Compular Simulator: Automated Computational Screening of Battery Electrolytes

Abstract

The vast compositional space and intricate ionic interactions of modern liquid battery electrolytes make the connected R&D very challenging. A comprehensive workflow for reliable computational screening of electrolytes would therefore be highly valuable, which motivated the development of CHAMPION [1-4] and laid the foundation for Compular [5].Within this context, we present Compular Simulator, a new web application automating advanced computational screening of electrolytes, based on molecular dynamics simulations, density functional theory calculations, and machine learning (ML) models. The Simulator can be used in three different modes with increasing degree of sophistication: 1) single system simulation, 2) manual screening, and 3) property optimization by a design of experiments (DoE) methodology.In the single system mode, the user specifies the stoichiometry and thermodynamic state. Simulator then automatically creates a simulation cell with a physically reasonable initial geometry, sets up and runs suitable simulations and CHAMPION analyses on our HPC cloud. Throughout, the user can follow the status of the computations. The results are presented on three levels: a) overall system, b) molecular species, and c) dynamic species. Examples of the former properties are total ionic conductivity, density, electrochemical stability window (ESW), rate of solvation dynamics, conformational entropy, degree of aggregation, and viscosity, while for molecular species, properties including concentration, diffusivity, partial ionic conductivity, ESW, solvation and coordination numbers, average lifetimes, and transference numbers. These are all computed and presented in tables and charts. Finally, dynamic species are structures formed by non-covalent bonds, e.g. solvation shells and ionic aggregates, with the same set of properties as molecular species and presented in the same manner. The tables and graphs can be sorted and filtered w.r.t. properties and constituents and all predictions are given with statistical error bars.Manual screening can be done either by repeatedly creating single systems with varying composition or state, or by creating several systems at once by sweeping a given parameter, such as temperature or a species concentration over a range. Once all systems have been simulated and analyzed as described above, the user can compare the properties of the systems in the screening view. Similarly to the single system view, the screening view consists of three levels: a) overall system properties, shown as a table of systems vs. properties and one chart per property, b) molecular properties, presented as one table per property with systems vs. species and corresponding graphs, and c) dynamic species, similarly displayed. Tables in the screening view can be sorted and filtered similarly to those in the single system view, but in addition, systems can be filtered based on whether they fulfill requirements on their properties, e.g. greater or less than some value, or whether they are on the Pareto front of the target properties.Finally, in the DoE mode, the user specifies: the candidate species, whether any components should be held constant (e.g. fixed solvent composition in a salt screening), if so the total concentration (range) of the candidate species, the state, and the target properties, as well as whether they should be maximized or minimized. The state variables can also optionally be varied within a range. Simulator suggests a batch of systems to run in order to learn as much as possible about the parameter space. Based on the results, it builds a model for how the target properties vary as functions of the input parameters. Simulator then suggests another batch of systems, to add further information where most needed and explore regions near the Pareto front. In addition to the computational predictions, experimental data of target properties can be uploaded and used in the optimization. After each batch, the user is presented with the set of Pareto optimal systems among those sampled, and an assessment of the quality of the current model.While the Compular Simulator web application has hitherto been optimized for liquid electrolytes, we plan to extend it in the future and would love to hear from anyone who has an idea of a different use case. In summary, this is our biggest contribution yet towards fulfilling Compular’s vision of making battery materials R&D digital first.