The CO2 power cycle (CPC) system is an efficient and environmentally friendly method for waste heat recovery (WHR). However, the traditional design and optimization process of a CPC system is very complex and time-consuming. This paper proposes a novel goal-oriented design method based on machine-learning methods for quickly designing an optimized CPC system with given performance indicators. And taking the design of the CO2 transcritical power cycle (CTPC) system for internal combustion engines (ICEs) as an example. Firstly, the net output power and the total cost of the system prediction models are trained by simulated data. Then the multi-objective optimization of the system is carried out by using the genetic algorithm coupled with the prediction models, and the optimization results are used to train a classification model. Finally, the given target indicators are input into the classification model for goal-oriented designing and getting the optimal configuration. The results of the goal-oriented design validation show that the goal-oriented design method can design the CTPC system well. And, once the classification model is trained, the CTPC system's future goal-oriented design process only needs to be calculated once, significantly reducing design time. In conclusion, the goal-oriented design method based on machine-learning proposed is a novel and promising method. This is a technology that combines computer science and energy science and can provide users with a quick and reliable CPC system design method.