Neural Network Analysis of Zr and U Diffusion in LiCl-KCl Molten Salt Using Cathodic Cyclic Voltammetry Data

This study developed three types of neural networks—a multilayer perceptron (MLP), a gated recurrent unit (GRU), and a long short-term memory (LSTM) network—to predict cathodic sweep currents from cyclic voltammetry (CV) experiments. The training data included measurements from ZrCl4 and UCl3 dissolved in LiCl-KCl molten salt. Zirconium chloride data covered a temperature range of 723 to 823 K and concentrations from 0.5 to 5 wt%. Uranium chloride data were collected at 773 K and high concentrations, from 5 to 10 wt%.To test model robustness, subsets of the data were created by randomly omitting data in 10% steps, reducing the data set incrementally from its full size down to 10%. The networks’ predictions of the cathodic current responses were then used together with the Delahay and Randles-Sevcik equations to estimate diffusion coefficients for Zr(II) and a figure of merit for U(III).The results showed that the best model architecture depended on the data set. For zirconium, the MLP delivered the most consistent performance; for uranium, the GRU was superior, handling its more complex data more effectively. Crucially, the study found that accurately predicting diffusion coefficients hinged on preserving detailed data near the cathodic peak rather than simply using a larger data set.Although the prediction quality declined at low analyte concentrations, all the models correctly estimated the Zr(II) diffusion coefficient at the highest tested ZrCl4 concentration. These findings support using neural networks to simulate cathode sweep currents and extract reliable diffusion coefficients from CV data.