Scale-Up and Design of Bioreactor Process Using Machine Learning Technique

Abstract

The increasing demand for biopharmaceutical products has intensified the need for efficient, predictable, and scalable bioprocess development. Stirred-tank bioreactors are central to these processes because they determine the local hydrodynamic and biochemical environments. However, bioreactor process design and scale-up remain challenging because operating conditions optimized at small scale do not necessarily translate to pilot or manufacturing scale. Changes in geometry, agitation, aeration, and liquid volume can generate substantially different local mixing, gas-liquid mass transfer, shear, energy dissipation, and phase distribution patterns. These local environmental variations can affect cell growth, viability, metabolism, and product formation, thereby influencing process productivity and product quality attributes. However, these coupled effects are not accounted for by traditional empirical equation-based scale-up approaches. While computational fluid dynamics (CFD) provides detailed spatially resolved descriptions of these multiphase hydrodynamic phenomena, its high computational cost limits its practical use in design optimization, uncertainty analysis, scale-up studies, and model-based decision-making, where repeated model evaluations are required. This thesis develops physics-informed neural network (PINN) surrogate modeling methods for aerated stirred-tank bioreactors to enable computationally efficient and physically consistent bioprocess prediction, optimization, and scale-up. The central hypothesis is that PINNs can serve as fast and physically consistent surrogates for bioreactor-relevant systems when important challenges such as their ability to describe two-phase flow and to address rotating fluid geometries can be tackled. Two related research streams are advanced: i) sensitivity-guided PINN methodologies for improved training efficiency, robustness, and sensor-informed data acquisition; ii) coupled hydrodynamic-biokinetic PINN surrogates for fast bioprocess prediction, optimal design, and scale-up under uncertainty. A sensitivity-based adaptive sampling (SBS) framework was developed to improve PINN training efficiency, accuracy, and robustness by reallocating collocation points according to the sensitivity of the training loss to perturbations in sampling locations. Unlike local residual-, gradient-, or sensitivity-based sampling methods, the proposed global sensitivity formulation accounts for coupled effects among collocation points, loss components, boundary conditions, governing equations, and network-weight updates. This provides a more informative criterion for identifying regions that strongly influence the overall training dynamics, leading to improved prediction accuracy in areas with strong spatial variations. The sensitivity principle was further used to guide sensor placement for PINNs under structural uncertainty. In a channel-flow case with uncertain inlet conditions, sensors were selected at high-sensitivity locations, and both the uncertain operating variable and the selected sensor measurements were incorporated as network inputs. This strategy improved generalization to boundary conditions beyond the training range while adding minimal model complexity. A two-phase PINN surrogate was developed for Euler-Euler gas-liquid hydrodynamics in aerated stirred-tank bioreactors and applied to both two-dimensional rotating-vessel and three-dimensional baffled, impeller-driven geometries. By incorporating agitation speed and aeration rate as operating-condition inputs embedded into the physics-informed residual constraints, the model enabled pointwise prediction of pressure, volume fraction, and gas- and liquid-phase velocity fields across operating conditions. These spatially resolved predictions provide access to local mixing, gas holdup, shear exposure, and interphase transport, which are critical for evaluating oxygen transfer, nutrient distribution, cell stress, and process productivity during process design, optimization, and scale-up. Fourier feature mapping improved the representation of high-gradient structures, and neural tangent kernel analysis motivated a partitioned dual-network architecture to balance the training of coupled hydrodynamic variables. Relative to purely data-driven baselines, the PINN surrogate achieved more accurate and physically consistent predictions, reducing the risk of nonphysical surrogate behavior when predicting new operating conditions or coupling with process-level biokinetic models. Coupling this hydrodynamic surrogate with dynamic transport-reaction biokinetic equations for biomass, substrate, and dissolved oxygen enabled fast process-level prediction at substantially reduced computational cost. The coupled PINN-biokinetic surrogate was embedded into a three-level Optimal Experimental Design for Optimal Process Design (OED-OPD) framework for robust bioreactor optimization under parametric uncertainty. The outer level selects informative experimental conditions for parameter identification, the middle level identifies worst-case parameter scenarios, and the inner level determines optimal process operating conditions. This structure integrates model-based experimental design and robust process optimization into a unified framework, while the PINN surrogate makes the nested optimization computationally tractable compared with direct CFD-based biokinetic simulations. The framework was demonstrated for single-scale optimal design in a 50 L bioreactor and for scale-up from a 250 mL vessel to a 50 L bioreactor. The optimized operating conditions showed close agreement with CFD-based benchmarks, and CFD validation confirmed that the genetic-algorithm solutions improved the objective-function values relative to neighboring candidate designs. The framework also selected different informative experiments at the 250 mL and 50 L scales, showing that optimal experimental designs are scale-dependent. These results demonstrate that the proposed OED-OPD framework integrates model-base experimental design for parameter identification with process optimization, enabling robust and computationally efficient bioreactor design under uncertainty. To summarize, this research establishes a new physics-informed surrogate modeling workflow for bioreactor applications by tackling different challenges related to the development of the model for rotating two-phase fluid conditions. By reducing reliance on repeated high-fidelity CFD simulations while preserving governing-equation structure and physical consistency, the proposed methods provide a foundation for surrogate-assisted bioreactor optimization.