Optimization of Broadcast Beams in Massive MIMO: Learning from A Digital Twin