Choosing a solvent blend or additive package is fundamentally a search problem: a handful of components, each at a continuous concentration, combine into a space with far more candidate formulations than anyone could test by hand. The Electrolyte Solvent Ratios and Electrolyte Additives pages describe what’s in that space and why each component matters. This post looks at how labs are now searching it computationally, rather than by intuition and grid screening, using Bayesian optimisation coupled to automated experimentation.
Why This Is a Hard Search Problem
A three-salt or three-solvent mixture, discretised into fine concentration steps, produces tens of thousands of candidate blends; a four-component mixture pushes into the millions. Testing every point in that space by conventional grid screening is impractical, and a formulation’s electrochemical response to composition is rarely linear or intuitive. A blend of two ordinary components can outperform either pure component for reasons that aren’t obvious from first principles. This combination (huge search space, non-intuitive response surface, expensive individual experiments) is exactly the setting Bayesian optimisation was built for: it builds a probabilistic model of how performance responds to composition, and uses that model to choose the next experiment most likely to improve on the best result so far, rather than sampling the space uniformly.
A Worked Case: Autonomous Discovery of an Aqueous Electrolyte
Carnegie Mellon University, working with Toyota Research Institute, built a robotic electrolyte test-stand named Otto and connected it directly to Dragonfly, a Bayesian optimisation package, so the software could request a formulation, receive the measured result, and choose the next formulation without a person in the loop (Dave et al., 2020). Otto mixed lithium and sodium salts into water-in-salt aqueous electrolytes and measured electrochemical stability window and ionic conductivity for each blend, at roughly 25 minutes per iteration including washing and flushing steps between runs.
Given four common sodium salts, the exhaustive search space at the test-stand’s working volume worked out to on the order of a million candidate mixtures. Dragonfly instead searched it over 40 hours of continuous, unattended operation and converged on a blend of 96% NaClO₄ to 4% NaNO₃ by volume. A combination that outperformed pure NaClO₄, the previously published state-of-the-art sodium electrolyte, by 24 mV of extra stability margin against oxygen evolution at a fixed leakage current, and suppressed OER current density at 2V by 58% relative to the pure salt. Neither salt alone reaches that performance; the authors note it isn’t obvious from first-principles reasoning why blending them should help, and the leading hypothesis is that the two anions interact to form a more effective passivating film or to suppress water activity at the electrode surface more than either does alone. That’s the case for autonomous search over hand-picked formulations: the optimiser found a non-intuitive blend that a grid survey guided by chemical intuition had no particular reason to try.
Where This Is Heading
The Otto/Dragonfly result is now a few years old. The same closed-loop pattern, a black-box optimiser proposing formulations, automated or robotic hardware executing and measuring them, the result feeding back into the model has since been applied more broadly across electrolyte design.
Multi-objective Bayesian optimisation to trade off ionic conductivity against Coulombic efficiency and interfacial stability simultaneously rather than one property at a time, warm-started optimisation that reuses results from a related formulation family to converge faster on a new one, and solid-state and polymer electrolyte screening alongside the liquid systems the earlier work focused on. The common thread across all of these is treating formulation as a black-box function to be searched efficiently, rather than a space to be understood exhaustively before testing begins. Useful specifically because full first-principle prediction of properties like the stability window of a concentrated blended electrolyte remains computationally out of reach for routine screening.
The limitation worth keeping in mind: these methods optimise within the design space and measurement objectives they’re given. They’re very good at finding the best blend of the salts or solvents you decided to test, at optimising the properties you decided to measure. However, choosing which components belong in the candidate pool and which properties actually matter for a real cell (as opposed to a symmetric test cell), is still a human judgment call upstream of the optimisation loop.
References
- Dave, A.; Mitchell, J.; Kandasamy, K.; Burke, S.; Paria, B.; Poczos, B.; Whitacre, J.; Viswanathan, V. Autonomous Discovery of Battery Electrolytes with Robotic Experimentation and Machine-Learning, Carnegie Mellon University, 2020 (supported by Toyota Research Institute’s Accelerated Materials Design and Discovery program)
- Autonomous Discovery of Polymer Electrolyte Formulations with Warm-Start Batch Bayesian Optimization, Chemistry of Materials, 2025
- Machine Learning in Electrolyte Design: Accelerating Discovery and Multidimensional Optimization for Next-Generation Batteries, Chem, 2026
See also: Electrolyte Solvent Ratios, Electrolyte Additives