A Critical Evaluation of Using Physics-Informed Neural Networks for Simulating Voltammetry: Strengths, Weaknesses and Best Practices

The recent explosion of applications of physics-informed neural networks (PINNs) as a discretization-free tool to solve partial differential equations (PDEs) shows great potential for applications in electroanalytical simu-lations. However, a simple, naive PINN approach may fail to make analytical level predictions in even only moderately complicated systems. Here, we explore eight test cases, spanning 1D to 3D simulations, including both cyclic voltammetry and chronoamperometry, and a wide selection of electrode geometries from macro -electrode, (hemi-)spherical electrode, microband electrode, cube electrode and microdisc electrode to serve the dual-purposes of expanding PINNs to more challenging electrode geometries and to recommend best prac-tices in the context of electroanalytical simulation. These best practices, include the use of dimensionless parameters, non-zero conditioning times, mathematical transformation of PDEs, sequence-to-sequence train-ing, adaptive weights algorithms, optimal batch sizing, domain decomposition, learning rate scheduling and transformation of coordinates. These suggested best practices are the intended key contribution of this paper, as to position future PINN users with a well-informed starting position for generic electroanalytical PINN sim-ulations, to avoid known difficulties and to skip the trial-and-error phase with hyperparameter tuning. We believe that these recommendations can serve as primers for PINN simulations for sophisticated Multiphysics problems, and make PINN simulations more accessible.

Automated summary

This study explores various electrode geometries and challenging test cases, recommending best practices in electroanalytical simulations to enhance the accuracy and reliability of PINNs simulations.