What should we do if 2 out of 10 CFD models do not converge? We need to compare CFD results amongst all 10 cases but 2 of them not converged. If we change under-relaxation factors for 2 of these, will our comparison be valid? What do people normally do when encounter this problem?

Your final result should not depend on relaxation factor. For rectifying convergence problem, have look on the poorley formed / diverged solution, and try to find regions of unrealistic numerical values of physical parameters, for your case of application. That can give you an idea for reason of divergence.