Robust bioreactor management strategies for the production of food ingredients using adaptive control and digital twins under stochastic process perturbations

Abstract

The article is devoted to improving the stability of biosynthetic processes in bioreactors of food biotechnology under stochastic disturbances caused by heterogeneity of raw materials, drift of metabolic characteristics of producers, fluctuations in physico-chemical conditions and scaling effects leading to gradients of substrate and oxygen and population heterogeneity. The transition from rigid control schemes to intelligent circuits is considered, combining robust control, parameter adaptation and digital twins synchronized with the object via sensor data streams and supporting scenario prediction of control actions before their physical implementation. The methodological part is based on computational experiments on 150 fermentations in 5 and 100 L apparatuses for the biosynthesis of Xanthomonas campestris exopolysaccharides, complicated by a highly viscous medium and nonlinear rheology; Temperature, pH, pO2, optical density, and residual substrate were recorded in 1-minute increments, and disturbances were modeled by additive and multiplicative noise with specified spectral properties. The digital twin is implemented as a hybrid of the mechanistic kinetics of Mono—Yerusalimi and a neural network LSTM module that compensates for structural uncertainty.; The parameters were identified by genetic algorithms, and the comparison of the strategies of PID, MPC, and adaptive H∞-regulator was performed using integral quality indicators and dynamic metrics in Monte-Carlo series (5000 runs) with ±20% parameter variations and real-time computational delay analysis. It is shown that as the measurement noise of the substrate increases to 11.1%, the standard deviation of the product concentration increases from 0.458 to 4.892 g/l for PID, while adaptive H∞ keeps the spread at 0.112–0.492 g/l, demonstrating significantly lower sensitivity than MPC. In pO2 transient modes with viscosity drift, intelligent regulators provide a dramatic reduction in overshoot and static error compared to deterministic settings, and the hybrid twin reduces the forecast MAPE of biomass at the 24-hour horizon from 32.45% to 7.89%, while maintaining a more stable residue structure at medium horizons. The spectral decorrelation of the error with external disturbances, negative Lyapunov exponents in the operating range, and the practical feasibility of the algorithms on modern controllers are noted, which supports their use for quality standardization, optimization of aeration and mixing energy costs, and replication to industrial fermenters with the prospect of integration into the digital infrastructure of the enterprise.