Source-linked AI summary
Autonomous discovery of battery electrolytes with robotic experimentation and machine-learning
Adarsh Dave, Jared Mitchell, Kirthevasan Kandasamy, Sven Burke, Biswajit Paria, Barnabas Poczos, Jay Whitacre, Venkatasubramanian Viswanathan
TL;DR
The study examines electrolyte-mixture design within a constrained testing volume and uses repeated robotic runs to generate experimental data. It reports a four-day machine-learning optimization experiment.
Problem
Electrolyte-mixture testing is constrained by the available 7 mL testing volume.
Method
The experiments used repeated wash, initial, and production runs, with the production run reported as data.
Results
The machine-learning optimization consisted of four days of experimentation divided into daily runs.
Takeaways & Limitations
The experimental setup supports repeated, standardized collection of electrolyte-mixture data within a fixed-volume design space.
Abstract
from arXiv · showhide
Innovations in batteries take years to formulate and commercialize, requiring extensive experimentation during the design and optimization phases. We approached the design and selection of a battery electrolyte through a black-box optimization algorithm directly integrated into a robotic test-stand. We report here the discovery of a novel battery electrolyte by this experiment completely guided by the machine-learning software without human intervention. Motivated by the recent trend toward super-concentrated aqueous electrolytes for high-performance batteries, we utilize Dragonfly - a Bayesian machine-learning software package - to search mixtures of commonly used lithium and sodium salts for super-concentrated aqueous electrolytes with wide electrochemical stability windows. Dragonfly autonomously managed the robotic test-stand, recommending electrolyte designs to test and receiving experimental feedback in real time. In 40 hours of continuous experimentation over a four-dimensional design space with millions of potential candidates, Dragonfly discovered a novel, mixed-anion aqueous sodium electrolyte with a wider electrochemical stability window than state-of-the-art sodium electrolyte. A human-guided design process may have missed this optimal electrolyte. This result demonstrates the possibility of integrating robotics with machine-learning to rapidly and autonomously discover novel battery materials.
Methods · Materials
The materials comprised commonly used sodium and lithium salts purchased from VWR International and Sigma Aldrich. The salts were specified by chemical grade and, where stated, used without further purification.
- Materials: Lithium perchlorate (CAS 7791-03-9, ACS reagent, >95.0%) and lithium nitrate (CAS 7790-69-4, Reagentplus) were purchased from Sigma Aldrich.
- Materials: Lithium sulfate (CAS 10377-48-7, Titration >98.5%) and lithium bromide (CAS 7550-35-8, anhydrous, >99.%) were purchased from Sigma Aldrich.
- Materials: The purchased salts were used without further purification.
Material storage
Materials were stored and handled under specified atmospheric and container conditions, while electrolyte solutions remained sealed at ambient laboratory temperature.
- Bromide solids were stored and massed in dry argon, whereas sodium perchlorate remained in unopened containers.
- All other solids were stored in ambient laboratory atmosphere in parafilm-sealed containers and massed under ambient conditions.
- Electrolyte solutions were stored in sealed Fisher Scientific Kimble Kimax GL 45 containers at 22◦C ± 2◦C.
- Deionized water for test-stand dilution was stored exclusively in Fisher Scientific Kimble Kimax GL 45 containers.
Preparation of solutions
Solutions were prepared under sealed, ambient-atmosphere conditions with magnetic stirring and temperature control tailored to their thermal behavior. All solutions were mixed for at least 30 minutes before density measurements.
- Solutions were mixed in ambient atmosphere and sealed with parafilm after all solids were added to deionized water.
- Magnetic stirring continued for at least 30 minutes after dissolution of the last visible solid.
- Temperature was controlled with hot plates capped at 30 ◦C for endothermic solutions and ambient-temperature water baths for exothermic solutions.
- All solutions were mixed for at least 30 minutes at ambient conditions before density measurements.
Experimental details
Experiments used a standardized three-run Otto procedure with a constant 7mL testing volume, while optimization proceeded over four days of daily 10–15-iteration runs. Platinum surfaces were cleaned between selected electrolyte evaluations.
- Experimental procedure: Each experimental iteration comprised three Otto runs: deionized-water washing, an initial requested-mixture run, and a production run reported as data.This procedure had the highest fidelity against benchmark cases.
- Experimental procedure: 7mL testing volume was held constant across experiments.
- Optimization schedule: Machine-learning optimization ran for four days, with daily runs of 10–15 iterations.Solutions were restocked as needed and prepared according to the preceding preparation section.
- Surface preparation: Platinum surfaces were cleaned between NaClO4 and Blend E evaluations using 1200 grit wet paper, isopropyl alcohol, and deionized water.
pH corrections
Reported potentials were standardized to pH zero using the Nernst-equation correction, which changes half-cell potential by 59 mV per pH unit.
- pH corrections: 59 mV per pH unit is the Nernst-equation change in half-cell potential.This correction supports performance standardization across the Pourbaix diagram.
- pH corrections: All reported potentials, unless otherwise noted, were shifted to pH zero for standardization across the Pourbaix diagram.The adjustment establishes a common pH reference for comparing reported potentials.
- pH corrections: Ereported = Emeasured+0.0591pHmeasured defines the pH correction applied to measured potentials.The equation expresses the shift to the pH-zero reference.
Tafel equation
The Tafel equation is derived from Butler–Volmer kinetics by neglecting the backward reaction. The electrolyte design space is modeled as constrained mixture volumes and discretized at 0.1 mL resolution for combinatorial analysis.
- Tafel equation: The Tafel equation derives from the Butler–Volmer kinetic law by applying an approximation that ignores the backward reaction.This yields the functional form used for Tafel analysis.
- Tafel equation: Tafel plots convert current density to A/cm2 and report electrode potentials on the standard hydrogen electrode scale.These conventions were used for Figure 3 and Extended Data 5, 6, and 7.
- Combinatoric estimation of design spaces: 7mL constrains the mixture design space, represented by feeder-solution volumes and dilution water satisfying x1 + x2 + ... + xd + xwater = 7.Each xi denotes the volume of feeder solution i, with up to d salts and optional water.
- Combinatoric estimation of design spaces: 0.1 mL discretization was used for all electrolyte optimizations, converting the volume constraint to 10x1 + 10x2 + ... + 10xd + 10xwater = 70.The scaled equation supports combinatorial complexity analysis across d salts, plus one water variable.
Data availability
Supporting data for the paper’s graphs and other study findings are available from the corresponding author upon reasonable request.
- Supporting data for the paper’s graphs and other study findings are available from the corresponding author upon reasonable request.
Extended Data
The extended data detail Otto’s robotic electrolyte-mixing and electrochemical-testing setup, validate its staircase-potentiometry assay against literature, and document follow-up Tafel analyses for Blend E and NaClO4.
- Robotic test stand: Otto combines near-saturation single-salt feeders, a jacketed mixing vessel with pH metering, conductivity measurements, and electrochemical testing.The test stand includes pumps, valves, a waste vessel, a rotary mixer, conductivity and pH probes, a potentiostat, a multiplexer, and a power supply.
- Electrochemical measurements: Staircase potentiometry measures potential at successive current densities using platinum electrodes and an Ag/AgCl reference; one example tests 7mL of 16 molal NaClO4.Derived quantities use the average of the last 2 seconds of each current step after 10seconds of 111 mA/cm2 preconditioning on each side.
- Method validation: 25 experiments benchmarked Otto’s staircase-potentiometry Tafel extrapolation against 1M KOH literature results, reproducing the 1966 Tafel slope within reported experimental uncertainty.A 50mV potential difference was attributed to possible electrode-surface or geometry differences.
- Follow-up electrolyte evaluation: 10 sequential recirculating runs evaluated Blend E and NaClO4 using half-logdecade current-density steps from 10−1 to 10−5 A/cm2.Only current densities from -2 and -5 were fitted, and the first three runs for each electrolyte were excluded because of platinum-oxide conditioning or formation.
- Tafel analysis: 7 retained runs were each fit with a separate linear Tafel equation, then averaged into the single curve shown in the main paper’s Figure 3.The retained-run parameters provided the averaged Tafel curve for the main-paper comparison.