Also I wonder how much scientific rigour went into the recent progress in AI/ML. At least in principle there might be ways to determine invariants and boundary conditions (without modifying models), also considering there had been efforts to change ML models and make them explainable.