Following the discussion on geoschem/geos-chem#2443, we have determined that the LSODE integrator is not thread-safe. That is, if you try to call the KPP Integrate function from within an OpenMP !$OMP PARALLEL DO loop, the integration will eventually fail with convergence errors. However, LSODE still works properly in single-thread instances.

I believe this is caused by an old-school F77 COMMON block that is buried deep inside the int/lsode.f90 routine. I am going to do a bit of testing to see how we can replace this legacy construct with modern Fortran while not causing the output to be changed. Stay tuned.