Designing new solid-state electrolytes for batteries through AiiDA workflows

A substantial improvement in performance and safety of electrochemical storage systems is mandatory in order to reduce the amount of greenhouse gases emitted by our vehicles. This can only be achieved by finding new materials to be used as electrolytes for lithium ions and protons. Solid inorganic Li-ion conductors are ultimately the only solution that can ensure safety of Li-based cells – no liquid organic electrolytes can withstand a runaway thermal event. Moreover, Li-metal anodes are key to high-energy batteries (Li-ion, Li-S, Li-air), but solid-electrolyte interphase layers formed in-situ in today’s organic electrolytes do not protect the Li anode against morphology changes during cycling. A protective solid membrane is required that is flexible, strong, and has high conductivity and chemical stability. Currently there is no material that satisfies these requirements.

In order to optimize new classes of solid organic Li-ion electrolytes with high ionic and low electronic conductivity, and good electrochemical stability, the MaX EPFL team has developed a new protocol where structures are automatically pulled from crystallographic databases, translated into the appropriate self-consistent first-principles calculation (including Car-Parrinello molecular dynamics MD), and where the simulations are automatically monitored for optimal performance. The last step, that concerns the computational screening on ionic transport, electronic structure, and electrochemical properties is realized by implementing a set of ready-to-use workflows realized within the python-based Automated Interactive Interface for computational science (AiiDA) platform [1,2], that allows the management of massive amount of calculations.

The approach is based on the development of a new hybrid/ab-initio potential able to describe the physic of lithium diffusion in a vast range of materials. The developed model (“pinball” model) allows very fast simulations of lithium carrier diffusion by molecular dynamics. The potential is optimised through a procedure based on regression and force-matching, where regression parameters mimic the screening effect of charge polarization, allowing us to describe the evolution of forces with an accuracy comparable with more sophisticated and expensive DFT procedures (see Fig. 1). The implementation of a new algorithm based on the Delaunay Triangulation on crystal structures then leads us to extract conductivity pathways in materials from relative short molecular dynamics simulations.

Fig. 1. Forces on lithium in Li10GeP2S12 calculated with DFT against the forces calculated with the ‘pinball model’, yielding excellent agreement. The top left inset shows the error in the forces made (against DFT) for the proposed pinball model, to be compared with a previously reported level of approximation (bottom right).

In order to compute diffusion coefficients in an automated fashion we combine the potentialities offered by the AiiDA platform with the MD tools implemented in the DFT plane-wave Quantum ESPRESSO package. Our aim is to maximize ease of use while providing robust and reusable workflows that can be shared. We have produced a workflow that eases the cumbersome task of launching and monitoring molecular dynamic simulations by automatizing the following steps:  to replicate the unit cell to limit spurious interactions, to perform calculations at the Gamma point, to check if a structure is computable with the resources at hand and, if compatible, to thermalize the system in a canonical simulation using ab-initio MD on the Born-Oppenheimer surface as implemented in the Quantum ESPRESSO code. This step is subdivided in multiple short simulations executed iteratively. This loop interrupts when the system is considered thermalized. The thermalized structure is afterwards simulated in a micro canonical ensemble to eliminate any effect of the thermostat. Aggregated micro canonical trajectories are then analyzed and transport coefficients are calculated (a diagrammatic representation of the MD workflow is reported in Fig. 2).

Fig. 2. A diagrammatic representation of the MD workflow.

The resulting computational high-throughput screening approach was used in a first screening of a subset of 100 structures from the ICSD and COD structural databases. This small set already resulted in one material that has a similar diffusion coefficient to the best material known (Li10GeP2S12), leading us to believe that our screening approach will yield more viable candidates. After the analysis, we will move on to screening a larger set of materials (~10,000 structures).

The workflows and tools are publicly available on BitBucket under MIT license.

[1]  Pizzi, G., Cepellotti, A., Sabatini, R., Marzari, N. & Kozinsky, B. Aiida: automated interactive infrastructure and database for computational science. Comp Mat Sci 111, 218–230 (2016).

[2]  AiiDA platform,