> What happens when the best methods for computational fluid dynamics, molecular dynamics, nuclear physics are all uninterpretable ML models?

A better analogy is "weather forecasting".