A thermodynamics-consistent machine learning approach for ammonia-water thermal cycles

The integration of physics with data-driven models has emerged as a promising approach. However, the application of such hybrid approaches to ammonia-water thermal systems remains underexplored. We addressed this gap by developing a framework that integrates physical constraints through hard constraint layers, soft loss penalties, and anomaly detection techniques. We validated this framework across three case studies including heat exchanger, absorption refrigeration, and Rankine cycle. To evaluate the performance, we introduced a novel Thermal Fitting Score (TFS) that combines the coefficient of determination, R2, and thermal inconsistency metrics. Our key contributions include (1) comprehensive exploratory data analysis for thermal cycle understanding, (2) thermal constraint formulation based on thermodynamic laws, and (3) constraint integration at architecture, training, and data levels. The constrained models achieve 100 % thermal law compliance with TFS improvements of 23 % and 63.5 % for heat exchanger and absorption refrigeration cases, respectively. This methodology advances the integration of thermal domain knowledge with data-driven approaches, ensuring both prediction accuracy and thermal consistency.