Neural networks applied in kinetic analysis of complex nucleation-growth processes: Outstanding solution for fully overlapping reaction mechanisms

Abstract

Performance of several neural network architectures (convolutional neural network CNN, multilayer perceptron MLP, CNN/MLP hybrid CDD) was evaluated for kinetic analysis of complex processes with overlapping independent reaction mechanisms based on the nucleation-growth Johnson-Mehl-Avrami (JMA) model. Theoretically simulated data used for the testing covered absolute majority of real-life JMA-JMA solid-state kinetics scenarios. The performance of the tested architectures decreased in the following order: MLP > CDD >> CNN. For partially overlapping processes the CDD and MLP architectures provided accurate estimates of the JMA model kinetic parameters, performing on par with traditional methods of kinetic analysis. For the fully overlapping kinetic processes, the accuracy of the estimates provided by the neural networks significantly worsened, however still largely outperforming the traditional approaches of kinetic analysis based on the standard non-linear optimization, such as mathematic or kinetic deconvolution. The corresponding kinetic predictions were of suitable precision for majority of real-life applications preparation (glass-ceramics).