Maximizing Supercapacitor Capacitance Through Adopting Convolutional Neural Network Guided Bayesian Optimization of 3D Printing

Supercapacitors are desired in the field of energy harvesting for its high capacitance. Although material properties are the primary factors in determining capacitance, the physical structure is also a dominant factor. 3d printed lattice structures are among preferred structures of supercapacitors as they allow a large surface to volume ratio (SAV). Once the structural rigidity is guaranteed, the printed lines must be narrow to maximize the surface area. Thus, in direct ink writing (DIW) 3d printing, printing parameters are tuned to find an optimal condition under which a uniform and thin line can be printed. However, this process is not only laborious but also costly due to many parameters that can influence printing quality. Recently, there has been a novel trial to implement Bayesian optimization, a technique used in machine learning, to optimize printing parameters [1]. With sample sizes less than 32, the Bayesian optimizer suggested printing conditions of different inks with higher printing resolution. In this work, we introduce a convolution neural network (CNN) guided Bayesian optimization to enhance the performance of supercapacitor through maximizing SAV of 3d-printed lattice electrode. The two optimization goals of high uniformity and narrow width in 3d printed lines are reached through a single scoring method: standard deviation of line widths. Through a camera with 600 x 800 resolution, 800 width measurements are made for each 6.5mm line printed from a 3d printer (Pro4, Nordson) and standard deviation of the measurements are calculated. The Bayesian optimizer attempts to find the global maxima of a surrogate model built from these standard deviation values, fed in negative. Throughout the process, the CNN model guides the Bayesian optimizer to keep its search space within straight line printed regions, preventing costly explorations and reducing total optimization time. The CNN model categorizes the printed lines to discontinuous, straight, coiled, or aggregated and resets the Bayesian optimizer’s parameter boundaries when it detects a non-straight line. 7mm x 7mm x 0.72mm lattice structure supercapacitors are made using either graphene oxide (GO) ink or poly(3,4-ethylenedioxythiophene):polystyrene sulfonate (PEDOT:PSS) ink. Three printing parameters: nozzle speed, nozzle height, and pressure, are optimized. Throughout 50 iterations of Bayesian optimization, the CNN model reset the search space 6 times for GO ink and 7 times for PEDOT:PSS ink. Supercapacitors were 3d printed with the optimized parameters and compared to supercapacitors prepared with conventional DIW printing method [3]. For GO and PEDOT:PSS supercapacitor, ligament widths decreased by 78.3% and 82.5% respectively. While ligament spacing was set to 0.3mm, number of unit cells increased by 158.2% and 156.3%, enhancing the pore architecture of the lattice structures. The capacitance, measured by cyclic voltametric method, increased by 121.5% and 108.4%.<br/><br/><b>Acknowledgement</b><br/>This work was supported by the National Research Foundation of Korea(NRF) Grant funded by the Korean<br/>Government(MSIT) (No. 2020R1A2C3013158)<br/><br/><b>REFERENCES</b><br/>[1] Kalani Ruberu, et al, <b>Applied Materials Today</b>, 2020, 22, 100914<br/>[2] Timothy Erps, et al, <b>Applied Science and Engineering</b>, 2021, 7, 22<br/>[3] Hyunwoo Yuk, el al, <b>Applied Materials</b>, 2018, 30, 6