Skip to main content

Our machine learning-based digital transformation study, “Multiphysics generalization in a polymerization reactor using physics-informed neural networks,” has been published in Chemical Engineering Science.

Congratulations, Yubin and Sunkyu!


  • We applied physics-informed neural networks (PINN) to an ethylene radical polymerization reactor.
  • Continuity, Navier-Stokes, and species balance equations are incorporated in the PINN.
  • PINN more accurately captured changes of operating conditions than conventional NN.
  • PINN satisfactorily extrapolated ethylene conversions by capturing the reversed tendency of conversion beyond the trained range.


Multiphysics engineering has been a crucial task in a chemical reactor because complicated interactions among fluid mechanics, chemical reactions, and transport phenomena greatly affect the performance of a chemical reactor. Recently, physics-informed neural networks (PINN) have been successfully applied to various engineering problems thanks to their domain generalization ability. Herein, we introduce a novel application of PINN to multiphysics in a chemical reactor. Specifically, we examined the effectiveness of PINN to reconstruct and extrapolate ethylene conversion in a polymerization reactor. We ran CFD for the polymerization reactor to use in the training process; thereafter, we constructed the PINN by combining the loss of conventional neural networks (NN) with the residuals of the continuity, Navier-Stokes, and species transport physics equations. Our results showed that the PINN more accurately predicted the overall ethylene concentration profile, which is the primary result of multiphysics in the reactor; PINN showed 18 % lower mean absolute error (0.1028 mol/L) than NN (0.1267 mol/L). Furthermore, the PINN satisfactorily predicted the conversion concaveness effect, which is a unique multiphysical effect in a radical polymerization reactor, while NN couldn’t. These results highlight that multiphysics in a chemical reactor may be efficiently predicted and even extrapolated by harnessing physics in neural networks.

Leave a Reply