Refroidissement élastocalorique optimisé en temps réel par IA : amélioration de l’efficacité et de la durabilité dans les systèmes à compression à base d’alliage Ni-Ti.

Elastocaloric cooling based on stress-induced phase transformations in shape memory alloys (SMAs) offers a promising solid-state alternative to vapor-compression refrigeration. In this study, a laboratory-scale compression-mode elastocaloric cooling system utilizing Ni-Ti SMA tubes was developed and dynamically optimized through real-time artificial intelligence (AI) control. Baseline testing under fixed operational parameters (4.5 % compressive strain, 0.2 Hz cycle frequency, 0.6 L/min HTF flowrate) demonstrated a net cooling capacity of 520–550 W and a coefficient of performance (COP) of 2.8–3.1, with cold-side outlet temperatures dropping by 7–9 ◦C. A genetic algorithm (GA) search identified optimal operational regions, improving steady-state COP to 3.6–3.7 and cooling capacities to approximately 600–625 W. Building upon these findings, a reinforcement learning (RL) agent was deployed for real-time cycle-by-cycle optimization, dynamically adjusting strain amplitudes, cycle timing, and HTF flowrates. Under AI supervision, the system achieved a stabilized COP of 3.8–3.9 and cooling capacities of 640–660 W, while demonstrating robust adaptability to step changes in external thermal loads with minimal transient performance penalties. Long-term durability tests over 104 cycles uncovered only a ~5 % decline in cooling performance, linked to adaptive strain management strategies that mitigated SMA fatigue progression. Compared to conventional fixed-parameter operation, the AI-enhanced system showed a 20–30 % improvement in efficiency and extended functional lifetime. These results demonstrate that integrating real-time AI control into elastocaloric systems can noteworthily enhance both cooling performance and material durability, providing a critical step toward scalable, sustainable solid-state cooling technologies.