Using ML for interactive topology optimization
Altair EmployeeI still remember the first project in which I tried topology optimization. I had spent the first years of my career thinking that finite element analysis itself was awesome - and don't get me wrong, I still think it's really cool being able to accurately predict what will happen with complex solid mechanics problems. But when I was told that I could just give the computer a blank canvas, some requirements, and a goal, then sit back and watch the part being designed for me... MIND. BLOWN. Needless to say, I was hooked. Fast forward 8 or so years and many many projects using the technology, my fascination remains unchanged. Starting any project it would still be the first tool I'd reach for in the tool box.
Problem formulation is the most important thing when it comes to topology optimization. The set of requirements you set up is really the only thing preventing all the material from being removed if your objective is to minimize mass. Or all the space being filled with material if your objective is to minimize compliance (=maximise stiffness). But what if you don't know the exact requirement up front? What if you just want an indication of what an optimal topology would look like? What if you want to a trade-off between mass and stiffness? What I have generally ended up doing is running multiple topology optimizations with different settings (brownie points here for using a systematic way of setting this up - like a DOE in HyperStudy) and then looking through each of them to get inspiration for designing the part.
If you, like me, are thinking about what machine learning can do for product design, then you might have realized that we're sitting on some data here. What if instead of looking through each result individually, we can create a machine learning model that interpolates between the results? In the video below you can see a simulation surrogate produced by the physicsAI library that predicts the topology result of varying volume fraction. Rather than having to look through loads of result files we have one ML model loaded in on screen that predicts the result based on the current settings.
Which parameters do you vary when running topology optimization? Could your setup benefit from applied machine learning? Please comment below or get in touch.