Dassault Systèmes Isight is a powerful tool for system integration that lets you link various design and analysis software packages together into integrated models. The output of one package can be fed to another, for instance CAD geometry can be updated parametrically and then passed to an FEA package for analysis, and the software is able to cascade the changes to input parameters through the system to update and extract outputs from each component. This package has been designed to allow many different types of studies like DOEs and parametric sweeps and it offers a suite of tools for post-processing, data modeling and optimization. Although these tools were intended to be used in conjunction with integrated design loops they can just as easily be used to build predictive models from existing data, which this article will demonstrate.
To begin, downloading the sample data set as a CSV file from ConnectMV. We will use the LDPE data set for this example.
Once you have downloaded the file open the Isight Runtime gateway and import it by clicking File from the top menu and then “Import Data From File…”. In general your data should be in columns, with the first row containing the names of each variable. The HDPE data set is conveniently in this format already, although the first row contains an integer identifier for each data point so we can ignore it. The file contains 14 process variables and 5 outputs, so when we import the data we need to tell Isight to use the first 15 columns as inputs. It will automatically set up the last 5 as the outputs.
After this is complete, switch to the “Visual Design” tab and click “Create Approximation”. The default approximation will be executed, but we can configure it to achieve a more accurate predictive model. To do that, simply click the “Configure This Approximation” icon, which is next to the Approximation1 label. Switch to a User Defined approximation and click “Next”, then change the Approximation technique to a “RBF Model” by selecting it in the drop-down box. Click “Next” to accept the model type, “Next” again to accept the inputs and outputs we have already established when we imported the data, “Next” again to accept the default RBF options, “Next” again at the info prompt. We want to specify the number of points used from the data set to cross-validate the results, so on the Error Analysis Method screen make sure “Cross-validation” button is selected before clicking “Next” to accept. Change the number of points to 15 and press “Next”, check the “Perform Sequential Sampling” box then press “Next” to accept the Approximation Improvement Options. The approximation wizard is now ready to run, so click “Finish” to execute it.
Now that the approximation model is complete, you can explore the results in the Visualization tab. The Local Effects is a particularly useful series of plots, it shows you the impact of each input variable on all of the output variables. So, you can see that Conv is driven by F12, Tcin2, z2 and Fi1 while LCB is driven primarily by just z1 and z2. And if you want to see how well the fit was achieved you can click the Error Analysis tab and the Response Fit sub tab to see how the predicted values of the 15 validation points compare to the actual values. Keep in mind, Isight did not use these points in the model, it simply plugged the inputs in to the neural network and extracted the outputs.
As you can see, all of the points are very near the line that indicates a 1:1 mapping of prediction to actual output, so for this dataset, that means that all output variables could be predicted with just a couple of percent error from the inputs. This is a pretty remarkable result, especially given the fact that it took just a few minutes to set up and execute this model.
From here, you can perform optimizations and search for inputs to achieve desired process outputs by using the “Design Search” function. This would let you do things like minimize the amount of expensive components while maximizing the performance attributes you care about. We will cover the design search features in greater detail in a later blog post, but hopefully this gave you a sense of just how powerful and how user-friendly Isight is as a predictive modeling tool.
If you have any questions or would like to learn more about how Isight could add value to your business please let us know.