I took on the role of maintaining such a package a while back. It's called jnumeric (available on github and installable via maven).
JNumeric kind of has a weird history though, dating back to the early 2000s. It's never really been functionally equivalent to NumPy (or even numeric, which is what it's actually trying to emulate), and while it was "good enough" for what we were using it for, to use it as the primary number-cruncher in a Java program is probably not a good idea. It was a bad enough idea that we rewrote our application from scratch in Python so that we could use NumPy instead of trying to do vector math in Java. For that reason, jnumeric is undermaintained, and should probably silently fade into non-existence.
I recently noticed a new project pop up on Github, Numpy4J, which may have a brighter future.
While I know it doesn't quite address your question, I am curious why you would want to move to Jython for scientific code. Java does not have the nice number crunching and plotting libraries that Python has. ML libraries like Weka have Python equivalents in scikit-learn. Imaging stuff like ImageJ has an equivalent in scikit-image. Statistical packages exist in pandas and statsmodels. What is your scientific itch that Python does not scratch?
If you want to move to Jython in order to interface with an existing Java library that cannot be easily ported to Python, I would consider JPype rather than Jython.