Analyzing Structural Contributions of 3D-Printed Porous Electrodes
We have successfully streamlined the generation of state of charge maps. Our Attention Residual U-Net CNN achieved 70% accuracy in semantic segmentation, demonstrating a strong baseline for classifying key regions within the porous electrode system. We also successfully clustered tails using DBSCAN and quantified their average intensity, allowing for structural analysis of reduced electrolyte regions.
Khoi Bui1, Sofía Edgar2, Michael Aziz2
1Orange Coast College, Costa Mesa, California
2Harvard John A. Paulson School of Engineering & Applied Sciences, Cambridge, MA
Department of Energy, DE-SC0020170
1Jin, Shijian, et al. “Near Neutral pH Redox Flow Battery with low Permeability and Long-Lifetime Phosphonated Viologen Active Species.“ Advanced Energy Materials, vol. 10, no. 20, 2020, p. n/a-n/a, https://doi.org/10.1002/aenm.202000100.
Results
Methods
Introduction
Conclusion
References
Acknowledgments & Contact
Fig. 3: 3D image of the electrode (hidden),
bulk electrolyte, and tails with color map
y
z
x
ZX Slice at Y = 129
SOC Map
Segmentation of a tail cluster into individual tails using DBSCAN
Tail Map (with color map)
Fig. 4: Tails and their SOC in the 129th slice
x
z
ZX Slice at Y = 442
SOC Map
Tail Map (with color map)
Fig. 5: Tails and their SOC in the 442nd slice
Segmentation of a tail cluster into individual tails using DBSCAN
x
z
Our research investigates how the structure of redox flow battery's electrodes affects electrolyte flow and reduction efficiency by integrating electrode design with image-based machine learning analysis. To eliminate the heterogeneity of conventional electrodes, we fabricate 3D-printed electrodes with predefined mesh geometries. As the electrolyte flows through these structures, we capture high-resolution 3D images of the "tails" of reduced electrolyte, which reflect the localized electrolyte reduction behavior. From these images, we created and trained a convolutional neural network to segment images into three classes: electrode, bulk electrolyte, and tails. Finally, we quantify the segmented tails by extracting features such as tail area and state of charge, which allows us to assess mass transfer limitations. This approach enables us to systematically study how electrode geometry influences mass transfer behavior and identify the core principles that govern this relationship. This fundamental understanding can serve as a stepping stone toward developing optimal electrode structures that maximize electrolyte utilization while minimizing flow resistance.
Step 1: Generate State of Charge (SOC) Map
Step 3: Create a machine learning model
Step 4: Quantify the tails
Step 2: Slice 3D Image Stack Along the Y-dimension to Obtain
ZX-Plane Views
Use DBSCAN to identify tails and calculate their average intensities.
Use Trainable Weka Segmentation to label the image into 3 classes: electrodes, bulk (oxidized) electrolyte, and tails (reduced electrolyte).
Then, use segmented images to develop an Attention Residual U-Net CNN.
Fig. 2: Process of making a SOC map
Open Circuit Voltage
0% SOC
Full Charged
100% SOC
Charged
Variable SOC
SOC Map
SOC Map (with color map)
Apply color map
Thank you to the following people for this amazing opportunity!
Khoi Bui: khoi.tm.bui@gmail.com
Sofía Edgar: sedgar@g.harvard.edu
Michael Aziz: maziz@harvard.edu
Fig. 1: Scheme of BPP−Vi | ferrocyanide flow battery with enlarged reaction chamber.1