Air route networks can no longer meet operational efficiency requirements because of the rapid growth of complex traffic flows. Machine learning is employed to investigate the evolutionary mechanism of congestion in such networks in view of their high complexity and high density, and a reasonable network optimisation scheme is presented. First, deviations between nominal and actual routes are investigated with reference to radar track data, and a network reflecting actual route operations is constructed using adversarial neural networks. Second, flight time is used to characterise congestion in route networks. Actual network operations are considered, and congestion is defined from the perspective of road traffic engineering. The effects of the operational properties of traffic flows on flight times are analysed to establish various congestion indicators. A gradient boosting model is used to select indicator characteristics and analyse patterns in the variations of indicator values for each flight segment in distinct periods. The indicator–time relationship is leveraged to explore the evolutionary mechanism of congestion in the route network. Third, on the basis of this mechanism, a multiobjective optimisation model of congestion is formulated, and a particle swarm optimisation algorithm is executed to adjust the route passage structure, thereby solving the optimisation model. Finally, calculation validation is conducted using radar track data from the control sector of the Yunnan region. The average flight time in a route segment is 10% shorter in the optimised route network than in the nonoptimised route network, which confirms that the optimisation solution is practicable.