Reaction engineering with recurrent neural network: Kinetic study of Dushman reaction

Abstract

A recurrent neural network (RNN) was used to predict the time evolution of the Dushman reaction for the first time. The structure of the RNN was optimized to reduce the training time and amount of data required by introducing four significant features: concentration change as an output; log-scale calculation in the neural network; mass balance calculation; and pre-training of an RNN cell using two different rate laws, followed by fine-tuning.The pre-trained RNN models showed similar accuracy to that of a method using integration of the original rate laws (mean squared error of 3.30 and 1.28 μM2 for each rate law), as expected. The fine-tuned models exhibited significant improvements (mean squared error of approximately 0.10 μM2 for both rate laws) regardless of the accuracy of the original rate laws. No overtraining was observed during the fine-tuning because of the implementation of log-scale and mass balance calculations, which helped the RNN models to effectively learn from training data scattered by several orders of magnitude and avoid theoretically invalid curvature. The proposed RNN model is suitable for practical applications because it can achieve high accuracy with limited training time and a limited amount of experimental data (only 88 conditions were used to prepare training dataset in this work) and does not require detailed investigation of the reaction mechanism. The results show that if a theoretical background is properly introduced by the features mentioned above, the RNN models are highly versatile in terms of reaction kinetics.