Physics-informed Neural Network to predict kinetics of biodiesel production in microwave reactors

Abstract

Microwaves are a process intensification (PI) method to deliver energy to reactive systems. Microwaves act directly on molecules’ dipolar moment, generating volumetric heating that allows temperature to rise rapidly, which directly impacts the overall rate of a reaction. Because reaction rates adhere to non-linear rate laws, predicting them is challenging. We use a Physics-informed Neural Network (PINN), a physics-driven model, to identify the reaction kinetics of a biodiesel production process. PINNs perform a regression on very few experimental data points and try to fit the physics at hand. We use a microwave reactor with a constant power input to perform the transesterification reaction, measure the infrared temperature and analyze the concentration of glycerides using GC-FID at different reaction times. We train the PINN to predict the reaction rates with respect to the Arrhenius equation. Results show that the PINN successfully identifies the rate constants, including their temperature dependency. Furthermore, the PINN can extrapolate its predictions to other power inputs without ever seeing the concentration data, generating a digital twin of the microwave-assisted reaction.