Just a quick project idea.
Using machine learning for solving differential equations might sound hard. I think I got a simple idea. Just use machine learning ( learning from historic data ) to time predict the future points after sampling the previous calculated solution points by a known algorithm. For instance rk.