What you see from that though is that there is a growing set of scientific fields (systems biology, some areas of particle physics, numerical analysis research, etc.) that are really embracing Julia in a big way. Those fields are not as big as machine learning, so some metrics may see that as a loss, but in reality, having a strong core of dedicated users who happen to develop the core packages for their smaller fields is not a bad place to be. At the end of the day, scientists will use whatever language has the packages developed for them. That's why MATLAB and Mathematica are still mainstays: high level open source software ecosystems never really reviled the support actual mathematical languages have in areas like differential equations (here's a benchmark that shows MATLAB is still orders of magnitude ahead of things like SciPy: https://benchmarks.juliadiffeq.org/html/MultiLanguage/wrappe... ). Julia seems to be the first language with package developers taking the "(non machine learning + non data scientist) technical computing scientist who stayed on MATLAB+Fortran because of scientific tooling" seriously, and the ability to be fast enough to write those codes in Julia is the key that binds that all together.
There are still enough web developers in the world that even if a language gets every scientist to use it, then it will still not make more than a blip in the ratings. However, scientists and mathematicians are loyal to their tools, loyal enough to still be majority MATLAB and Fortran, so any inroads means staying power for many many years.