Transfer learning-based prediction and evaluation for ionic osmotic energy conversion under concentration and temperature gradients

Abstract

Ionic osmotic energy conversion under concentration and temperature gradients synergistically utilizes osmotic and thermal energies to drive the directional migration of ions in charged nanochannels for power generation. The current research conducts preliminary experiments and simulations to determine the impact of a single parameter on output performance while lacking prediction models to reflect the link between comprehensive parameters and outputs. The complex partial differential relationship restricts the establishment of prediction models, which can be addressed by combining engineering and data science like transfer learning. This study presents a data-driven insight into ionic osmotic energy conversion to establish a transfer learning-based prediction model for comprehensive parameters using small sample sizes. Based on the trained source task model, the transfer learning-based deep neural network (TL–DNN) model with 17 inputs and 3 outputs is trained by freezing four hidden layers with 600 samples acquired from finite element method (FEM) simulations. The determination coefficients of diffusion potential, maximum power, and energy conversion efficiency are predicted to be 0.97, 0.98, and 0.97, respectively, by the TL–DNN model based on 5-fold cross-validation. Compared with FEM results, the TL–DNN model displays an exceptionally high speedup ratio of 1.37 × 10⁶ with errors less than 4 %. Besides, low concentrations and nanochannel radius exhibit high descriptor importance exceeding 0.70, indicating the dominant influence on performance. The multi-objective optimization is performed by non-dominated sorting genetic algorithm II to obtain 10 sets of parameter combinations with the highest entropy weight scores. This study has provided an alternative prediction model based on transfer learning and promotes theoretical development by applying data science to engineering science.