The global shift toward low-carbon energy underscores an urgent need for efficient heating in cold regions. CO₂ heat pumps, leveraging eco-friendly refrigerants and excellent low-ambient performance, show promise but suffer from significant throttling losses due to CO₂’s low critical temperature, limiting cycle efficiency. Existing optimizations (e.g., recuperators, two-stage cycles) rely on expert-guided modifications and fail to holistically integrate refrigerant selection, cycle configuration, and heat exchange networks. To overcome this, we propose the intelligent GraPHsep method, which: (1) generates compressor/throttling valve topologies via graph theory; (2) constructs refrigerant flow structures using splits, junctions, and phase separators; (3) synthesizes heat exchangers by coupling refrigerant branches with heat sources/sinks and optimizing layouts via pinch analysis combined with a genetic algorithm. Applied to space heating at −30 ℃ ambient/45 ℃ supply water, GraPHsep designed a novel CO2 cycle achieving a coefficient of performance of 2.17, which is 77.9% and 16.7% higher than classical single/two-stage transcritical cycles, respectively. This demonstrates its capability to develop high-efficiency heat pump systems for extreme cold climates.