The growing global population and heightened concern for climate change leads to increased interest in utilizing microbial fermentations to replace polluting production processes for e.g., plastics, fuels, and animal proteins. Computational fluid dynamics (CFD) is a valuable tool for accelerating the scale-up and optimization of large-scale bioprocesses. However, the design correlations underlying most of these CFD models are validated with air-water systems, not accounting for the distinct hydrodynamic properties of microbial fermentation broth. In this review, we provide an extensive overview of the current understanding of how various biotechnologically relevant solutes impact the hydrodynamics of bubble columns. We examine the effects of components found in fermentation broths, including salts, surfactants, viscoelastic solutes, alcohols, acids, ketones, sugars, biomass, and proteins, on mass transfer, bubble formation, bubble interactions, and flow regime transitions. These components all exhibit unique effects, yet their combined influences remain poorly understood. Future research should prioritize identifying the concentration at which coalescence inhibition occurs for different compounds, especially in mixtures, and exploring the role of proteins in bubble column hydrodynamics from micro- to macroscale.

@en

Bioprocesses are scaled up for the production of large product quantities. With larger fermenter volumes, mixing becomes increasingly inefficient and environmental gradients get more prominent than in smaller scales. Environmental gradients have an impact on the microorganism's metabolism, which makes the prediction of large-scale performance difficult and can lead to scale-up failure. A promising approach for improved understanding and estimation of dynamics of microbial populations in large-scale bioprocesses is the analysis of microbial lifelines. The lifeline of a microbe in a bioprocess is the experience of environmental gradients from a cell's perspective, which can be described as a time series of position, environment and intracellular condition. Currently, lifelines are predominantly determined using models with computational fluid dynamics, but new technical developments in flow-following sensor particles and microfluidic single-cell cultivation open the door to a more interdisciplinary concept. We critically review the current concepts and challenges in lifeline determination and application of lifeline analysis, as well as strategies for the integration of these techniques into bioprocess development. Lifelines can contribute to a successful scale-up by guiding scale-down experiments and identifying strain engineering targets or bioreactor optimisations.

@en

An optical fibre probe was tested in liquids with biotechnologically relevant compounds, to determine if any limits arise in bubble property determination due to these compounds. The data from this fibre probe was compared to the data from a camera for which an image processing algorithm was developed. This algorithm was able to successfully filter objects using configurable parameters. Objects had to be in focus, within a certain size range and elliptical in shape. The filtering was able to work with partially out of focus bubbles, and bubbles at the edges of the image within tolerances specified by the parameters. To provide insight into the relation between the fibre probe data and camera data a control measurement was performed in pure water, after which measurements were performed in a mixture containing 100g/L ethanol and a mixture containing 0.4M NaCl. The data sets from both measurement methods were compared for bubble velocity and bubble size. The velocity would be directly comparable,
the size was however the local height at the point where the fibre probe pierced the bubbles. As such the height directly below the fibre probe of the bubbles was determined from the images to serve as comparable data. The comparison between the probe data and camera data was inconclusive due to a dissimilarity between the distributions. Mainly the overestimation of the comparable size from the camera data was a problem. Which was likely due to the camera showing only a 2D projection of the bubble and as such the entire height would be used for this comparison. For the velocity the camera data had a narrower distribution as compared to the fibre probe data. This could be caused by the camera averaging the speed of bubbles between two or even three frames if a bubble was not detected in the middle frame, the probe might be more sensitive to these variations and as such yield a wider distribution. The image processing algorithm showed a correct processing of the images to yield usable data, the exact comparable data however still needs to be improved for this specific application.

The traditional Eulerian view of biomass in bioprocess modelling results in issues when modelling large-scale bioreactors in which heterogeneous conditions are common. As the field increasingly moves from “scale-up” to “scale-down” philosophy, in which such heterogeneities are included from the start of process design, accurate modelling of the biomass response to these varying conditions is essential. Euler-Lagrange (EL) simulations provide a means of modelling the microbial lifelines of cells traversing heterogenous conditions of a bioreactor. Lapin et al. were the pioneers of EL simulation in their 2004 paper where a metabolic model of glycolysis is coupled to a Lagrangian biomass phase. This BSc thesis focusses on reproducing their model using modern computational fluid dynamics (CFD) techniques. Specifically, by using a dynamic Lattice Boltzmann Method using Large Eddy Simulation model for CFD as opposed to a frozen-flow Finite Volume Reynolds Averaged Navier- Stokes model. The resulting differences in the overall behaviour of the cell metabolism through the lens of glycolytic oscillations are discussed. In addition, possible pitfalls in model validity such as grid dependence, the effects of heterogeneous particle distributions and the effects of particle numbers were explored. The synchronisation and desynchronisation of glycolytic oscillations as observed in Lapin et al. 2004 were able to be reproduced.

Scaling up bioprocesses is one of the most crucial steps in the commercialization of bioproducts. While it is known that concentration and shear rate gradients occur at larger scales, it is often too risky, if feasible at all, to conduct validation experiments at such scales. Using computational fluid dynamics equipped with mechanistic biochemical engineering knowledge of the process, it is possible to simulate such gradients. In this work, concentration profiles for the by-products of baker’s yeast production are investigated. By applying a mechanistic black-box model, concentration heterogeneities for oxygen, glucose, ethanol, and carbon dioxide are evaluated. The results suggest that, although at low concentrations, ethanol is consumed in more than 90% of the tank volume, which prevents cell starvation, even when glucose is virtually depleted. Moreover, long exposure to high dissolved carbon dioxide levels is predicted. Two biomass concentrations, i.e., 10 and 25 g/L, are considered where, in the former, ethanol production is solely because of overflow metabolism while, in the latter, 10% of the ethanol formation is due to dissolved oxygen limitation. This method facilitates the prediction of the living conditions of the microorganism and its utilization to address the limitations via change of strain or bioreactor design or operation conditions. The outcome can also be of value to design a representative scale-down reactor to facilitate strain studies.

@en

We assess the effect of substrate heterogeneity on the metabolic response of P. chrysogenum in industrial bioreactors via the coupling of a 9-pool metabolic model with Euler-Lagrange CFD simulations. In this work, we outline how this coupled hydrodynamic-metabolic modeling can be utilized in 5 steps. (1) A model response study with a fixed spatial extra-cellular glucose concentration gradient, which reveals a drop in penicillin production rate q_p of 18–50% for the simulated reactor, depending on model setup. (2) CFD-based scale-down design, where we design a 1-vessel scale down simulator based on the organism lifelines. (3) Scale-down verification, numerically comparing the model response in the proposed scale-down simulator with large-scale CFD response. (4) Reactor design optimization, reducing the drop in penicillin production by a change of feed location. (5) Long-term fed-batch simulation, where we verify model predictions against experimental data, and discuss population heterogeneity. Overall, these steps present a coupled hydrodynamic-metabolic approach towards bioreactor evaluation, scale-down and optimization.

@en

CFD simulations of mixing in single-phase multi-Rushton stirred tanks based on the RANS methodology frequently show an over-prediction of the mixing time. This hints at an under-prediction of the mass exchange between the compartments formed around the individual impellers. Some studies recommend tuning the turbulent Schmidt number to address this issue, but this appears to be an ad-hoc correction rather than physical adjustment, thereby compromising the predictive value of the method. In this work, we study the flow profile in between two Rushton impellers in stirred tank. The data hints at the presence of macro-instabilities, and a peak in turbulent kinetic energy in the region of convergent flow, which both may promote inter-compartment mass exchange. CFD studies using the steady-state multiple reference frame model (unsteady simulations are treated in part II) inherently fail to include the macro-instability, and underestimate the turbulent kinetic energy, thereby strongly over-estimating mixing time. Furthermore, the results are highly mesh-sensitive, with increasing mesh density leading to a poorer prediction of the mixing time. Despite proper results for 1-impeller studies, we do not deem MRF-RANS models suitable for mixing studies in multi-impeller geometries.

@en

Steady state multiple reference frame-RANS (MRF-RANS) simulations frequently show strong over-predictions of the mixing time in single-phase, multi-impeller mixing tanks, which is sometimes patched by ad hoc tuning of the turbulent Schmidt-number. In Part I of this work, we experimentally revealed the presence of macro-instabilities in the region between the impellers, as well as a peak in the turbulent kinetic energy in the region where the flow from the individual impellers converges. The MRF-RANS method was found unable to capture both. In this second paper, we show that the sliding-mesh RANS (SM-RANS) approach does capture the effect of macro-instabilities, while still underestimating the turbulent kinetic energy. Consequently, the SM-RANS method mildly over-estimates the mixing time, while being less sensitive to the exact mesh geometry. Large eddy simulations with the dynamic Smagorinsky model reasonably capture the kinetic energy contained in macro-instabilities, and properly assess the turbulent kinetic energy in the region between the impellers, even for crude meshes. Consequently, the mixing time is reasonably assessed, and even under-predicted at the crudest meshes. However, the turbulent kinetic energy and energy dissipation in the impeller discharge stream are poorly assessed by the dynamic Smagorinsky model.

@en

While various multiphase flow simulation techniques have found acceptance as predictive tools for processes involving immiscible fluids, none of them can be considered universally applicable. Focusing on accurate simulation of liquid-liquid emulsions at the scale of droplets, we present a comparative assessment of the single-component multiphase pseudopotential lattice Boltzmann method (PP-LB, classical and modified) and the Volume of Fluid method (VOF, classical and modified), highlighting particular strengths and weaknesses of these techniques. We show that a modified LB model produces spurious velocities 1–3 orders of magnitude lower than all VOF models tested, and find that LB is roughly 10 times faster in computation time, while VOF is more versatile. Simulating falling liquid droplets, a realistic problem, we find that despite identical setups, results can vary with the technique in certain flow regimes. At lower Reynolds numbers, all methods agree reasonably well with experimental values. At higher Reynolds numbers, all methods underpredict the droplet Reynolds number, while being in good agreement with each other. Particular issues regarding LB simulations at low density ratio are emphasized. Finally, we conclude with the applicability of VOF vis-à-vis PP-LB for a general range of multiphase flow problems relevant to myriad applications.@en

Euler–Lagrange computational fluid dynamics simulations offer great potential for the integration of transport dynamics and metabolic dynamics in fermentation systems. Since the seminal work of Lapin et al. [1,2], progress has been made, mainly in the analysis of CFD data and translation to laboratory setup designs. Different large-scale processes require different analysis methods; in this paper we discuss which analysis methods are best suited for given reactor types, by reviewing prior simulation cases as well as introducing new test cases. Furthermore, we address challenges in the translation from Euler–Lagrange simulations to laboratory scale systems, and propose methods to work around these shortcomings. Based on the current state of the art, we propose guidelines for the selection of data analysis methods, and we discuss the design of rational scale-down simulators. We conclude with a brief discussion regarding the requirements and possibilities of next-generation scale-down simulators, such as microfluidic single-cell analysis, and possible ways to approximate cellular lifelines from invasive intra-cellular measurements.

@en

In the present work, by performing chemostat experiments at 400 and 600 RPM, two typical power inputs representative of industrial penicillin fermentation (P/V, 1.00 kW/m³ in more remote zones and 3.83 kW/m³ in the vicinity of the impellers, respectively) were scaled-down to bench-scale bioreactors. It was found that at 400 RPM applied in prolonged glucose-limited chemostat cultures, the previously reported degeneration of penicillin production using an industrial Penicillium chrysogenum strain was virtually absent. To investigate this, the cellular response was studied at flux (stoichiometry), residual glucose, intracellular metabolite and transcript levels. At 600 RPM, 20% more cell lysis was observed and the increased degeneration of penicillin production was accompanied by a 22% larger ATP gap and an unexpected 20-fold decrease in the residual glucose concentration (C_s,out). At the same time, the biomass specific glucose consumption rate (q_s) did not change but the intracellular glucose concentration was about sixfold higher, which indicates a change to a higher affinity glucose transporter at 600 RPM. In addition, power input differences cause differences in the diffusion rates of glucose and the calculated Batchelor diffusion length scale suggests the presence of a glucose diffusion layer at the glucose transporting parts of the hyphae, which was further substantiated by a simple proposed glucose diffusion-uptake model. By analysis of calculated mass action ratios (MARs) and energy consumption, it indicated that at 600 RPM glucose sensing and signal transduction in response to the low C_s,out appear to trigger a gluconeogenic type of metabolic flux rearrangement, a futile cycle through the pentose phosphate pathway (PPP) and a declining redox state of the cytosol. In support of the change in glucose transport and degeneration of penicillin production at 600 RPM, the transcript levels of the putative high-affinity glucose/hexose transporter genes Pc12g02880 and Pc06g01340 increased 3.5- and 3.3-fold, respectively, and those of the pcbC gene encoding isopenicillin N-synthetase (IPNS) were more than twofold lower in the time range of 100–200 hr of the chemostat cultures. Summarizing, changes at power input have unexpected effects on degeneration and glucose transport, and result in significant metabolic rearrangements. These findings are relevant for the industrial production of penicillin, and other fermentations with filamentous microorganisms.

@en

In a 54 m³ large-scale penicillin fermentor, the cells experience substrate gradient cycles at the timescales of global mixing time about 20–40 s. Here, we used an intermittent feeding regime (IFR) and a two-compartment reactor (TCR) to mimic these substrate gradients at laboratory-scale continuous cultures. The IFR was applied to simulate substrate dynamics experienced by the cells at full scale at timescales of tens of seconds to minutes (30 s, 3 min and 6 min), while the TCR was designed to simulate substrate gradients at an applied mean residence time ((Formula presented.)) of 6 min. A biological systems analysis of the response of an industrial high-yielding P. chrysogenum strain has been performed in these continuous cultures. Compared to an undisturbed continuous feeding regime in a single reactor, the penicillin productivity (q_PenG) was reduced in all scale-down simulators. The dynamic metabolomics data indicated that in the IFRs, the cells accumulated high levels of the central metabolites during the feast phase to actively cope with external substrate deprivation during the famine phase. In contrast, in the TCR system, the storage pool (e.g. mannitol and arabitol) constituted a large contribution of carbon supply in the non-feed compartment. Further, transcript analysis revealed that all scale-down simulators gave different expression levels of the glucose/hexose transporter genes and the penicillin gene clusters. The results showed that q_PenG did not correlate well with exposure to the substrate regimes (excess, limitation and starvation), but there was a clear inverse relation between q_PenG and the intracellular glucose level.

@en

The broth in industrial scale fermentors may contain significant gradients in, for example, substrate concentration, dissolved oxygen and shear rates. From the perspective of microbes in this fermentor, these gradients translate to temporal variations in their environment that may affect their metabolic response. As a result, there may be differences in process yield between laboratory scale fermentations and their industrial counterpart. Rather than scaling-up bioprocesses based on equivalence, it is recommended to scale-down: mimic the large-scale environment in lab scale setups, to account for hydrodynamic-metabolic interaction from the start. In this thesis, the use of Euler-Lagrange computational fluid dynamics to capture the large-scale fermentation environment is explored. Lagrangian simulations offer to study processes from the microbial perspective (so-called “lifelines”), and enable coupling of metabolic models describing the response to external variations. With this, it is possible to take the history of the trajectory of the microbe into account, as organisms may not adapt to their surroundings instantaneously. Guidelines for the setup of fermentor simulations are presented, and several means for processing the lifelines are discussed. The obtained information is used to design lab-scale fermentations that mimic large-scale conditions. It is furthermore shown how coupled hydrodynamic-metabolic simulations can be used to predict yield-loss, assess process improvements, and study the onset of population heterogeneity in large-scale fermentors. Additionally, a more fundamental towards the role of the turbulent Schmidt number in multi-impeller mixing is included.@en

Large substrate concentration gradients can exist in chemical or biochemical reactions, resulting from a large circulation time compared to the turnover time of substrates. The influence of such gradients on the microbial metabolism can significantly compromise optimal bioreactor performance. Lapin et al. (2004) proposed an Euler–Lagrange CFD method to study the impact of such gradients from the microbial point of view. The discrete representation of the biomass phase yields an advantageous perspective for studying the impact of extra-cellular variations on the metabolism, but at significant computational cost. In particular, the tracked number of particles, as well as the applied time resolution, have a large impact on both the accuracy of the simulation and the runtime of the simulation. In this work we study the influence of these parameters on both the simulation results and computation time, and provide guidelines for accurate Euler–Lagrange bioreactor simulations at minimal computational cost.

@en

The trajectories, referred to as lifelines, of individual microorganisms in an industrial scale fermentor under substrate limiting conditions were studied using an Euler-Lagrange computational fluid dynamics approach. The metabolic response to substrate concentration variations along these lifelines provides deep insight in the dynamic environment inside a large-scale fermentor, from the point of view of the microorganisms themselves. We present a novel methodology to evaluate this metabolic response, based on transitions between metabolic “regimes” that can provide a comprehensive statistical insight in the environmental fluctuations experienced by microorganisms inside an industrial bioreactor. These statistics provide the groundwork for the design of representative scale-down simulators, mimicking substrate variations experimentally. To focus on the methodology we use an industrial fermentation of Penicillium chrysogenum in a simplified representation, dealing with only glucose gradients, single-phase hydrodynamics, and assuming no limitation in oxygen supply, but reasonably capturing the relevant timescales. Nevertheless, the methodology provides useful insight in the relation between flow and component fluctuation timescales that are expected to hold in physically more thorough simulations. Microorganisms experience substrate fluctuations at timescales of seconds, in the order of magnitude of the global circulation time. Such rapid fluctuations should be replicated in truly industrially representative scale-down simulators.

@en