Machine Learning Models vs Physics Models: The Battle for acceptance

"All models are wrong but some are useful" (George Box) For many centuries, physics models helped us to understand the universe, visualize processes, grasp the main ideas in hypotheses and theories. They were simplified representations of reality, which, otherwise, could be too complex for human mind to comprehend. They were useful because we could understand them. Now we are dealing with different type of models: Machine Learning Models. They are not designed to make a picture simpler. They are not based on first principles or even hypotheses. They are mainly created with one specific goal – to make predictions. In some situations, they can be understood, but in many cases. they are just black boxes that accept some input and generate predictions as output. Are those models “useful” ? Can we accept predictions without completely understanding how these predictions are made? Are there ways to make those predictions more “transparent”? And if they are not transparent enough, is there still a place for them in our pursuit of knowledge?