Next generation batteries will require new electrolyte concepts. Two interesting alternatives are ionic liquids (ILs) and high entropy electrolytes, molten salts (MSE). However, the chemical space of these materials are vast and it is impossible to fully explore them using classical methods. In this project, we will model ILs and MSE on a molecular level using machine learning (ML) and symbolic regression (SR). A particular aim is to relate the molecular structures to macroscopic properties that are relevant for batteries, such as ionic conductivity. Previous works and initial experiments have shown that graph neural networks (NNs) can be successful in these tasks - but to take the work further more computational resources will be needed. During the previous round of this project, initial experiments with classical ML methods and NNs served to verify previous works and helped developing a good data pipeline. Unfortunately we then turned our focus to SR and an implementation that uses genetic algorithms, something that required us to make use of CPU- rather than GPU-resources. However, this project is now very close to an end (with an almost finished manuscript) and in future works we would like to explore alternative ways to implement SR. Some of the more interesting approaches are much more reliant on NNs and thus GPU-resources, something we could use this project for. One of the areas that we would like to explore is the use of variational autoencoders (VAE) as a means of doing symbolic regression. Here you would encode equations into a latent space, which then could be used to sample equations in a clever way. This could potentially have several advantages, including computational speed and the possibility to make use of a Bayesian approach, perhaps posing a prior over the desired equations we are searching for.