Parameter estimation of distributed activation energy models via chemical reaction neural network

Kinetic parameter estimation is of fundamental importance in modeling the biomass pyrolysis process for biofuel production. In this work, a neural network architecture, named chemical reaction neural network (CRNN), was utilized to learn kinetic parameters (pre-exponential factor and distribution) in distributed activation energy models from the measurement of conversion rate without prior knowledge of the reaction. The Arrhenius equation is reformulated as the activation function of a neuron in the hidden layer of a three-layer neural network. The gradients of loss with respect to kinetic parameters can then be derived analytically, with which a gradient-based training algorithm is employed to optimize the kinetic parameters. The CRNN performance was evaluated based upon systematical numerical investigation of reactions with a double-Gaussian distribution function. The results show that by transforming the optimization problem into neural network training, the CRNN can accurately and efficiently recover the distribution and pre-exponential factor due to the embedded chemical knowledge. The applicability of CRNN in the pyrolysis of rice straw under different heating rates is examined by experimental measurements. It is shown that with the estimation provided the Kissinger method as the starting point, the CRNN is capable of reconstructing the conversion rate curve. We anticipate, as a feasible, efficient, and accurate model, the CRNN will benefit in enhancing the practice of biomass pyrolysis analysis.