There are many historic HPC applications in various scientific and technological fields, and they often belong to different development communities or target issues at different detail levels. Integrating/unifying multiple legacy HPC applications can sometimes be very costly. Also, in the fusion of HPC and AI, it is important to utilize the simulation data of legacy HPC applications for machine learning and improve the performance of HPC applications by machine learning methods. To solve these problems, we are developing a library for "coupling." The coupling library h3-Open-UTIL/MP realizes coupled calculations by connecting HPC applications that have been individually used for simulation and analysis. Starting from combining closely related physics simulations, such as atmospheric and ocean simulations, we have extended h3-Open-UTIL/MP. One is ensemble coupling, and it is now possible to proceed with coupled simulation while both or one of the components performs ensemble calculations. The other is multi-approach coupling. We will introduce our study to gradually replace the physical model with the data-driven model, combining physics simulation written in Fortran and learning in a modern machine learning library written in Python online using h3-Open-UTIL/MP.