Deep Learning-Assisted Elucidation of Structure-Performance Relationships in MOF-NH3 Adsorption Refrigeration Working Pairs.

Adsorption refrigeration systems using MOFs-ammonia working pair can convert low-grade thermal energy into cooling capacity, thereby emerging as promising future carbon-neutral energy systems. However, the adsorption mechanism of MOFs under saturated ammonia conditions remains unclear due to the lack of sufficient samples and dedicated databases. To mitigate this limitation, this data-driven study established a MOFs-NH3 adsorption database containing 9835 MOFs structures via high-throughput grand canonical Monte Carlo calculations. Further, machine learning and deep learning modeling were conducted, with 42-dimensional MOFs parameters serving as feature variables and the cyclic adsorption capacity as the target variable. Among all the developed models, the Convolutional Neural Network (CNN) model yielded the highest prediction accuracy (R2 = 0.880 ± 0.030), as its local receptive field generates MOFs feature fingerprints to capture the deep characteristics of the structure-performance relationships. Subsequently, via the dependence analysis of SHAP values, the MOFs parameters exerting the most significant impacts on adsorption refrigeration performance were ranked as volume-related parameters, surface area-related parameters, and pore size-related parameters. Notably, among the volume-related parameters, the specific pore volume exhibited the best predictive performance, as it captures both the steric effect and adsorption potential energy of MOFs. Furthermore, the optimal ranges of each parameter that contributes most to refrigeration performance were quantified. This study deepens the understanding of MOFs-ammonia adsorption mechanisms and provides guidance for the design of next-generation carbon-neutral materials in the field of adsorption refrigeration.