in this optimization work.

In this particular work, I will change the topic slightly to create an app already to change the output conditions, such as the objectives, and achieve the optimized design utilizing a web app. So, this schematic work frame is shown with this figure here. So, we have two scripts in this work.

We have a train model script, which utilizes the geometries from the previous work, as well as the results files, again from the previous work. And the Stochos trained model script utilizes a DIMGP regression model, a deep infinite mixture of Gaussian processes, to create a fitted model.

Once the fitted model is generated, we use a second script to create an application. The application uses the model, as well as the previously generated results files from the CFD work.

Finally, we opt-in in the app browser, so that we can go through changes on the objectives and immediately see the results. I would like to briefly show you the script, particularly this is for the train model script.

As I mentioned, this script reads multiple files, both geometric and scalar data-wise, and it generates the 3D model with the regression analysis.

And if I scroll down, I will show you the corresponding section of the code where the functions are read, as well as this DimGP regression function creates the trained model, and it saves the model with this name: PressWellModel.

In the second phase, we're going to go to the second script called App Generation 3D. As you see, this is a pretty intense code, but it basically again reads the CSV files, as well as the previously generated trained model, to create an app.

The train model script is run under the Spyder platform, and when it was run, it took a couple of minutes, maybe 5-10 minutes, totally to complete, to create the optimized model. After it is done, we have to use a different strategy to run the app generation 3D code.

Since this is going to be performed under the streamlit function, we are going to have to use a separate bat file. So, I tried to zoom in slightly. So, we have two different sliders here.

As I said, we're going to get the optimized design, and the corresponding parameter distribution, such as pressure, in this case, and move the sliders to get the most meaningful objective for us. For this case, it actually is the higher pressure drop. It can be highest or higher.

So, as you see, when I put the slider, the perform function runs, and then it creates the file, which we haven't seen yet, because we have to push this update mesh and prediction button.

And also, for the uniformity wise, we want to have a less pressure differential, and that's why it's actually giving us a better uniformity. So, once we select those with the sliders, clicking this button is going to bring the corresponding design with the pressure distribution.

Yeah, so that design apparently requires large holes. They are not even, apparently, depending on the size.

Keeping in mind that I deliberately made this shape of the exterior shape non-symmetric, so I just wanted to see the impact of that non-symmetry on these internal holes when it gets to the pressure, when it gets to the objectives. So, this is the optimum design to achieve these conditions.

So, let's assume I want to get totally opposite end and bring the uniformity all the way to the right side. And click again. This is the corresponding design. In this case, the pressure scale is fixed.

Pretty much it doesn't change, but you can see the distribution is impacted, as well as the design. It's as fast. I have performed this live demonstration so that you can actually see how fast we can get the optimized design with the suggested objectives.

So, this concludes this work, as well as the previous works. I hope you enjoyed this. Thank you. Please contact us at https://ozeninc.com/contact for more information.