Molten carbonate salts are growing in popularity in applications such as heat transport and thermal energy storage, as well as molten carbonate fuel cells and rechargeable batteries, due to their high ionic conductivity, low viscosity, and reduced environmental impact. Experimentally, characterizing the electrolyte structure and properties is very difficult, and is often limited to the measurement of a single observable. While computational methods for predicting the physicochemical properties of molten salts have been recently employed, many fail to accurately capture relevant physical properties, such as diffusivity, ionic conductivity, and equilibrium composition. In addition, many computational approaches struggle to scale to larger systems, as the computational cost becomes intractable with increasing system size. To address these challenges, Anirban Mondal, Athanassios Z. Panagiotopoulos and colleagues developed a deep potential machine learning model with active learning to simulate the chemical reactions of alkali carbonate–hydroxide electrolytes containing dissolved CO2.
In this deep potential generator (DP-GEN) workflow, ab initio molecular dynamics (AIMD)-generated configurations were used to produce a training dataset. Four independent deep neural network (DNN) models were then trained with varying parameters, and the DNN models were improved by adding additional training data via iterative exploration of the configurational space with deep potential molecular dynamics (DPMD) simulations under varying thermodynamic conditions. For snapshots that failed to converge, the previous steps were repeated iteratively. This iterative process helped to improve the overall accuracy by increasing the number of data points. At the end of the DP-GEN process, the training set contained over 15,000 structures, covering temperatures in the range of 825–1,225 K.
This is a preview of subscription content, access via your institution