Learning the laws of lithium-ion electrolyte transport using symbolic regression

Abstract

High-throughput experiments (HTE) enable fast exploration of advanced battery electrolytes over vast compositional spaces. Among the multiple properties considered for optimal electrolyte performance, the conductivity is critical. An analytical expression for ionic transport in electrolytes, accurate for practical compositions and operating conditions, would accelerate the process of i) co-optimizing conductivity alongside other desirable electrolyte properties, and ii) learning fundamental physical laws from data, which is one of the paramount goals of scientific big-data analytics. Here, we used symbolic regression with an HTE-acquired dataset of electrolyte conductivity and discovered a simple, accurate, consistent and generalizable surrogate expression. Notably, despite emerging from a purely statistical approach, the expression reflects functional aspects from established thermodynamic limiting laws, indicating our model is grounded on the fundamental physical mechanisms underpinning ionic transport. We prove the potential of using machine learning with HTE to find accurate and physically-sound models in complex systems without established physico-chemical theories.

Robotic cell assembly to accelerate battery research

“…The overall workflow of the herein presented automatic coin cell assembly robot focuses on the production process after electrode coating and electrolyte formulation. Though the system is principally amendable to manufacturing cells with different electrolyte mixtures 19,20 and electrodes, the herein presented study uses the same electrolyte (1M LiPF6 in 3:7 EC:EMC by weight formulated by Elyte, Germany) and electrodes throughout.…”

One‐Shot Active Learning for Globally Optimal Battery Electrolyte Conductivity**

“…The studies by Dave et al [11,12] consider a wide range of electrolyte formulations but within a narrow range of temperatures. Utilizing an existing dataset [13,14] spanning a wide range of formulations and temperatures, we aim to perform as few as possible additional experiments to discover formulations with maximum conductivity for a wide range of temperatures. This is performed in a workflow called one-shot active learning.…”

“…The existing dataset of lithium hexafluorophosphate (LiPF 6 ) in ethylene carbonate (EC), ethyl methyl carbonate (EMC) and propylene carbonate (PC) was totaling 80 electrolyte formulations with measured conductivities at À 30 to 60 °C. [13,14] The suggestion of new formulations was fully exploitative, [20] i. e., requested formulations were selected solely based on their predicted conductivity at a respective temperature with complete neglect of model uncertainty. Active learning in fully exploitative mode has been shown to significantly increase the so-called "enhancement factor" by Rohr et al [20] The enhancement factor describes the increase in probability of finding an optimum given a fixed budget of experiments.…”

“…There are other research modes [20] not explored in this study. However, a recent study by Flores et al [13] focused on the "understanding driven" research mode. Their symbolic regression approach [13] works well for high temperatures but fails for highly concentrated liquid electrolytes at low temperatures, indicating a change in the physicochemical behavior.…”

“…However, a recent study by Flores et al [13] focused on the "understanding driven" research mode. Their symbolic regression approach [13] works well for high temperatures but fails for highly concentrated liquid electrolytes at low temperatures, indicating a change in the physicochemical behavior. Organizationally, this study is the human-in-the-loop version of the fully autonomous active learning study presented by Rahmanian et al [21]…”