One thing we've explicitly thought about is using TCAD models to improve the corresponding SPICE models. All our simulators are differentiable, so it's pretty easy to throw it into a gradient based optimizer and get a spice model that's fit to the TCAD results