# High‐Conductivity Electrolytes Screened Using Fragment‐ and Composition‐Aware Deep Learning

## TL;DR

A new deep learning framework predicts electrolyte conductivity in batteries by combining molecular and functional unit information, enabling faster and more interpretable design.

## Contribution

A hierarchical deep learning framework that integrates molecular and functional unit attributions for accurate and interpretable electrolyte conductivity prediction.

## Key findings

- The framework achieves high accuracy in predicting ionic conductivity of lithium battery electrolytes.
- Fragment-level and composition-aware attentions provide chemically interpretable insights into electrolyte performance.
- The model enables large-scale virtual screening of electrolyte formulations.

## Abstract

Rising energy generation from renewables (e.g., wind, solar power) will drive global demand for >1.0 TWh of long‐duration energy storage by 2030 to stabilise grids and balance supply. Rechargeable batteries are central to this transition, with their performance critically governed by the properties of active materials and supporting electrolytes. However, designing electrolyte formulations remains a major challenge, as their performance arises from complex, non‐additive interactions among lithium salts and organic solvents, requiring elegant molecular design and selection. Conventional trial‐and‐error strategies still dominate electrolyte design, but they are slow and resource‐intensive. Recent machine learning approaches have improved electrolyte screening, yet many rely on coarse molecular representations that neglect fragment‐level chemistry and explicit ratios, limiting interpretability and their utility for guiding experiments. Here we introduce a deep learning framework that integrates intermolecular attributions across solvents with intramolecular attributions from functional units. The framework builds a hierarchical representation, decomposing formulations into molecules and their functional units, while integrating ratios, physicochemical descriptors, and salt identity to generate mixture‐invariant embeddings for accurate and interpretable conductivity prediction. Applied to benchmark datasets of lithium battery electrolytes, the framework achieves high accuracy in predicting ionic conductivity and enables large‐scale virtual screening. Crucially, it provides chemically interpretable insights: fragment‐level attentions align with functional units; composition‐aware attention reveals the impact of mixing ratios; and counterfactual perturbations confirm causal roles of key motifs. This framework paves the way for data‐driven, interpretable electrolyte design and can be generalized to broader formulation challenges in materials science.

We present a new deep learning framework that hierarchically links molecular and functional unit attributions to predict electrolyte conductivity. By integrating molecular composition, ratios, and physicochemical descriptors, it achieves accurate, interpretable predictions and large‐scale virtual screening, offering chemically meaningful insights for data‐driven electrolyte and formulation design across materials science.