Mini-channel heat exchangers (MCHEs) are attractive for supercritical CO₂ (SCO₂) Brayton cycles due to their high thermal efficiency, compactness, and high-pressure resistance. Conventional fabrication methods, such as chemical etching and diffusion bonding, limit improvements in structural compactness and performance, while additive manufacturing (AM) enables complex freeform structures but suffers from significant surface roughness and dimensional deviations. This study addresses these challenges by integrating selective laser melting (SLM) orthogonal experiments, deep neural networks, and reinforcement learning (RL) optimization. Square and rectangular channel specimens were fabricated under different process parameters, with channel dimensions and surface roughness measured via metallographic and confocal microscopy. A U-Net-based image recognition method extracted geometric features for dimensional evaluation, while a backpropagation neural network (BPNN) predicted the relationship between process parameters and fabrication quality. An RL agent based on the Deep Deterministic Policy Gradient (DDPG) algorithm identified optimal parameters, enabling high-quality MCHE fabrication. Full-temperature and full-pressure SCO₂ tests confirmed that the optimized MCHE achieved dimensional and surface roughness deviations within ±5%, heat transfer correlation accuracy within ±5%, and a maximum volumetric power density of 300 MW/m³.
 

Selective laser melting, Mini-channel heat exchanger, Deep neural network, Reinforcement learning