Machine Learning-Driven Optimization of Deuterium Water Upgrading Process Using Multi-Tower Cascade Distillation

Deuterium water(D₂O) is an essential chemical material widely used in the nuclear energy field, particularly as a neutron moderator in nuclear reactors, and it also plays an important role in medical, biological, and chemical research. D₂O concentration by a water distillation cascade process, however, due to the extremely small vapor pressure difference between D₂O and H₂O, very high reflux ratios(typically ranging from 50 to 200) are required. This results in substantial energy consumption and elevated operating costs. In addition, traditional cascade experiments are expensive and difficult to conduct, and the complex coupling among multiple operating parameters makes it challenging to comprehensively observe isotope distribution and identify optimal operating conditions. Consequently, more effective control of key operating parameters is of great significance for optimizing production design and achieving cost control. In this study, a novel two-column cascade water distillation process coupled with a heat pump is developed to reduce energy consumption. To address the complex nonlinearity and multi-variable coupling characteristics of this process, an optimization and prediction model based on a back propagation(BP) artificial neural network was proposed. The model architecture was optimized with 7 hidden neurons and trained using 2085 sets of rigorous simulation data, in which the number of theoretical stages, compression ratio, and circulating working-fluid flow rate are selected as input variables, enabling accurate predictions of heavy water purity, energy-saving efficiency, and total cost. The results demonstrate the model’s exceptional reliability. The predicted values for all output variables showed high consistency with the experimental data, with the correlation coefficients(R) exceeding 0.999 and the mean squared error(MSE) being minimized. Furthermore, a sensitivity analysis was conducted using Garson’s and Yoon’s algorithms to quantify the relative importance of input variables. The analysis reveals that the compression ratio is the dominant factor influencing energy-saving efficiency, whereas the reflux ratio(correlated with circulation flow) significantly impacts product purity and total cost. Based on these findings, an optimized operational strategy is proposed: a higher compression ratio should be prioritized to minimize energy consumption, while a higher reflux ratio is recommended to balance equipment investment and operational expenses. Overall, this study demonstrates that machine learning techniques can be used as powerful optimization tools for complex multivariable separation processes. The proposed framework reduces the reliance on extensive experimental trials, simplifies the exploration of operating conditions, and provides an efficient strategy for process design and optimization in deuterium water distillation systems.