Once in a while someone gets lucky and makes a guess that happens to match experimental reality. But it's more like a lottery than a demonstration of the predictive power of maths. They get a career and possibly a Nobel prize, and all the many, many people who made equally plausible predictions that happened not to match experimental reality are quietly forgotten.
Game changers like Einstein and Newton turn up occasionally, but they're very rare. We're desperately in need of one now, but the current academic system would probably exclude someone like Einstein. ("Works in a patent office" won't get you into Nature.)
The problem now is it's getting harder and more expensive to test new possibilities. The Standard Model is obviously incomplete, SUSY doesn't seem to be working out, there are a lot of loose ends, but it's hard to sell funding bodies on bigger accelerator designs if there's no serious prospect of finding new stuff.
Great examples you got, because those two did very little of the "here's a theory, let's see what it can explain" the GP is taking about, and focused basically on "here's some weird data, let's get a theory for explaining it".
https://www.dartmouth.edu/~matc/MathDrama/reading/Wigner.htm...
Also these quotes are interesting:
https://en.wikipedia.org/wiki/The_Unreasonable_Effectiveness...
https://en.wikipedia.org/wiki/Unreasonable_ineffectiveness_o...
It's not that we shouldn't use math, but I suggest that when using Math we are predisposed to want clean, simple equations that explain everything.
There aren't many clean, simple equations (relatively speaking), compared to long, ugly ones.
Each time you make an equation longer, you have a huge variety of choices for the new bits; if we take "ugliness" to include things like asymmetry, arbitrary constants, lack of relationships between the parts, etc. then there are more ugly equations than beautiful ones.
It's more practical to look for simple, pretty laws since there are far fewer of them to try. Once we've exhausted them, we can try making things a little uglier, and seeing if that works; and keep incrementing.
After all, if you have enough epicycles you can explain any motion you like (since you're basically taking a fourier transform)
Where is the math that predicted Dark Matter and Dark Energy?