Molecular Dynamics simulations can be very useful in optimizing battery electrolytes, being cheaper than costly experiments as well as providing a more atomistic understanding of the electrolyte system. Unfortunately, the calculation of the ionic conductivity and transference numbers, which are important for electrolyte screening, require longer simulations for sufficient statistics.

This project aims to circumvent that problem by using machine learning techniques to predict the dynamic properties, like the ionic conductivity, from structural properties, which can be derived from shorter simulations. Additionally, analysis of which features the machine learning models deem important for their prediction of dynamic properties can give insight into the connection between structure and dynamics.

Ionic liquids were chosen as simple model systems to explore machine learning predictions. Since ionic liquids only consist of anions and cations, it is straightforward to construct a database of simulated systems by freely combining anions and cations. For a description of the structure that can be read by machine learning models, the radial distribution function (RDF) was chosen due to its physical interpretability. As for machine learning models, a linear regression, a random forest and a neural network were utilized to predict the ionic conductivity.

After the initial results, the quality of prediction was improved by feature engineering and by refining the input data. Using the logarithm of the RDF, which corresponds to an effective free energy landscape, as well as the heights and positions of the peaks of the RDF, emphasizing coordination shells, delivered the most promising results. Especially the results using the peaks as input were very helpful in mapping structures to dynamics, implying that the most important information is derived from the first peak, i.e., the first coordination sphere.

Overall, this project showed the importance of having some physical understanding in the data preparation. Because the random forest model and linear regression model outperform the neural networks, our analysis also shows the importance of considering simpler models. While these models were trained on ionic liquids, they could also be adapted to different electrolyte systems or even help in the discovery of new electrolytes by providing estimates of their ionic conductivity based on short simulations.