Are you going to put the source-code of your programs in there too? I'm not sure how a ML model would ever know that... say... Pin A5 is a rarely-changing GPIO that's effectively DC. Or that Pin-B6 is configured as an analog-DAC in this particular application (and therefore slowly changing).

Or hell, that PinC1 is an input vs output.

Modern chips have very flexible pins. Many microprocessors (and MCUs) can configure any pin to be input, output, tri-state output, pull-up output, or pull-down output, for example, with different properties in each state. Some pins can be analog (on-board op-amps), others can be digital and high frequency, still others can be digital but rarely-changing / effectively DC for the lifetime of the application.

Each scenario changes how you'd route a trace.

In general, you can organize specifications and verification information in a way that can be consumed by other programs for the purpose of synthesizing designs or generating tests or whatever. At that point, an optimizer could do a a decent job of routing. Maybe ML could add value somewhere in that process by handling some unstructured aspect.