I am aware about LunchBox, Pug and Owl. I will start by reading the documentation and read tutorials - yet it will be extremely helpful to see a solution to a problem similar to mine.
I have also posted this on Grasshopper forum. Thanks in advance!
I have worked on a similar workflow using Ladybug/Honeybee simulation outputs to create a multi-objective dataset. After collecting simulation results and design parameters, I used Python-based ML/data analysis workflows to investigate relationships between inputs and performance indicators. For your case, I would suggest treating Ladybug outputs as your training data and developing a surrogate model to replace expensive simulations.
I can send my LinkedIn link and see my work , if you would like.