Uni-ELF: A Multi-Level Representation Learning Framework for Electrolyte Formulation Design

TL;DR

Uni-ELF addresses the challenge of designing electrolyte formulations by introducing a two-stage pretraining framework that unifies molecular- and formulation-level representations. It combines molecule-level 3D structure reconstruction using Uni-Mol with formulation-level RDF prediction derived from MD simulations, employing molar-ratio weighting and temperature embeddings in a transformer backbone. The approach yields state-of-the-art performance on both molecular properties (e.g., melting point, boiling point, dielectric constant, refractive index, density, synthesizability) and formulation properties (e.g., Coulombic efficiency, conductivity), and demonstrates robust few-shot generalization and the ability to rediscover high-performance solvents like FAN. This work paves the way for AI-assisted automated electrolyte design and has potential applicability to other formulation-heavy domains, enabling closed-loop integration with experimental validation.

Abstract

Advancements in lithium battery technology heavily rely on the design and engineering of electrolytes. However, current schemes for molecular design and recipe optimization of electrolytes lack an effective computational-experimental closed loop and often fall short in accurately predicting diverse electrolyte formulation properties. In this work, we introduce Uni-ELF, a novel multi-level representation learning framework to advance electrolyte design. Our approach involves two-stage pretraining: reconstructing three-dimensional molecular structures at the molecular level using the Uni-Mol model, and predicting statistical structural properties (e.g., radial distribution functions) from molecular dynamics simulations at the mixture level. Through this comprehensive pretraining, Uni-ELF is able to capture intricate molecular and mixture-level information, which significantly enhances its predictive capability. As a result, Uni-ELF substantially outperforms state-of-the-art methods in predicting both molecular properties (e.g., melting point, boiling point, synthesizability) and formulation properties (e.g., conductivity, Coulombic efficiency). Moreover, Uni-ELF can be seamlessly integrated into an automatic experimental design workflow. We believe this innovative framework will pave the way for automated AI-based electrolyte design and engineering.

Figures (5)

• Figure 1: Electrolyte formulation representation learning framework.a, Electrolyte design at multiple levels. At the atomic level, individual atoms and their interactions form molecular geometric structures, creating molecular-level representations. Based on these, individual molecular species, their proportions, and their interactions (depicted by red lines) within the mixtures create formulation-level representations, which are then used to predict device-level properties. b, Multi-level representation learning: b1. Molecule-level representations are learned through self-supervised tasks, including recovering masked atom types and denoising atom pair distances. b2. These refined representations are then fed with mixture ratios into the Uni-ELF backbone. c, Uni-ELF backbone model architecture. The Uni-ELF model is based on a transformer encoder design. Molar ratios are used as weights for molecular representations, and pair representations are maintained for mixture-level pretraining. Symmetrical elements in the pair representation matrix are summed and combined with the radial features obtained from the Gaussian kernel. These combined features are then used to predict radial distribution functions (RDFs), a pretraining task to recover the structural properties of the mixed system.
• Figure 2: Prediction of molecular pairwise RDFs as a formulation-level pretraining task, using the LiPF$_6$/PC/EMC system with a molar ratio of n(Li$^+$) : n(PF$_6^-$) : n(PC) : n(EMC) = 0.12 : 0.12 : 0.54 : 0.22 as an example. The plots compare the true values obtained from molecular dynamics (MD) simulations (blue) with the predicted values from the Uni-ELF model (orange) for various molecular pairs: PF$_6^-$, Li$^+$, PC, and EMC, including all pairwise combinations forming a lower triangular matrix. The right panel illustrates the system configuration. The strong agreement between predicted and true RDFs demonstrates the accuracy of the Uni-ELF model during pretraining.
• Figure 3: Comparative performance in predicting molecular properties for electrolyte design. Uni-ELF (in purple) surpasses previously reported state-of-the-art (SOTA) methods (in blue) in predicting seven molecular properties (melting point, boiling point, vapor pressure, dielectric constant, refractive index, density on R² scores, and synthesizability on the AUC), which are essential for the inverse molecular design of electrolytes. Each concentric circle represents an interval of 0.05, with the outermost boundary corresponding to a perfect score of 1.0.
• Figure 4: Regression plots for electrolyte formulation property prediction using Uni-ELF.(a) Results of the Coulombic efficiency dataset. (b,c) Liquid electrolyte conductivity dataset, with (b) representing the random split and (c) the group split. The regression plots show the parity between experimental and predicted values in the test sets, with insets showing the results in the training sets. To illustrate data distribution, kernel density estimation is displayed at the top and right of each plot. The color gradients in the plots indicate the magnitude of prediction errors.
• Figure 5: Conceptual electrolyte design using Uni-ELF.a, Set-up of the conceptual experiment: The objective is to achieve high ionic conductivity, a wide liquid range, and ease of synthesis. The molecular space to search is constrained by some practical expert criteria. b, 1,165 candidates generated by graph-theoretic enumeration, visualized using t-SNEvan2008visualizing to reduce molecular representations to two dimensions, and color-coded by predicted maximum conductivity and synthesizability. The red circle highlights the high-conductivity FAN (fluoroacetonitrile) molecule discovered by the model, while the blue circle highlights a series of four-membered ring molecules with high predicted conductivity but low predicted synthesizability, which were thus screened out. c, Top 10 molecules from zero-shot formulation-level prediction, emphasizing FAN's superior performance. Positive and negative values indicate model-predicted standard deviations, with parentheses showing experimental values. d, Few-shot learning: Conductivity vs. concentration and temperature for LiFSI/FAN and LiTFSI/FAN systems. The model accurately predicts the conductivity-concentration relationship using data from only three experimental points: the initial concentration point (0.1 M), the final concentration point (4 M), and the peak conductivity concentration point (1.3 M for LiFSI/FAN and 1.2 M for LiTFSI/FAN) predicted by the model, which notably aligns with the experimental results. For the conductivity-temperature relationship, the model accurately predicts the high conductivity performance of FAN at low temperatures, fitting well to the Arrhenius relationship (red text).