(189cm) Using Artificial Neural Networks to Model Diffusion Characteristics in Lithium Solid State Electrolytes

There is great interest is
developing solid state lithium electrolytes for use in an all solid-state
battery to replace the flammable organic electrolyte. Previous efforts trying to
understand the structure-function relationships resulting in high ionic
conductivity materials have relied on ab-initio molecular dynamics (AIMD).
Such simulations, however, are computationally demanding and cannot be applied
to large systems containing more than a few hundred atoms in a reasonable time
frame. Herein, we propose using machine learning artificial neural networks
(ANN) to supply the forces and energies used during the MD simulations, and to
eliminate the need of costly ab-initio force and energy evaluation methods,
such as density functional theory (DFT). After carefully training a robust
artificial neural network for four and five element systems, we obtain nearly
identical lithium ion diffusivities for Li₁₀GeP₂S₁₂
(LGPS) when benchmarking the ANN-MD results with DFT-MD. We find that ANN-MD
simulations allow the study of systems that require high number of atoms
because the calculation of forces and energies scales linearly with the number
of atoms rather than cubically in DFT (Figure 1), To demonstrate the power of
the outlined ANN-MD approach we apply it to a chlorine doped LGPS system to
calculate the effect of concentrations of chlorine on the lithium diffusivity
at a resolution in Li₂₀Si₃P₃S₂₃Cl
with decreasing concentrations of chlorine: 2% 1% and 0.67%. Such low
concentrations would be unrealistic to model with DFT-MD. We found that an
optimal 1% chlorine substitution resulted in the highest predicted room
temperature conductivity of 24.22 mS cm^-1 which is in remarkable
agreement with the experimental value of 25mS cm^-1 (Figure 1).¹
We additionally test the overall robustness of these ANNs and find their
accuracy is highly dependent upon the training data with some force fields
being unable to reproduce 0 K structure optimization while others accurately
reproduce both 0 K and finite temperature effects. Overall, we find that ANN-MD
simulations can provide the framework to study systems that require a large
number of atoms more efficiently and with high accuracy within the trained
regime.

Figure 1: Plot of the time needed for one force
calculation as a function of number of atoms. Above 50 atoms the artificial
neural network is more efficient. In this region we calculate diffusion
coefficients for a doped LGPS system and find an optimal diffusion at 1%
Chlorine.

References:

1. Kato, Y. et al. High-power
all-solid-state batteries using sulfide superionic conductors. Nat. Energy
1, 16030 (2016).