Everything that is physics is also just networks. Working together.
Then it could be a an idea to model some parts of the algorithm in a 3D modeler like Blender.
For instance I suspect you need simple simulated physical wheel suspension systems for periodic objects.
?Everything can be written in a periodic way. Then I suspect that using machine learning models. With lots of error correction. Could possible brake the network. I guess the process could then be improved with proven real world examples of say suspension. Breaking the network is similar to the error goes to inf.
I got this idea from mechanical inventions trying to solve simple network type problems like showing time. Maybe there exist some advantage in modeling some parts in a 3D simulator than trying to come up with a limited math function.