[api]

Outsider here but — is it conceivable that instead of writing a geometry kernel one could be trained? Use the existing geometry kernels and all the open CAD documents you can find to generate gigantic amounts of training data and train a geometry model.

Of course the catch here is going to be the precision required for real world use. A ton of impressive AI demos are just that — demos. They are good enough to wow as a demo. Still, if the data set is big enough, and you’d probably want to run the model itself in at least fp32 precision, maybe you could get something.

There is a body of machine learning work that’s been done on precision annealing of models. Basically you train to min loss and then go into a finer grained domain using something like simulated annealing to fine tune parameters.

[ddingus]

That is what I was writing about above.

Use current CAD and a script to output a LOT of cases to be used for training data.

[zihotki]

Sure, as soon as we could get predictable and repeatable results from a model in reasonable time

[ddingus]

That is circular.