Machine learned spice components

Schematic Simulation (Ngspice)
1

A wild idea…

Machine learning can be applied to any logical model. The net can be taught to replicate the logical behaviour of that model to a very high accuracy. With a small test rig, any physical component can be mapped as well.

Perhaps this methodology could be applied to derive spice models for any component. It would enable emulation of encrypted spice models as well, provided they can be ”tested” by the learning sw. Since the code is unique, the copyright would not apply. All manufacturers components could be replicated and a huge open source spice library can be created. No more compatibility issues either for NGSpice

As a bonus, this ”machine learned” component would run much much faster than the original model. The higher the complexity the larger the gain. All the ml program does is to test all the pins and determine the response.

Thoughts?

2

When are you getting started ?

1 Like
3

You don’t need ML/AI to make a blackbox model of an analog electronic component… It’ called Levenberg-Marquardt (or any other gradient descent algorithm). Of course it won’t model complex internal states (i.e. figure out the instruction set of a CPU based on the signals from its pins :P). But neither will AI/ML…

T.

4

The extraction of behaviour could get very close to breaking the EULA and would keep lawyers employed.