Optimization of Batch Polymerization Reactors Using Neural-Network Rate-Function Models

Abstract

A simulated verification and validation of the neural-network rate-function (NNRF) approach to modeling the nonlinear dynamic systems is provided. The NNRF modeling scheme utilizes some a priori process knowledge and experimental data to develop a dynamic neural-network model. Based on the obtained neural-network model, an optimal temperature trajectory was computed via the two-step method to drive a batch free-radical polymerization reaction to a prescribed molecular weight distribution (MWD). Evaluation of the quality of the end product suggests that the proposed NNRF modeling approach can be applied in dynamic modeling of a complex and nonlinear reaction system.

Dynamic modeling of batch polymerization reactors via the hybrid neural-network rate-function approach

“…Model equations for batch MMA polymerization reactions and the related kinetic parameters can be found in the literature [1,[14][15][16]. Especially, one can find the details of mathematical description of MMA solution polymerization batch reactor in the appendix of Chang and Hung [15]. The loading conditions (W M (MMA) = 3.492 kg; W s (toluene) = 5.239 kg; [I] (AIBN) = 0.05-0.1 mol/dm 3 ) were used in this study.…”

“…In this way, the postulated functional forms in the NNPF model will no longer be required. The predictive capa- bility of the NNRF model was found to be acceptable and optimal temperature trajectories were calculated accordingly [15]. However, the NNRF model developed [15] was built only on the measurable states.…”

“…The predictive capa- bility of the NNRF model was found to be acceptable and optimal temperature trajectories were calculated accordingly [15]. However, the NNRF model developed [15] was built only on the measurable states. It did not include the state variables difficult to measure, among which was the initiator concentration during the polymerization reaction.…”

“…Direct application of NNPF modeling approach or hybrid modeling approach to the polymerization reaction system will be restricted. Chang and Hung [15] proposed the neural-network rate-function (NNRF) modeling approach to obtain a reliable dynamic neural network model characterizing the complex MMA polymerization reactions carried out in a batch reactor. In the proposed NNRF modeling approach, the rate change of the states as a function of the selected states could be built directly by an FANN for an unknown process.…”

“…It did not include the state variables difficult to measure, among which was the initiator concentration during the polymerization reaction. Therefore, in this work, we will extend the modeling methodology of Chang and Hung [15] and propose the hybrid neural-network rate-function (HNNRF) modeling approach to tackle this problem, which is usually encountered in complex reactor systems. In addition, both the way to measure the causal importance of individual input variables and the computational algorithm to cope with the noisy measurements in building the NNRF or HNNRF model are also addressed in this work.…”