(195b) Machine Learning the Fundamental Tradeoffs between Conductivity and Voltage Stability in Solid State Electrolytes
AIChE Annual Meeting
2019
2019 AIChE Annual Meeting
Topical Conference: Applications of Data Science to Molecules and Materials
Applications of Data Science in Molecular Sciences I
Monday, November 11, 2019 - 3:45pm to 4:00pm
All solid state batteries provide
many safety advantages over traditional lithium-ion batteries by replacing the
combustible organic liquid electrolyte with a ceramic solid-state electrolyte.
However, reported superionic conductors with conductivities approaching that of
liquid electrolytes are unstable in contact with a lithium anode leading to
increased internal cell resistance, and poor cyclability. Conversely, compounds
stable at the anode or cathode interfaces do not exhibit useful bulk ionic
conductivities. Although ab initio methods exist to study each ionic
conductivity and voltage stability range, there is no established theory to
connect these two properties. Here, we leverage machine learning (ML) to
investigate the role of crystal structure in the tradeoff between voltage stability
and ionic conductivity. To this end, we trained a partial least squares (PLS)
machine learning algorithm using the valence electronic density as a descriptor
of 60 known solid-state electrolytes along with their corresponding ionic conductivity,
anodic voltage limit, and cathodic voltage limit. The trained model has an 80%
prediction accuracy and suggests that within the search space of crystal
structures, the voltage stability and ionic conductivity are inherently
inversely correlated (Figure 1a). A multi-objective optimization also suggests
that materials with positively correlated ionic conductivity and voltage
stability may be highly anisotropic. Our PLS machine learning model, compared
to the more conventional neural network, has the benefit of being able to
predict and explore the relationship between multiple properties and retains a
high level of interpretability versus other âblack boxâ models. The PLS model
successfully identifies BCC anion substructure (Figure 1b) and channels as
effective descriptors, which is in good agreement with prior work.1
Using this model, we screened through a database of ca. 14,000 materials and
identified five new promising solid state electrolyte candidates to have
conductivities greater than 16 mS/cm. The model predictions were subsequently verified
with ab initio molecular dynamics simulations. The proposed ML model and
electron density descriptor may be used in future studies to elucidate structure-property
relationships for other applications with high accuracy and without sacrificing
interpretability.
Figure 1: (a) Latent
variable space determined by the PLS model used to predict properties and (b)
the BCC anion substructure identified by the PLS model.
References:
1. Wang, Y. et al. Design principlesfor solid-state lithium superionic conductors. Nat. Mater. 14,
1026â1031 (2015).