Developing clean energy technologies requires designing materials with specific functional properties. Traditional trial-and-error methods are slow and expensive. Data-driven methods using machine learning (ML) offer potential acceleration, but are limited by the scarcity of large, high-fidelity experimental datasets. Redox flow batteries (RFBs), particularly non-aqueous RFBs (NRFBs), are promising for grid-scale energy storage due to their ability to separate energy storage and power generation components and their potential for higher energy density through tuning the solubility of redox-active organic molecules (ROMs). A major challenge is the low solubility of ROMs in organic solvents, hindering their performance. Accurate solubility determination is difficult, influenced by solute properties, solvent composition, time, and temperature. While "excess solvent" methods offer speed, "excess solute" methods, though slower, provide more accurate thermodynamic solubility measurements. High-throughput experimentation (HTE) platforms improve the efficiency of the "excess solute" method by automating sample handling and reducing analysis time. However, adapting these methods to non-aqueous systems, and particularly binary solvent mixtures, is challenging. This work utilizes a machine learning-guided HTE platform, employing active learning (specifically Bayesian optimization, BO) to efficiently generate solubility data for ROMs in organic solvents.

The literature highlights the challenges in materials discovery for clean energy technologies, emphasizing the time and cost associated with traditional methods. Data-driven approaches using machine learning are gaining traction, but face limitations due to a lack of extensive, high-quality experimental data. Redox flow batteries are presented as a promising energy storage technology, with a particular focus on the need for improved solubility of redox-active organic molecules in non-aqueous systems. The existing literature details the complexities of solubility measurements, comparing 'excess solvent' and 'excess solute' methods, highlighting the advantages and disadvantages of each. The need for high-throughput experimentation (HTE) combined with machine learning techniques to accelerate the process is also discussed.

This research uses 2,1,3-benzothiadiazole (BTZ) as a model ROM to investigate its solubility in a library of over 2000 organic solvents. A closed-loop workflow integrates a high-throughput experimentation (HTE) platform with Bayesian optimization (BO). The HTE platform automates sample preparation (powder and solvent dispensing, mixing, monitoring for saturation, and sampling for analysis) using a robotic system. Solubility measurements are performed via quantitative 1H-NMR spectroscopy using 1,4-dinitrobenzene as an internal standard. The BO algorithm uses a Gaussian Process Regression (GPR) surrogate model and an expected improvement (EI) acquisition function to guide the selection of solvent candidates for experimental evaluation. The GPR model is trained using physicochemical descriptors and quantum chemistry-derived features (molecular weight, topological polar surface area, solvation free energies, dipole moments, etc.). The BO algorithm iteratively suggests new solvents based on model predictions and uncertainties, aiming to maximize BTZ solubility. The workflow begins with an initial set of solubility measurements, which are used to train the GPR model. The model then predicts the solubility of other solvents in the library, and the algorithm selects the solvents with the highest expected improvement for experimental evaluation. This closed-loop process continues until the desired solubility is achieved or a predefined stopping criterion is met. Additionally, a high-throughput viscosity measurement workflow was also incorporated, integrating automated sampling with a high-throughput viscometer.

The integrated platform successfully identified 18 binary solvent systems exhibiting BTZ solubility exceeding 6.20 M, after evaluating only 218 candidates (less than 10% of the total library). The study demonstrated that binary solvent mixtures, particularly those containing 1,4-dioxane, significantly enhanced BTZ solubility compared to individual solvents. The Gaussian Process Regression (GPR) model achieved an R² of 0.81 for predicting BTZ solubility in the training dataset. Bayesian optimization (BO) proved significantly more efficient than random solvent selection in identifying high-solubility solvents. The synergistic effect of solvent mixing on BTZ solubility was highlighted, with some binary solvent combinations yielding unexpectedly high solubilities even when one component alone showed poor solubility. For example, the combination of 1,4-dioxane and glutaronitrile resulted in a surprisingly high solubility of 6.84 M, despite glutaronitrile having a low solubility of 1.86 M on its own. The viscosity of the saturated solutions remained below 2.5 cP, indicating that high solubility can be achieved without a significant increase in viscosity.

The findings demonstrate the effectiveness of the integrated HTE and BO approach for accelerating materials discovery. The efficiency of the method is highlighted by the small number of experimental measurements required to identify high-performing electrolytes. The success in identifying binary solvent systems with unexpectedly high solubility underscores the importance of exploring synergistic solvent effects. This platform not only advances the development of high-performance NRFBs, but also offers a generalizable approach for expedited discovery of functional materials across various domains. The methodology's ability to handle complex solvent systems and incorporate other essential properties will further enhance the practical applications of NRFBs.

This work presents a highly efficient, ML-guided high-throughput robotic platform for the accelerated discovery of optimal electrolyte formulations. The platform successfully identified numerous binary solvent systems with significantly enhanced solubility of the model ROM, BTZ. The integration of HTE with Bayesian optimization showcases a powerful approach for materials discovery in energy storage and beyond. Future work will focus on extending the model to more complex systems and integrating additional critical parameters like viscosity, ionic conductivity, and chemical stability for a more comprehensive electrolyte performance evaluation.

The study focused solely on the solubility of BTZ and did not fully investigate other crucial properties of the electrolyte systems, such as ionic conductivity, viscosity, and long-term stability, which are vital for overall battery performance. The model's accuracy may be affected by the accuracy of the input features, which might not be perfectly reflective of real-world conditions. The generalizability of the current model to more complex solvent mixtures or other redox-active molecules beyond BTZ requires further investigation.