Model Predictive Control — simulating future outcomes over a receding horizon to steer viable cell density, titer, and lactate in real time
Model Predictive Control cannot simulate a future it cannot describe. Before any optimization happens, engineers build and calibrate a mechanistic model of the bioreactor: mass-balance equations for viable cell density, substrate consumption, and product formation, fit against historical batch and perfusion data until the model reproduces observed dynamics within acceptable error. Every prediction MPC will ever make inherits the accuracy of this first step.
A bioreactor process model is a set of coupled ordinary differential equations tracking how state variables change over time as a function of the current state and the applied control action (typically the feed rate, F):
dX/dt = μ(S)·X − D·X (viable cell density) dS/dt = −(μ(S)/Yxs)·X + F·Sf/V − D·S (substrate/glucose) dP/dt = (α·μ(S) + β)·X − D·P (product titer, Luedeking–Piret)
Where μ(S) is the specific growth rate — typically Monod-limited, μ(S) = μmax·S/(Ks+S) — Yxs is the biomass yield on substrate, Sf is feed substrate concentration, D is the dilution rate from feed addition, and α, β are growth- and non-growth-associated product formation constants.
A parallel equation tracks lactate, the key byproduct constraint: overflow metabolism kicks in when glucose is fed faster than the cells can oxidize it, diverting flux to lactate production instead of biomass and product.
Raw kinetic equations are only a skeleton; μmax, Ks, Yxs, α, β and the lactate overflow threshold must be estimated from real data. Engineers run parameter estimation (nonlinear least squares, or Bayesian methods for uncertainty quantification) against 8–20 historical batches spanning different feed strategies, seed densities, and media lots.
The model is validated on batches excluded from fitting: if predicted VCD and titer trajectories track the held-out batch within roughly 10% RMSE across the full run, the model is considered fit for control use. Models that pass calibration but drift on a new cell line or raw-material lot are a recurring operational risk — which is exactly the model-plant mismatch problem stage 3 and 4 confront directly.
A model does not need to be perfect to be useful for control — it needs to be accurate enough, over the short prediction horizon MPC actually uses, that the ranking between candidate control actions stays correct even if absolute values drift slightly.
This calibrated kinetic model does not sit in a spreadsheet — it becomes the computational core of the bioreactor's digital twin, wired directly into the Process Analytical Technology (PAT) stack: at-line and in-line sensors, soft sensors, and the historian feed it fresh calibration data every batch, while the control system calls it thousands of times per control interval to simulate candidate futures. Building this model well is the single highest-leverage investment in the entire MPC deployment — a well-calibrated model with a simple optimizer consistently outperforms a sophisticated optimizer wrapped around a poor model.
A prediction is only as good as its starting point. Before the model can simulate forward, the control system must know precisely where the bioreactor actually is right now: viable cell density, residual glucose, dissolved oxygen, pH, and product titer, fused from in-line probes, at-line assays, and soft-sensor inference into a single, continuously updated current-state vector.
Dissolved oxygen (DO) and pH probes report continuously and are the fastest, most trusted signals in the loop — DO reflects real-time oxygen uptake rate (a proxy for metabolic activity), while pH reflects both metabolic acid/base production and the control system's own base/CO2 additions. Capacitance probes (dielectric spectroscopy) give a near-real-time, non-invasive estimate of viable cell density by measuring the polarizability of intact cell membranes — cells that have lysed no longer contribute capacitance, making it a genuinely viable-selective signal, not just a total-cell-count proxy.
Glucose, lactate, and titer traditionally required at-line sampling (a technician pulls a sample, runs it on a bioanalyzer or HPLC) on the order of once or twice per shift — far too infrequent for a control loop that re-optimizes every 1–4 hours.
Soft sensors close this gap: statistical or model-based estimators that infer slow-sampled or unmeasured variables (glucose, lactate, titer, sometimes even metabolic flux) from the fast continuous signals (DO, pH, capacitance, off-gas CO2/O2) using the same kinetic model built in stage 1, updated recursively with an estimator such as an Extended Kalman Filter or Moving Horizon Estimator.
Every time a new at-line assay result does arrive, it is used to correct the soft-sensor estimate and reduce accumulated drift — the estimator continuously reconciles what the model predicted state to be against what was actually measured, the same reconciliation logic that keeps a receding-horizon controller honest batch after batch.
Every branch simulated in the next stage starts from this single state vector — biased or noisy state estimation propagates directly into every candidate future trajectory, regardless of how good the kinetic model or the optimizer is. This is why serious MPC deployments invest as heavily in state estimation infrastructure (calibrated soft sensors, sensor fusion, outlier rejection) as they do in the optimizer itself: MPC's celebrated foresight is worthless if it is looking forward from the wrong starting point.
With a calibrated model and a trustworthy current state in hand, MPC does what gives it its name: it looks forward. The model is run not once but many times, each run driven by a different candidate sequence of future feed-rate actions, forecasting how viable cell density, titer, and lactate would evolve under each scenario across the prediction horizon.
Simulating literally every possible feed-rate trajectory is intractable, so candidate profiles are parametrized — piecewise-constant feed levels over a handful of future control intervals, or a smooth exponential/step ramp described by 2–3 tunable coefficients. This turns an infinite-dimensional search (an arbitrary function of time) into a finite optimization over a manageable number of decision variables, typically one feed level per control interval within the horizon.
For each candidate profile, the stage-1 kinetic model is integrated forward from the stage-2 current state, step by step, producing a full predicted trajectory — not just a final number, but the entire path VCD, titer, and lactate would take under that feed policy.
A longer prediction horizon lets the controller see further ahead — anticipating, for example, that an aggressive early feed rate will exhaust dissolved oxygen capacity two days out, and moderating now rather than reacting only after the problem appears. But every additional hour of horizon compounds model uncertainty: small kinetic errors integrate over more time, and the number of feed intervals to search grows, increasing computational cost per cycle.
Short horizons react well to what is already happening but can be myopic, walking into constraint violations a longer view would have avoided. Practical bioprocess MPC deployments typically use horizons long enough to span the slower dynamics of interest (feed-limitation and byproduct accumulation) while staying short enough that the model remains trustworthy — commonly a fraction of total batch duration, re-evaluated every control interval.
Horizon length is not a free foresight dial. Every extra hour of prediction trades reduced short-term reactivity and increased computational load against greater anticipation of slow-developing constraint violations — the right value depends on how fast the real process dynamics and the real model errors both evolve.
The branching trajectories fanning out from the current point are not decoration — each dashed line is a genuinely distinct future the model has simulated, corresponding to a specific feed-rate policy the optimizer is considering. Some branches feed aggressively and reach high predicted titer but cross the lactate limit; others feed conservatively and stay safely within constraints but under-deliver on titer. The optimizer's job, taken up in the next stage, is to search this fan of futures for the one that best satisfies the objective without violating any constraint.
Simulating candidate futures is only half of MPC — the other half is optimization: scoring every predicted trajectory against an explicit objective function and a set of hard constraints, selecting the sequence that wins, and then implementing only its very first control action. This single discipline — plan the whole sequence, act on just the first step — is what separates MPC from a simple lookup-table controller.
The optimizer minimizes (or maximizes) a cost function that encodes what "good" means for this batch, typically a weighted combination of terms: maximize predicted final titer, minimize deviation from a target VCD trajectory, penalize aggressive feed-rate changes (to protect actuators and avoid oscillation), and penalize proximity to constraint boundaries even before they are crossed. This multi-objective formulation is precisely why MPC handles competing goals — titer versus byproduct accumulation versus oxygen demand — more gracefully than a single-loop feedback controller ever could: all objectives are weighed simultaneously across the whole predicted horizon, not reactively one at a time.
Alongside the objective, the optimizer enforces hard constraints: lactate must stay below its inhibitory threshold, dissolved oxygen must stay above the level that would trigger hypoxic stress, feed pumps cannot exceed their physical flow limits. Formally, this is a constrained optimization problem (often a nonlinear or quadratic program) solved numerically at every control cycle: minimize the cost function over the space of candidate feed policies, subject to the predicted trajectory never leaving the feasible region defined by these constraints.
When every candidate feasible under the constraints has been scored, the branch with the lowest cost (here, effectively the highest safe predicted titer) is declared the winner — turning solid and bright in the visualization while the rejected candidates fade.
Even though the optimizer just computed an entire future feed-rate sequence spanning the whole prediction horizon, the controller deliberately throws most of it away: only the first control interval's feed rate is actually sent to the pump. The rest of the winning sequence was never a commitment — it was scaffolding the optimizer needed to correctly judge the consequences of that first step.
This matters because the further into the future a plan reaches, the less trustworthy it becomes — model error compounds, and real disturbances the model could not anticipate will have appeared before that future arrives. Committing only to the immediate action, then re-planning once new data arrives, is what keeps MPC grounded in reality rather than executing an increasingly stale open-loop script.
Plan the whole horizon, act on only the first step, then immediately re-plan: this receding-horizon discipline is the core mechanism that lets MPC be both far-sighted in its reasoning and disciplined in its execution.
The defining property of Model Predictive Control is not that it predicts — plenty of controllers use forecasts — but that it never stops re-predicting. As soon as new sensor data confirms (or contradicts) what the model expected, the entire predict-optimize-implement cycle repeats: the horizon slides forward in time, a fresh state estimate anchors a new fan-out of candidate futures, and a new optimal action is selected. This is receding horizon control.
Classical feedback control (PID) reacts to the error between a single measured variable and its setpoint — effective for simple, near-linear loops like maintaining DO or pH, but structurally unable to proactively balance multiple competing objectives or anticipate hard constraints before they are violated. A PID loop regulating feed rate to a titer target has no native way to know that doing so is about to push lactate over its limit three hours from now; it will only respond once the violation has already begun.
MPC, by contrast, evaluates the entire predicted trajectory of every relevant variable simultaneously, over a horizon long enough to see the consequence coming, and chooses actions that respect the constraint before it is ever crossed. For a nonlinear, multivariable, constrained process like a bioreactor — where growth kinetics are nonlinear, feed affects multiple outputs at once, and both titer maximization and byproduct avoidance matter — this proactive, whole-horizon reasoning is a categorically different and generally more capable control strategy.
MPC's entire advantage rests on the model being a reasonably faithful stand-in for the real bioreactor — and no model ever is, exactly. Model-plant mismatch arises from unmodeled dynamics, cell-line variability, raw material lot changes, sensor drift, and simplifications made for tractability. Left unmanaged, mismatch causes the optimizer to consistently over- or under-predict outcomes, degrading control performance and, in the worst case, causing the controller to select an action it believes is safe but that actually violates a constraint in reality.
Receding horizon re-optimization is itself the primary defense: because the cycle repeats every control interval using a freshly measured state (not the model's stale prediction of that state), accumulated model error is continuously corrected rather than allowed to compound across an entire batch. Robust MPC implementations add further safeguards — back-off margins that keep predicted trajectories away from hard limits, online model re-estimation, and ensembles of models reflecting parameter uncertainty.
Solving a constrained optimization over many candidate trajectories, at every control interval, for the entire duration of a multi-day or multi-week bioreactor run, is a nontrivial computational commitment. Each cycle must simulate the kinetic model many times (once per candidate feed profile) and solve a nonlinear or quadratic program to convergence — all comfortably inside the control interval, since the whole point is to act before the next round of data invalidates the plan.
Bioprocess control intervals of one to several hours give real-time optimization far more computational slack than, say, a robotics or chemical-plant MPC loop running every few seconds, which is part of why bioprocess MPC has become practical even with commodity computing hardware, off-the-shelf nonlinear solvers, and models complex enough to be genuinely useful.
Receding horizon MPC represents the culmination of the PAT and soft-sensor infrastructure built in the earlier stages — none of it is useful without a controller capable of consuming it intelligently, and MPC is precisely that consumer. It sits atop calibrated kinetic models, real-time state estimation, and constrained optimization to deliver something a fixed recipe or a single-loop feedback controller cannot: a control strategy that continuously adapts to the batch actually unfolding in the vessel, rather than the batch that was planned on paper months earlier.
MPC does not need to be right about the whole future — it only needs to be right enough about the next control interval, because it will check its work and correct course again before the next interval begins. That single design choice is what turns an imperfect model into a genuinely reliable controller.