(71a) Addressing Feedstock Variability in the Delignification Process through Direct Application of Kmc for Predictive Control

Tailoring the properties of lignin is essential for its optimized applications. The complex macromolecular structure of lignin, often exceeding tens of kilograms per mole in molecular weight, presents unique challenges in controlling their properties, such as their molecular weight distribution (MWD) and monomeric compositions [1]. Since these characteristics can fluctuate widely during fractionation, understanding of system dynamics becomes more challenging. Moreover, conventional analytical techniques often do not provide detailed insights and can be time-consuming [2,3].

To address these challenges, this study utilizes the kinetic Monte Carlo (kMC) algorithm [4], known for its ability to capture detailed system states. Here, we designed the model with a two-layer structure to comprehensively represent both macroscopic and microscopic phenomena [5]. In our model, the macroscopic layer accounts for the dissolution of lignin into the reaction medium, adhering to global mass and energy balances. The microscopic layer employs the kMC method to simulate microscopic reactions, including de/repolymerization and monomeric conversion, meticulously tracking MWD and monomeric compositions based on the reaction rates computed at each time step. The simulation results align well with the experimental outcomes, providing more detail about the lignin fragments in the reactor, including the lignin populations by individual chain lengths and monomeric compositions.

Although the kMC model offers detailed and accurate system representations, it suffers from a high computational burden [6], limiting its real-time application within model-based control frameworks. To overcome this, we integrated a machine learning strategy into the standard kMC algorithm. Specifically, we circumvented the primary simulation bottleneck – calculation of reaction rates for all existing chains within the system – by training a deep neural network (DNN). This approach bypassed the time-consuming rate calculations, allowing the kMC algorithm to directly choose and execute the event using the DNN-predicted reaction rate distributions. Consequently, the simulation time decreased to below 1% of its original duration while preserving high accuracy.

Finally, the DNN-assisted kMC (DNN-kMC) model was validated against the experiments. Subsequently, the model was directly embedded into a model predictive control (MPC) framework, enabling precise regulation of key lignin properties. This allows the MPC to adjust system parameters to achieve desired characteristics, maintaining the detailed information from the kMC model. Especially, this MPC framework with DNN-kMC can regulate not only average molecular weights but specific MWD and monomeric composition, demonstrating broad applicability in lignin studies. This novel controller will pave the way for delicate customization of lignin properties, possibly handling the feedstock variabilities.

Literature cited:

[1] Gentekos D.T., Sifri R.J., & Fors B.P., (2019). Controlling polymer properties through the shape of the molecular-weight distribution. Nat. Rev. Mater., 4, 761-774.

[2] Montgomery J.R.D., Bazley P, Lebl T, & Westwood N.J., (2019). Using fractionation and diffusion ordered spectroscopy to study lignin molecular weight. ChemistryOpen, 8, 601.

[3] Xiao T., Yuan H., Ma Q., Guo X., & Wu Y., (2019). An approch for in situ qualitative and quatitative analysis of moisture adsorption in nanogram-scaled lignin by using micro-FTIR spectroscopy and partial least squares regression. Int. J. Biol. Macromol., 132, 1106-1111.

[4] Gillespie D.T., (1976). A general method for numerically simulating the stochastic time evolution of coupled chemical reactions. J. Comput. Phys., 22, 403-434.

[5] Lee C.H., Kim J., Ryu J., Won W., Yoo C.G., & Kwon J.S.-I., (2024). Lignin structure dynamics: Advanced real-time molecular sensing strategies. Chem. Eng. J., 487, 150680.

[6] Dybeck E.C., Plaisance C.P., & Neurock M., (2017). Generalized temporal acceleration scheme for kinetic Monte Carlo simulations of surface catalytic processes by scaling the rates of fast reactions. J. Chem. Theory Comput., 13, 1525-1538.