Bioprocess & Scale-Up

Bioprocess control and scale-up for cellular agriculture are inherently complex, dynamic, and non-linear: many interacting parameters — feed rates, dissolved oxygen, pH, temperature, shear, mixing — jointly determine yield, and they must be balanced against scalability, cost, and environmental sustainability if cultivated-cell production is to reach commercial scale. Accurate models and adaptive control strategies are central to that effort, and AI offers several routes to both. A caveat frames the literature below: much of the foundational work was done in microbial and conventional fermentation rather than mammalian cell culture, so several of these papers are analogous to cell-ag rather than native to it — but the methods (surrogate modeling, small-data prediction, soft sensing, reinforcement-learning control, and autonomous experimentation) transfer directly, and the most recent work is increasingly mammalian.

Real-time monitoring and soft sensors

Closed-loop control needs fast, non-destructive readouts of what is happening inside the reactor. Tamburini et al. (2014, ref #7) built near-infrared spectroscopy (NIRS) soft sensors — both on-line (flow-cell) and in-line (fiber-optic probe) — calibrated with multilinear regression and partial least squares to track glucose, organic acids, and biomass during bacterial fermentations, reaching calibration/validation R² around 0.95–0.99 for the key analytes and feeding those estimates into automated control. NIRS could not discriminate cells by shape (water absorption dominates the spectrum), but that same insensitivity let one calibration transfer across several strains under matched conditions. As a process-analytical-technology pattern, this maps directly onto real-time substrate and biomass monitoring in cultivated-meat bioreactors.

Three recent studies push the same soft-sensing idea into mammalian cell culture with machine learning. Hevaganinge et al. (2023, ref #206) trained an interpretable convolutional “metabolite regression” model on contactless short-wave-infrared hyperspectral images of cell-free spent media from CHO cultures, predicting glucose and lactate at test RMSEs (~27 and ~20 mg/dL) competitive with submerged Raman probes — a fouling-free optical sensor for closed-loop feedback. Behdani et al. (2024, ref #207) built a random-forest soft sensor that predicts total sialic acid — a glycosylation quality attribute — from routine process parameters (viable-cell density, glutamate, ammonium, phosphate, basal-medium type) across multiple CHO scales, replacing a multi-hour wet assay with an inference behind a dashboard. Xu et al. (2025, ref #208) benchmarked ensemble regressors (random forest, XGBoost, CatBoost) against partial least squares for online viable-cell-density and viability monitoring, finding a random forest improved viability accuracy by roughly 32% over PLS (R² 0.58 → 0.81) and that Raman outperformed capacitance spectroscopy for viability monitoring (while capacitance was the better signal for viable-cell density). All three are mammalian and transfer directly to cultivated-cell process-analytical technology, though each is bounded to its instrument and cell line.

Surrogate modeling and CFD acceleration for reactor design

Computational fluid dynamics captures the mixing, shear, and mass-transfer behavior that governs how a bioreactor scales, but high-fidelity CFD is far too slow to sit inside a design loop. Del Rio-Chanona et al. (2019, ref #29) addressed this by training a convolutional-neural-network surrogate on a coupled CFD-plus-kinetics model of a photobioreactor, then searching the surrogate with a hybrid stochastic optimizer to co-optimize reactor geometry and operating conditions; at the verified optimum the surrogate reproduced the integrated model’s biomass and biofuel predictions to within about 1%, and its one-time training (~2 hours) replaced the roughly six days of CFD each configuration would otherwise take, collapsing a months-long design study to about a day. Rojek et al. (2021, ref #33) work the same theme from the computing side, presenting AI-accelerated CFD simulation built on OpenFOAM with combined CPU/GPU execution. Together they illustrate the central move for reactor design under cell-ag constraints: replace the expensive physics solver with a cheap learned emulator so that mixing and transport can be optimized at scale.

Predicting and controlling fermentation with limited data

Process datasets in this field are typically small, which shapes the methods that work. Peng et al. (2013, ref #30) paired an artificial neural network (with fermentation time as an explicit input) with a genetic algorithm to derive a time-dependent control trajectory for bacteriocin production, improving titre by 26% over a fixed protocol and identifying dissolved oxygen as the dominant parameter. Zhang et al. (2020, ref #31) found that a support-vector-machine model generalized better than a back-propagation neural network on a small lipid-fermentation dataset, then used a genetic algorithm on the SVM to maximize biomass, at which optimum the lipid yield also peaked. Roell et al. (2022, ref #32) made the small-data lesson explicit: across seven ML families benchmarked on syngas fermentation — with a key reframing that pairs current conditions with instantaneous production rates rather than time — random forests and SVMs generalized best to unseen conditions (test R² ≈ 0.85) while neural networks overfit, and the model’s feature importances pointed toward model-predictive control. The recurring pattern — classical ML over deep nets when data is scarce, GA or MPC for the control layer — is directly applicable to cultivated-cell process development.

Two further studies extend this small-data prediction pattern. Takahashi et al. (2016, ref #209) used a genetic algorithm to optimize an artificial neural network’s architecture from uniform-design experiments describing rabies-virus production in BHK-21 cells, reaching a held-out test correlation coefficient of 0.99 on a multilevel design — a mammalian-cell GA-plus-ANN-plus-DoE template for modeling production from few runs (a viral-vaccine endpoint here rather than a cultivated-meat one). On the microbial side, Du et al. (2025, ref #27) trained a two-stage gradient-boosting pipeline to predict volatile-compound profiles in Saccharomyces cerevisiae fermentation simulating canned-meat aroma — a precision-fermentation prediction task that bridges to Sensory Prediction.

Reinforcement learning for bioprocess control

A distinct and fast-growing thread treats feeding, temperature, and setpoint control as a sequential decision problem and solves it with reinforcement learning (RL) — learning a control policy by trial and error against a process model rather than tuning a fixed controller. Petsagkourakis et al. (2020, ref #200) is the foundational demonstration: a policy-gradient method with a recurrent-network policy, pre-trained on a preliminary model and then refined batch-to-batch on the “true” plant, outperformed nonlinear model-predictive control on simulated microbial batch fermentations under model mismatch. Subsequent work hybridizes RL with classical control and broadens the objectives: Oh et al. (2022, ref #201) set a model-predictive controller’s terminal cost from a learned action-value function (a double-deep-Q-network), cutting operating cost on a simulated penicillin semi-batch process with less data than model-free RL; Rajasekhar et al. (2024, ref #202) used a twin-delayed DDPG agent for setpoint-tracking and disturbance-rejecting temperature control of a simulated ethanol bioreactor, beating DDPG on tracking error; and Wu & Cui (2025, ref #203) applied PPO to multi-objective fed-batch lysine control, recovering a better-spread Pareto set of yield-versus-productivity trade-offs in far fewer iterations than weighted-sum and normalized-normal-constraint baselines. Every one of these is a simulation study on a microbial or model system rather than mammalian cell culture, so the caveat is sharp — but the methodological transfer to cultivated-cell feeding and environmental control under uncertainty is direct, and RL is the control-layer counterpart to the surrogate and soft-sensor work above.

Mechanistic models and digital twins

A complementary route keeps the mechanism explicit. Antonakoudis & Richelle (2026, ref #59) developed a data-driven pipeline that reduces a genome-scale metabolic model into a compact, dynamic form suitable for bioprocess simulation, using Bayesian flux estimation to propagate the uncertainty of sparse, noisy exo-metabolomics data into the reduction. Applied to time-course data from twelve fed-batch CHO cultures, a single reduced model remained feasible across all conditions and reproduced the observed extracellular fluxes — an uncertainty-aware foundation for the kind of digital twin that mammalian cultivated-cell bioprocesses will need. Because CHO is the canonical mammalian production host, this is among the most directly cell-ag-relevant entries here.

Two hybrid mechanistic-plus-ML studies sit close to this digital-twin goal and are squarely mammalian. Chiu & Du (2025, ref #204) built a neural-ordinary-differential-equation model-predictive controller to regulate glucose in a simulated fed-batch CHO bioreactor, reporting lower control error and smaller model–plant mismatch than neural-network or LSTM MPC at higher computational efficiency. Yang et al. (2024, ref #205) fitted a physics-informed neural network — first-principle kinetics with neural-network-parametrized terms, recalibrated daily — to real 750-litre pilot CHO fed-batch data, extrapolating better than purely data-driven models and degrading gracefully as instrumentation is removed. Both are the hybrid, uncertainty-aware control and modeling layer a mammalian cultivated-cell digital twin will be built from.

Autonomous experimentation and robot scientists

The newest line closes the experimental loop with agents that plan, execute, and inspect cell-culture work. Wang et al. (2025, ref #61) describe Agentic Lab, a multi-agent, human-in-the-loop system in which a virtual-PI agent delegates to subagents for literature retrieval, protocol design, data analysis, and training-free phenotyping, with an augmented-reality module giving operators real-time, context-aware guidance; the authors demonstrate it on a multi-week protocol differentiating human pancreatic organoids. Qiu et al. (2025, ref #62) present BioMARS, a fully-robotic counterpart in which Biologist, Technician, and Inspector agents (built on LLMs and vision-language models, driving dual-arm robotics) autonomously perform cell passaging and culture, matching or exceeding manual performance on viability and morphology and outperforming conventional strategies when optimizing retinal-pigment-epithelial differentiation. Both point at the same destination — autonomous, reproducible cell-culture operations — and differ mainly in how much human oversight stays in the loop.

Tools and data

CFD and reactor simulation: OpenFOAM and the broader Bioprocess Modeling & Scaling section in Software.md — the physics solvers that surrogate models like ref #29’s accelerate.
Metabolic modeling: the Metabolic Modeling & Strain Design tooling in Software.md and the Pathways, Metabolism & Metabolic Models databases (e.g. BiGG Models, BioModels) underlying the genome-scale work in ref #59; CHO-specific model files are inventoried in Datasets/CHOReference.md.
Agentic lab software: the AI Agents & Foundation Models section in Software.md for the agent frameworks adjacent to refs #61–#62.

Open challenges for cell-ag

The strongest results sit at the two ends of the spectrum — well-characterized microbial fermentation on one side, and very recent mammalian-cell systems on the other — with comparatively little validated work on mammalian bioreactor control at scale in between. The reinforcement-learning control work (refs #200–#203) is promising but still entirely in simulation on microbial or model systems, and the mammalian monitoring and hybrid-modeling advances (refs #204–#208) come largely from CHO biopharma rather than food-cell production. Three gaps stand out. First, data scarcity is structural: cultivated-cell process runs are slow and costly, so the field will keep favoring small-data methods (kernel models, surrogates, transfer learning) and benefit from shared, standardized process datasets, which today barely exist. Second, the surrogate and digital-twin work needs to be grounded in mammalian-cell physiology and food-relevant endpoints rather than borrowed wholesale from microbial or algal systems. Third, the autonomous-lab demonstrations are still proofs of concept; turning them into reproducible, auditable cultivated-cell manufacturing will require integration with the monitoring, modeling, and control layers above. A monitoring-to-control stack that fuses soft-sensor readouts, a reduced metabolic model, and an autonomous experimentation loop on a mammalian cultivated-cell line is the open opportunity these papers collectively point toward.