Abstract: To address the problems of repeated iterations and long computational time of existing optimization methods for supercritical carbon dioxide (sCO₂) mini-channel heat exchangers under different design conditions, a reinforcement learning-based method for thermal and strength optimization of mini-channel heat exchangers under multiple design conditions is proposed. A coupled environment integrating variable-property segmented thermal design and strength verification is established. A reward function is developed with volumetric power density as the primary objective, considering compactness and pressure drop while treating structural strength as a hard constraint. A design optimization agent based on the deep deterministic policy gradient algorithm is constructed. The results show that, with 50 environment-model evaluations, the agent improves the composite reward by 13.7% and reduces the search time by 65.2% compared with the genetic algorithm (GA) under the same computational budget. When achieving a composite reward comparable to that of the conventional constrained GA, the number of environment-model evaluations and search time required by the agent are only 2.4% and 0.9% of those of the GA, respectively. For 300 design conditions, the average total optimization time of five repeated tests is 56.9 s. All optimized designs satisfy the pressure-drop and strength constraints, with volumetric power density of 48.23–127.61 MW/m³ and average compactness of 5449.1 m²/m³.