The focus of our work is the development of computational tools to accelerate the design of organic electrolytes for aqueous organic redox flow batteries (AORFBs). In our paper, we present the construction of a computational dataset of organic redox-active molecules for aqueous electrolytes. This dataset underpins the development of machine learning (ML) models for property prediction, with a longer-term vision of leveraging Generative AI to propose novel electrolyte candidates and accelerate sustainable materials discovery.

The growing demand for renewable energy emphasises the need for efficient and scalable energy storage. AORFBs are promising systems due to their sustainability and tunability. However, identifying suitable organic electrolytes remains a major challenge. [1] Molecules must satisfy strict criteria including appropriate redox potentials, solubility and chemical stability. Current discovery approaches rely heavily on experimental screening, which is both time-consuming and resource intensive.

This study aims to advance electrolyte discovery for AORFBs by integrating high-throughput computational methods. Specific goals include: (i) identifying molecular properties critical for electrolyte performance, (ii) benchmarking computational approaches to guide design and (iii) applying ML to accelerate computational screening and propose novel candidates.

We implemented a virtual screening workflow to evaluate existing ML models for key electrolyte properties, such as redox potential and solubility. A custom RDKit-based enumeration tool was developed to generate combinatorial derivative libraries of organic backbones through user-defined substitutions, generating over 10,000 unique molecules. Several state-of-the-art ML models were benchmarked, including MolGAT for redox potential, AqSolPred for solubility, and GASA for synthetic accessibility. [2-4] The enumeration tool enabled flexible, high-throughput exploration of structural modifications, while the ML benchmarking revealed both the strengths and limitations of current predictive models for redox-active systems.

We employed the enumeration tool with an automated DFT workflow to model the redox reactions of 7,500 organic redox-active molecules, yielding a consistent dataset of redox potentials and solvation free energies across diverse organic scaffolds. We take this dataset to train a ML model to predict redox potentials with DFT-level accuracy at a significantly lower computational cost. Ongoing investigations explore the use of transfer learning strategies for cross-domain redox potential predictions using graph neural networks (GNNs).

Overall, this study establishes a computational framework for aqueous organic electrolyte discovery, integrating molecular enumeration, automated quantum chemical calculations and AI. The resulting dataset provides a robust foundation for developing advanced AI models to support sustainable materials discovery. Future work will couple these predictive models with generative AI approaches to design novel electrolyte candidates.