Chemical Reaction Neural Networks

Chemical Reaction Neural Network (CRNN), is a type of neural network for automatic (template-free) reaction pathway modeling and uncertainty quantification, originally proposed by the DENG lab at MIT. 

CRNN hard-encodes the widely-applicable law of mass action and the Arrhenius law into the neural network architecture, inferring unknown kinetic parameters from experimental data. Previously published works (linked below) have demonstrated the capability of CRNN to learn meaningful kinetic and thermal-kinetic mechanisms and parameters from species concentration trajectories, pyrolysis mass histories, and heat release profiles. CRNN has also been demonstrated as a tool for uncertain model inference using uncertain or incomplete datasets.

CRNN is fully interpretable and is widely applicable, with previous application in various regimes including a range of temperatures, types of experimental data, amounts of prior knowledge, and data sparsity. As a new development in the field of scientific machine learning, this list is not comprehensive. We anticipate further development and exciting applications of CRNN in the area of autonomous reaction pathway modelling and uncertainty quantification. We invite others to download the example code from our Github repository linked below to try out CRNN, and see if it works for your modelling needs.

How CRNN works

An reaction system consisting of 4 species and 4 reactions is shown above for demonstration purposes. (a) A single reaction is represented as a neuron whose weights and biases can be directly interpreted as model parameters in the law of mass action and Arrhenius law; (b) Multiple reaction neurons are stacked into one hidden layer to formulate a CRNN for multi-step reactions. The number of intermediate species and reactions can be treated as a hyperparameter of the CRNN and determined by the principle of parsimony. Further capabilities detailed below include the learning of uncertain kinetic mechanisms, as well as thermal-kinetic parameter inference based on heat release data.