The development of catalysts for alkaline electrolysis (AEL), such as Raney-Nickel electrodes, is a time-consuming task which involves a lot of manual work and a large number of long experiments. To speed up the electrochemical experiments, this project aimed to explore both, lab automation techniques and machine learning. Lab automation is intended to decrease the time needed for the setup of one experiment. Machine Learning, for which the approach of Bayesian Optimization was used, is meant to reduce the measurement time, which is needed to characterize a certain catalyst material. All in all, this project aims to reduce the effort of the characterization of one material regarding its suitability as a catalyst for alkaline electrolysis.
## Activities
Activities involved a custom, 3D printed electrochemical cell and the development of a script for performing Bayesian optimization to determine optimal parameters.
Lab automation – custom electrochemical cell
A custom, 3D printable electrochemical cell, in which material characterization in three electrode setup can be performed, was designed and printed. The cell is optimized for being low cost, allowing quick and automated setup of experiments and having a high degree of reproducibility. To develop this design, a team of three persons worked together: Sterling Baird, myself, and Bojan Micevski (a technical designer). In a public github issue, all relevant aspects of the first design were discussed and test results were shared. During the time of this project, 3 cell designs and 2 designs for the sample holder were printed. This way, numerous improvements could be implemented within the prototyping stage at high development speed. The prints were carried out with two Stereolithography printers owned by the Acceleration Consortium and two printers owned by the National Research Council (NRC). The material compatibility of the resin from which this cell and sample holder were printed, was tested within this project for the “clear v4” and “rigid 10k” resin provided by Formlabs. Cubes with a volume of 1cm³ each were printed of both materials and were exposed to concentrated KOH solution (30w%). These tests were carried out both, at room temperature and at 80°C. The mass of the test cubes was measured before, during and after the exposure to KOH solution to evaluate any mass loss due to degradation within KOH solution.
You can see the cell in operation at the back of this picture:
Machine Learning for electrolysis life cycle testing
A script was developed for creating an accelerated stress test protocol. By running multiple experiments with the same material with stress parameters selected by the Bayesian Optimization algorithm, the “worst” parameters for a given material are intended to be found. This way a protocol can be established, which is known to stress the electrode to the maximum so that within a short experiment the performance of a given material under maximal stress can be evaluated. To create this script, the search space was defined by specifying the parameters and their constraints. Once this was completed, the Honegumi tool along with the Bayesian Optimization library “Ax Platform” were used to create the actual optimization script. The Bayesian Optimization script for developing this test protocol was completely written during this project. The actual experiments remain to be run in the labs of IFAM Dresden. If this approach turns out to be successful, it could greatly reduce the time for long term stability tests.
Scientific results
A 3D printed electrochemical cell was successfully designed, and a functioning machine learning script was written. Experiments using the printed cell and as suggested by the script are planned.
Lab automation – custom electrochemical cell
Tests in 3 electrode setup for the characterization of a material as a catalyst for AEL need to be carried out in a certain electrolyte solution. This electrolyte needs to be exchanged between all experiments to provide the same conditions at the beginning of each experiment. The cell design developed allows to automate this step completely. The electrolyte can be removed with a pump via an inlet into the cell and a subsequent washing procedure can be applied via an integrated washing nozzle in the inside of the cell which is connected to external water supply. The water of this cleaning step can once again be removed with the previously described pump. Next, fresh electrolyte can be filled into the cell through another inlet. Temperature control is implemented through double wall design, which allows an externally heated thermal fluid to be flowing between the inner and outer wall, thereby heating the electrolyte. Supply of inert gas into the electrolyte to purge solved oxygen is realised through a gas inlet tube which is connected to the cells tap. A Harber-Luggin-Capillary is also integrated into the tap, so that a HydroFlex Reversibel Hydrogen Electrode (Gaskatel) can be easily inserted into the setup. The sample holder, into which the tested material is mounted was designed from scratch to optimize it for the given requirements. These are: Electrical contact between the sample and the external Potentiostat, no contact between any cable/wire and the electrolyte and mechanically reliable mounting of the sample. The design created in this project fulfils all these requirements by combining 3D printed parts with several seals. The sample holder is the only part, which needs to be removed between experiments. Thanks to it being printable with an SLA printer, multiple sample holders can be easily printed, so that it’s easy to prepare several samples at once to reduce downtime between experiments.
Machine learning for electrolysis life cycle testing
The general idea of this approach is to stress a sample during a fixed period of time under discontinuous conditions as it is known that rapid changes in load can contribute to the altering of an electrode material. The time over which a specific current density is applied effects the material as well as the changes in current density. E.g. forcing the Hydrogen Evolution Reaction to occur at the electrodes surface with a current density j=0.5A/-cm² for 2 hours will stress the electrode material differently than running the HER with j=0.4A/cm² for 1h and subsequently with j=0.6A/cm² for 1h, even though the net amount of produced hydrogen is the same. It is of interest to have a test protocol which is known to stress the electrode material significantly in a short period of time, which involves numerous load changes. However, developing such a protocol would be a highly time consuming task as it would involve running a great number of possible test protocols if the number of cycles and their corresponding current density would be tested. That’s why Bayesian Optimization for the development of such a protocol was chosen.
The developed optimization protocol follows the following general structure:
-
Preconditioning (1h) to get the sample into a constant electrochemical state
-
Stress phase (12h)
-
Characterization phase (ca. 2h)
The stress phase will consist of a maximum of n cycles with varying time and current densities each. Current densities are allowed between -2 A and 0A, where 0A are effectively applied as OCP. The time of each cycle can be between 1s and the maximum stress phase time of 12h. The exact number of cycles, corresponding time intervals and current densities are chosen by the Bayesian Optimization Algorithm.
The characterization phase will consist of three steps, which analyse the electrochemical behaviour of the electrode material after the stress phase. These steps contain a galvanostatic step with -0.5A/cm² for 30 min, a Tafel-Analysis with measurements between 0 and -1A/cm² and a cyclovoltammetric step for the measurement of the double layer capacity. The results of this analysis phase are hence the overpotential, Tafel-slope, exchange current density and double layer capacity. These results will then be reported back to the optimization algorithm, which in turn will calculate the stress parameters of the next test protocol.
### Main achievements
-
Prototype of a fully equipped and automatable custom electrochemical cell
-
Many lessons learned from the design process of the custom cell. This experience is highly valuable for the next iterations of the design and is already speeding up the development significantly
-
Script for the development of an accelerated stress testing protocol for electrochemical analysis of electrode materials in a 3-electrode setup
-
Successful completion of a course from the Acceleration Consortiums educational program
-
Many discussions with scientists form University of Toronto, the Acceleration Consortium and the National Research Council. Several ideas for collaborations have already grown from these encounters.
Data management
All data created within this project is available on GitHub within the AccelerationConsortium/ac-training-lab repository. Data about the electrochemical tests in the custom cell will be published in the corresponding issues, too. Custom electrochemical cell (v1): Custom electrochemical corrosion cell for electrocatalysis experiments (v1) · Issue #62 · AccelerationConsortium/ac-training-lab · GitHub Custom electrochemical cell (v2): Custom electrochemical cell (v2) - enhancements and automation · Issue #104 · AccelerationConsortium/ac-training-lab · GitHub Machine learning for life cycle testing: ML for electrolysis life cycle testing · Issue #98 · AccelerationConsortium/ac-training-lab · GitHub
We plan to move the relevant issues from the training lab repository to GitHub - AccelerationConsortium/echem-cell: A custom, 3D printed electrochemical cell for autonomous electrochemistry, targeted at fuel cell electrolysis studies. and flesh out documentation and other details. This repo is currently a placeholder (as of 2025-04-04).