First, clearly a hard model as you are maximizing a convex function
Various options
You model it using binary variables and implies, although that will lead to a nasty nonlinear mixed-integer program.
You approcimate it using a pwa function and sos2 operators leading to a MILP
You model it using an anonymous black-box function using sdpfun (no derivatives will be available then)
You define your own overloaded operator and define derivatives etc (i.e,. copy logic in, e.g., sdpvar/atan)