Learning Qualitative Design Criteria

Machine learning can be used to incorporate qualitative factors like beauty and style into a design optimization.

“I know it when I see it”. That’s how U.S. Supreme Court Justice Potter Stewart characterize obscenity in 1964, and the logic applies to so many elements of our lives. Most people would be hard-pressed to define criteria for what makes something beautiful, stylish, or cool, but we always know it when we see it.

From a design perspective, these intangible criteria can be just as critical, if not more so, than the “hard” engineering criteria like velocity, fuel efficiency, or payload. So in a world where computational design algorithms are revolutionizing how products get made, how do we incorporate qualitative criteria like beauty into the process?

One idea is to turn to machine learning. At a high level, machine learning algorithms essentially just extract patterns from data. Here humans and ML models have complementary strengths: humans can identify examples of, say, beautiful or stylish designs, and ML models can extract patterns from those examples. The trained model can then guide and optimization towards a more aesthetically pleasing design.

To be clear, I don’t personally believe it’s possible to teach a machine learning model all there is to know about beauty, but in a limited context, I do believe ML models could be a useful tool for characterizing that which we otherwise have to “see to know”.

Let’s look at a simple example.

Say we wanted to design a bicycle frame that is both structurally sound and athletically pleasing. I’ve created a HyperStudy session with two models:

Model 1 is a good ol’ fashioned OptiStruct model which predicts displacements from static loading. After all, it doesn’t matter how cool your bike looks if it breaks when you sit on it! The design variables are the coordinates of the yellow nodes.

Model 2 is a Tcl script which launches HyperMesh, displays the design, and prompts the user for a score. For this example I decided to give bike frames that I thought looked “cool” a “1”, ok models a “0”, and ugly models a “-1”.

Now the last step is to setup a human-in-the-loop optimization. Throughout the process, the designer is presented with new designs which they can score based on their appearance, and the resulting data is used to train a machine learning model. That model can then steer the optimization towards what are hopefully more beautiful designs.

I chose to maximize the human score while constraining against excessive structural displacement. Below you can see the setup for 30 evaluations of the optimizer.

The optimal design satisfies all structural requirements and, in my humble opinion, looks pretty cool

Besides finding a cool design, this process can also reveal some insights about my own aesthetic preferences. Interestingly, I found that my human score is slightly correlated to three of my design variables, which indicates that these areas of the bike might have a greater influence on its look and feel.

To recap, we used a simple machine learning model to approximate my aesthetic design choices and incorporate them into a design optimization. The model also taught me something about how my preferences map back to the original design variables. Note that machine learning did not replace the human designer, rather, the two performed complementary tasks that result in a more efficient workflow and a better design.

Do you have a design challenge with both qualitative and quantitative criteria? How do you incorporate qualitative criteria when using computational design tools? Reach out to an Altair support representative if you’d like to learn more about the techniques used in this article, and happy designing!