(709c) From Physics-Informed GPR for Carbon Material Supercapacitors to Data-Driven Material Discovery

Amorphous porous carbons are one of the most popular electrode materials for energy storage owing to their high electrical conductivity, large specific surface area and low-production cost. Doping carbon electrodes with heteroatoms such as nitrogen and oxygen proves effective to improve the performance of aqueous supercapacitors. However, the optimal condition of N/O doping remains elusive due to the complexity of the porous structure and electrochemical behavior. Both physics-based models and machine learning (ML) methods have been used to correlate the electrochemical behavior of carbon electrodes. While physics-based models face challenges in capturing the pseudocapacitance effects, empirical correlation of the capacitance with ML methods may lead to erroneous predictions because of the lack of physical inputs, especially at high charging-discharging rates for electrodes with high mesopore surface areas. In this work, we introduced a Gaussian process regression (GPR) method using a physical model as the prior knowledge to limit the coupling effects of different input parameters. The physics-informed GPR proves effective in characterizing the capacitive behavior of N/O-codoped carbon electrodes in both 6 M KOH and 1 M H₂SO₄ aqueous solutions. The quantitative descriptions of the electrode capacitance offer valuable insights for the synthesis of the carbon-based electrode materials with better performance. Guided by our machine learning results, we designed a general sodium amide activation strategy for the synthesis of hyperporous carbons from hypercrosslinked polymers. These hyperporous carbons exhibit a high specific capacitance of 610 F/g in an acid electrolyte, highlighting the power of ML methods for materials discovery.