>It does get close to what I get with the full DiffEq solution but I will have to keep on experimenting with it.

This approach is most promising to come up with some approximation of LHS and RHS and then do sum of abs error (do not necessarily do squares).

You could use SVR's internal computational model on normed vector spaces and modify what passed to global optimizer and make many variations of SVR which deal with regressor which is actually fitting differential equations vs. polynomials.

This is within our reach today! (I know I am going to regret these words by the end of sept)