Source-linked AI summary
Autonomous optimization of nonaqueous battery electrolytes via robotic experimentation and machine learning
Adarsh Dave, Jared Mitchell, Sven Burke, Hongyi Lin, Jay Whitacre, Venkatasubramanian Viswanathan
TL;DR
The paper tackles efficient optimization of nonaqueous battery electrolytes, where electrolyte design is complex and conductivity is only one part of fast-charging performance. It couples the robotic platform Clio with a Bayesian experiment planner, finding improved fast-charging performance in pouch cells relative to an intuitively chosen baseline while acknowledging limits in the single-objective study.
Problem
Nonaqueous electrolyte design requires optimizing complex formulations, while the study’s fast-charging objective does not encompass all relevant electrolyte properties.
Method
The workflow couples the custom-built robotic platform Clio with a Bayesian experiment planner to optimize electrolyte conductivity and test selected electrolytes in pouch cells.
Results
The workflow produces an efficient design of experiments and finds better fast-charging performance in a cell than an intuitively chosen baseline.
Takeaways & Limitations
Closed-loop robotic experimentation can optimize nonaqueous electrolyte conductivity and connect the resulting formulations to pouch-cell performance.
Takeaways & Limitations
The study initially optimizes only bulk conductivity, an incomplete objective function, and the electrolytes are not optimized for long-term performance.
Abstract
from arXiv · showhide
In this work, we introduce a novel workflow that couples robotics to machine-learning for efficient optimization of a non-aqueous battery electrolyte. A custom-built automated experiment named "Clio" is coupled to Dragonfly - a Bayesian optimization-based experiment planner. Clio autonomously optimizes electrolyte conductivity over a single-salt, ternary solvent design space. Using this workflow, we identify 6 fast-charging electrolytes in 2 work-days and 42 experiments (compared with 60 days using exhaustive search of the 1000 possible candidates, or 6 days assuming only 10% of candidates are evaluated). Our method finds the highest reported conductivity electrolyte in a design space heavily explored by previous literature, converging on a high-conductivity mixture that demonstrates subtle electrolyte chemical physics.
1 Introduction
The paper addresses the difficulty of optimizing complex nonaqueous electrolytes by coupling robotic experimentation with machine-learning-guided closed-loop design. It focuses on conductivity optimization for fast charging within a defined ternary solvent, single-salt space.
- Motivation: Closed-loop battery-material optimization had not yet been demonstrated outside aqueous electrolytes.The paper positions this gap alongside prior demonstrations in related fields.
- Motivation: Electrolytes are difficult to optimize because solvent and salt choices, component proportions, and multiple performance objectives create a high-dimensional design space.The relevant properties include bulk transport and anode interfacial kinetics.
- Approach: Clio combines high-throughput electrolyte characterization with a machine-learning experiment planner to autonomously explore and optimize a chosen objective.The platform measures transport properties of nonaqueous solvent and salt blends while integrating returned experimental results with planning.
- Objective: The study initially optimizes bulk conductivity as a single objective for improving rate-capability and acknowledges that this is an incomplete objective function.The workflow is also described as enabling future multi-objective electrolyte optimization.
- Study design: Clio optimizes solvent mass fraction and salt molality in an EC-DMC-EMC-LiPF6 ternary-solvent, single-salt system before selected electrolytes undergo fast-charging pouch-cell tests.The results are compared with a baseline electrolyte selected a-priori from the design space.
2 Automated Electrolyte Characterization
Clio automates electrolyte preparation and characterization, returning measurements to support closed-loop experimentation. Its repeatability was assessed through rinsing, triplicate measurements, and contamination studies.
- Automated characterization: Clio doses electrolyte formulas from pre-made feeder solutions and measures conductivity, viscosity, UV-vis spectrum, and density.These measurements provide automated characterization of each specified composition.
- Measurement protocol: Each composition is tested in triplicate, with the system rinsed between experiments as part of contamination-control procedures.The contamination study is reported in Extended Data Figure 1.
- Measurement reliability: ±1.3% was the mean absolute difference between repeated conductivity measurements of the same electrolyte across a range of electrolytes.Samples can also be retained for follow-up pouch-cell testing.
- System integration: Clio is controlled over HTTP, with designs passed to the system in JSON and experimental results returned for integration with an experiment planner.Hardware, software, and operational details are provided in Methods.
3 Machine-Learning Guided Experimentation
The workflow couples Clio’s automated electrolyte experiments with Dragonfly’s Bayesian optimization to search a constrained ternary solvent design space. It converged on a conductivity optimum and identified the highest-conductivity blend reported from this heavily studied space.
- Workflow and design space: Clio creates electrolyte samples by mixing up to 19 feeder solutions and measures conductivity, density, viscosity, and UV-vis spectrum.Tested samples can be retained for follow-on cell testing, extending the workflow from property measurement to cell performance.
- Workflow and design space: Dragonfly uses Bayesian optimization to adaptively sample a three-dimensional EC-DMC-EMC, LiPF6 electrolyte design space.The axes encode EC mass fraction, DMC co-solvent ratio, and LiPF6 molality, with more than 1000 discretized candidate points.
- Optimization behavior: 15 experiments were sufficient for Figure 2 to illustrate convergence on a conductivity optimum.Dragonfly combined acquisition strategies that balance exploration and exploitation with periodic random sampling to favor exploration.
- Optimization results: 13.7 mS cm−1 was measured at EC:DMC 40:60 by mass with 0.9m LiPF6 at 26–28 °C.Follow-up measurements along EC mass-fraction contours confirmed the optimum within the studied design space.
- Optimization results: The 40% EC blend had higher predicted conductivity than 30% or 50% EC, consistent with improved ion dissociation and lower viscosity.The cited interpretation attributes the balance to improved ion dissociation relative to 30% EC and reduced viscosity relative to 50% EC.
- Optimization results: The workflow found the highest-conductivity blend yet reported from the heavily explored EC:DMC:EMC LiPF6 space.This result demonstrates the stated value of combining machine learning with automation for electrolyte optimization.
4 Pouch Cell Testing of Discovered Electrolytes
Selected electrolytes discovered by Clio were tested in small pouch cells using rate tests up to 4C charging. The Clio blends matched or exceeded baseline 4C discharge capacity, although long cycle life was not optimized.
- Testing protocol: Small pouch cells underwent formation, a five-step rate test up to 4C charge, and repeated 4C charge cycling until over-voltage or capacity-fade failure.Discharges between rate-test steps were performed at 0.5C.
- Rate-test results: All Clio-optimized blends showed greater or equal discharge capacity at the 4C step compared with the baseline.The result indicates greater usable capacity at this high charging rate for the tested Clio blends.
- Rate-test results: 5% was the improvement for the worst Clio cell versus the worst baseline cell after 4C charging, while the best-cell improvement was 13%.Both comparisons use discharge capacity after the 4C charge step.
- Scope boundary: The electrolytes were not optimized for long cycle life.The study instead used the 4C rate-test discharge capacity as a metric connected to five-cycle performance.
- Extended cycling: Discharge capacity after the 4C rate-test step strongly correlated with capacity after five back-to-back 4C charge cycles.The figure caption identifies unfilled markers as cells that did not complete five cycles.
5 Assessing Workflow Efficiency
Clio’s efficiency is evaluated against manual optimization using separate time and sample-efficiency measures, combined into an overall estimate. The workflow is estimated to reduce both work-days and experiments needed for optimization.
- Time efficiency: Time efficiency asks whether Clio or a graduate student completes 40 experiments faster.The comparison accounts for preparation, solution-making, and manual-testing time.
- Sample efficiency: Sample efficiency asks whether Clio reaches a fixed optimization target with fewer samples than a human-led design.The comparison uses Clio’s machine-learning-guided DOE against factorial designs.
- Time efficiency: 33% fewer work-days are required for Clio to complete 40 experiments than for a graduate student.Clio can operate 24 hours a day, compared with 8 hours a day for a researcher.
- Sample efficiency: 3x the samples are required by the human-run 5-level, 3-factor factorial compared with Clio’s DOE.The human factorial represents 10% of the design space posed to Clio, while Clio uses a 10-level, 3-factor discretization.
- Overall efficiency: Approximately 3x overall efficiency is estimated for Clio across time and sample efficiency.Clio is estimated to complete in 2 work-days what would take a graduate student greater than one work-week.
6 Conclusion
The study demonstrates a closed-loop workflow combining robotics, experiment planning, and device testing for nonaqueous electrolyte optimization. It identifies an unreported conductivity optimum and fast-charging candidates while estimating a 3x speed-up over manual optimization.
- Conclusion: The workflow performs closed-loop optimization of a nonaqueous battery electrolyte and integrates conductivity optimization with pouch-cell testing.The workflow couples a custom robotic platform with experiment planning and device testing.
- Conclusion: The Bayesian experiment planner produces a highly efficient DOE and finds a yet-unreported conductivity optimum in a well-studied design space.The result extends optimization beyond simply revisiting known candidate formulations.
- Conclusion: The workflow identifies candidates with better fast-charging performance in a cell than an intuitively chosen baseline.The conclusion connects electrolyte discovery to measured cell performance.
- Conclusion: Closed-loop experiments have potential to discover optimal material designs in well-explored and unexplored design spaces.The authors frame the workflow as relevant beyond this electrolyte study, including other autonomous discovery platforms.
- Conclusion: 3x overall speed-up is estimated relative to manually conducted optimization.The estimate combines time and sample efficiency.
Methods
The methods describe Clio’s automated electrolyte handling, conductivity measurement, software integration, calibration, and pouch-cell testing. Measurements use controlled temperature and atmosphere conditions, repeated evaluations, and standardized cell protocols.
- Clio platform: Clio automates electrolyte dosing, mixing, rinsing, and measurement through a shared closed volume.The platform uses a precision pump, programmable valve set, and impedance analyzer.
- Repeatability: Triplicate electrolyte evaluations use the final two runs averaged as the reported measurement.Repeatability is assessed across 120 samples spanning carbonate-solvent compositions.
- Experimental conditions: All reported conductivity measurements remain between 26 °C and 28 °C in a dry argon environment.Clio maintains moisture below 10 ppm during operation.
- Conductivity measurement: Conductivity is measured by impedance spectroscopy using symmetric platinum electrodes and a calibrated cell constant.The reported specific conductivity is calculated from inverse resistance divided by the cell constant.
- Composition control: Density estimates convert Dragonfly’s composition axes into feeder-solution volumes and are confirmed by Clio’s density measurement.The estimates come from the Advanced Electrolyte Model.
Extended Data
Extended data evaluates measurement repeatability, contamination, cycling behavior, composition-dependent conductivity, and model-based explanations of the conductivity optimum. These analyses support a maximum near 40% EC and identify high-salt formulations among the surviving cells.
- Measurement repeatability: Conductivity repeatability error is reported as uncorrelated with conductivity level.This supports using repeated measurements across the design space to assess experimental consistency.
- Measurement repeatability: 1.3% mean absolute percentage error is measured across repeated electrolyte evaluations, with a 95% confidence interval of 3.8% error.The repeatability study covers 120 electrolytes in the EC-DMC-EMC-DEC + LiPF6 design space.
- Cell cycling: Two Clio-discovered cells survive 30 repeated 4C-charge, 0.5C-discharge cycles without over-voltage or sudden death.Both surviving cells are high-salt-concentration electrolytes, identified as E and B.
- Conductivity optimum: The EC:DMC-only survey finds maximum conductivity near 40% EC at 26–28°C.The re-surveyed contours reproduce this maximum within 2% of each EC mass-fraction level despite a small density-calculation error.
- Model interpretation: The Advanced Electrolyte Model corroborates a conductivity maximum near 40% EC at 30°C.The model associates the behavior with single-ion population and viscosity covariates, while experiments find a higher peak difference at lower molality.