The exclusion of ANN-based and adaptive feedback architectures from AFRM®'s control design rests on three independent physical and statistical arguments. The first — that fouling kinetics operate on timescales shorter than any practical feedback latency (Δt < τmech) — is developed in the Technology page and the White Paper. A third — that digitally-driven control architectures currently lack standardised GMP validation protocols — is discussed in the final section of the Perspective article.
This page documents the second, distinct set of arguments: those arising from the statistical structure of closed-loop process data and the safety properties of nominal prediction models. The sources cited — Hassanpour et al. (2022), Krippl et al. (2020), and Oetomo et al. (2025) — are papers that advocate or implement machine learning for membrane process control. Their own documented limitations constitute the analytical basis for the discussion that follows. No claim here goes beyond what the cited authors state directly. The wider question of why a new control paradigm is judged differently from incremental improvements is taken up in Reading AFRM® through Innovation Theory.
The foundational premise of ANN-based Model Predictive Control is that a network trained on historical process data can predict future system states accurately enough to act as a predictive controller. This premise contains a structural assumption: that the data available for training provides adequate coverage of the operating space the controller will encounter in service.
In TFF, this assumption fails for a reason intrinsic to the process architecture, not to any specific implementation. Hassanpour et al. (2022) identify and formalise this failure mode in the context of ANN-based MPC applied to systems operating under closed-loop control.
The paper addresses a system identification scenario that is representative of TFF in normal operation: a nonlinear process running under closed-loop PI control, from which training data is collected. Because the controller continuously drives process outputs toward the setpoint, the resulting dataset covers only a narrow band of the operating space. Inputs and outputs are not independently distributed — they are collinear, because the controller correlates them. The training manifold is structurally thin.
This collinearity has two separable consequences for ANN-based MPC. The first is coverage: when the process deviates from the trained region — as it necessarily will during any atypical fouling event — the ANN is not interpolating within a known distribution. It is extrapolating. The authors document this formally:
"Implementation of an MPC, where the inputs are moved independently, may necessitate predictions by the model in regions where it was not trained, leading to poor closed-loop performance."
— Hassanpour et al. (2022), AIChE Journal, §2The second consequence is more subtle and more dangerous. Even within the trained region, if the MPC optimizer proposes setpoints that are inconsistent with the output collinearity structure of the training data — that is, setpoints that the process could not have reached under the control conditions from which data was collected — the controller attempts to drive the system to a physically inconsistent state. The result is not a smooth prediction error. It is closed-loop instability.
Hassanpour et al. demonstrate this with three RNN models trained on the same dataset, with statistically comparable predictive accuracy on the training distribution (SRMSE 0.78–0.84 across validation and test sets). Under arbitrary setpoints that ignore output collinearity, the three MPCs produce qualitatively different closed-loop responses: one stable, one oscillatory, one unstable. The root cause is not model quality. It is that the setpoints were prescribed without accounting for the correlations inherent in the training data:
"The reason for the oscillatory behavior is the MPC attempt to drive the process outputs to their set-points which are chosen without considering the inherent correlation between the outputs."
— Hassanpour et al. (2022), AIChE Journal, §2.3The proposed remedy — projecting arbitrary setpoints onto the training manifold through PCA-based Squared Prediction Error (SPE) constraints — is itself an acknowledgement of the structural problem. The corrective mechanism requires identifying the collinearity of the training outputs via Principal Component Analysis, computing the upper confidence limit of the SPE statistic, and solving a constrained optimisation to find the nearest achievable setpoint within the training hyperplane. The authors find that even the 95% confidence bound on SPE (SPEₔ = 1.8×10−1) may be insufficient in some scenarios; a tighter constraint (SPE ≤ 10−3) is required to eliminate residual oscillations. The corrective apparatus is substantial, and it is necessary precisely because the training data is inadequate to represent the full operating space the controller must handle.
Xtrain is the training manifold — a low-dimensional subspace of the full process state space, identified by PCA on closed-loop operational data. In TFF under normal control, Xtrain is narrow by construction: the regulator suppresses exactly the process variance that would be needed to expand coverage. Any fouling trajectory departing from the nominal operating band — the events that matter most in GMP manufacturing — lies outside Xtrain, and model predictions in that region carry undefined accuracy.
For TFF specifically, this argument carries additional weight because the process events that a controller is most needed to handle — an atypical lot composition, an elevated lipid fraction, a feed perturbation from upstream processing — are precisely those that deviate from the historical pattern from which the training set was assembled. They are underrepresented by construction in any closed-loop dataset. A controller that fails exactly where deviation from normal operation is largest cannot provide the bounded safety profile that GMP batch manufacturing requires.
Even if the extrapolation problem were solved — even if a sufficiently rich training dataset could be assembled through deliberate open-loop excitation across the full TFF state space — a second structural problem remains. ANN-based MPC controllers operate as nominal models: they produce point predictions of future system states without providing a quantified bound on prediction uncertainty. The MPC optimizer acts on the predicted trajectory with no representation of the probability that the true trajectory diverges from it.
In a chemical reactor context, where an unsuccessful batch can be discarded and the experiment repeated, this is an acceptable engineering trade-off. In GMP biopharmaceutical manufacturing, where a single batch of viral vectors, monoclonal antibodies, or extracellular vesicles represents material that may be irreplaceable within the campaign timeline, the question of what happens when the nominal model is wrong is not a statistical abstraction. It is a product loss event with direct patient and commercial consequences.
Oetomo et al. (2025), in a systematic review of machine learning applications in TFF, address this limitation directly and without qualification.
Oetomo et al. survey the current state of machine learning applied to TFF across concentration, diafiltration, and perfusion applications. Their assessment of fitness for GMP deployment is explicit on the safety and regulatory question:
"ML models are not well-suited for GMP manufacturing settings due to regulatory concerns around interpretability and the lack of mechanistic understanding."
— Oetomo et al. (2025), Journal of Membrane ScienceThe interpretability deficit is framed as a regulatory constraint, not an academic preference. A GMP validation package for an adaptive control system must demonstrate that the control logic is auditable, that its behavior is reproducible under defined conditions, and that deviations from expected behavior are detectable and bounded. A nominal ANN model, whose internal weight representations do not reduce to physical quantities, cannot satisfy these requirements through documentation alone. The validation burden falls on empirical demonstration across the full operating space — precisely the space that closed-loop training data does not cover.
| Architecture | Uncertainty representation | Behavior under off-nominal fouling | GMP validation basis |
|---|---|---|---|
| ANN-based MPC | Nominal only — no prediction bounds | Extrapolation; oscillatory or unstable response documented (Hassanpour et al.) | Interpretability deficit; not well-suited for GMP (Oetomo et al.) |
| Hybrid first-principles + ANN | Partial — ANN fouling component remains unbounded | Depends on ANN accuracy outside training region | Transport layer auditable; fouling residuals uncharacterised |
| AFRM® Deterministic | Bounded by fluid rheology — measurable physical quantities | dR/dt ≈ 0 pre-computed; no feedback latency dependency | IF-THEN physical logic; fully auditable and reproducible |
The recognised limitations of pure ANN approaches have led a research direction toward hybrid architectures: a white-box mechanistic model handles the physically well-characterised components of membrane transport, while a data-driven ANN approximates the residual — typically the fouling resistance term, which is the most process-specific and least tractable from first principles. Krippl et al. (2020) develop this approach for hollow-fiber ultrafiltration and report improved predictive performance relative to either component alone.
The Krippl model decomposes TFF flux behavior into a mechanistic transmembrane pressure component — derived from Darcy-Weisbach and membrane hydraulic permeability — and an empirically fitted fouling resistance term approximated by the ANN. The mechanistic backbone reproduces pressure-flux relationships across conditions with high fidelity; the ANN captures fouling dynamics that the mechanistic backbone cannot represent without empirical data.
The architectural significance of this decomposition lies in what it asserts implicitly. By retaining a white-box model for the transport physics and limiting the ANN to the fouling residual, the authors validate that the physical relationships governing membrane transport — permeability, transmembrane pressure, crossflow velocity — are sufficiently well-characterised analytically to not require empirical learning. The ANN is introduced precisely because fouling dynamics resist this characterisation: they are process-specific, history-dependent, and sensitive to feed variability that falls outside any training dataset assembled under controlled conditions.
The architecture is a response to a correctly identified problem. But the response assumes that fouling must be managed reactively — that the controller will encounter fouling resistance at levels requiring empirical approximation, and must predict and compensate for it in real time.
A hybrid model that retains white-box physics for transport and requires an ANN for fouling implicitly confirms that fouling is the addressable problem. AFRM® addresses it at its origin — before the fouling accumulation that would otherwise require empirical approximation can develop.
This is the white-box paradox of hybrid TFF control architectures. The mechanistic component is retained because it is correct: the physics of pressure-driven membrane transport are well-characterised, scale-invariant, and independently validated. The ANN component is added under the assumption that fouling must be controlled reactively. A deterministic architecture that prevents fouling accumulation from reaching the empirically intractable regime removes the premise on which the ANN component is introduced — while retaining the same physical foundations in full.
The three arguments documented across this page and its companion materials are independent in their logical structure. Each reaches the same architectural conclusion by a different route; none requires the others to hold.
The first argument — from control latency — establishes that adaptive feedback cannot respond to fouling events fast enough to prevent irreversible compaction in fast-fouling biofluids. The relevant timescale is τmech; the feedback latency Δt of any sensor-actuator loop exceeds it in the operating conditions that matter. τmech is not a fitted parameter: it is derived directly from the blocking law regime of the feed class, as formalised by Duclos-Orsello et al. (2006) and confirmed at industrial scale by Dhingra et al. (2026) — the same physical foundations documented on the Literature & Validation page. This argument is developed in full in the Technology page and White Paper.
The second argument — from data structure — establishes that the training data available for an ANN-based controller is structurally narrow, because closed-loop operation suppresses exactly the process variance that would be required to cover the operating space. The induced output collinearity produces oscillatory or unstable MPC behavior under novel operating conditions. Hassanpour et al. (2022) document this mechanism formally, quantify its effect across three RNN models with comparable validation accuracy, and propose a mitigation whose complexity is itself evidence of the structural inadequacy of closed-loop training data.
The third argument — from regulatory structure — establishes that nominal ANN models cannot satisfy GMP interpretability requirements without a physically grounded rationale for the control logic. This constraint is acknowledged directly in the membrane processing literature by Oetomo et al. (2025).
The convergence of these arguments reflects a common physical reality: fouling dynamics in TFF are fast relative to any practical feedback loop, physically tractable below the compaction threshold, and process-specific above it. A control architecture that exploits the tractable regime — by pre-computing the intervention schedule from fluid-class parameters before the batch begins — avoids all three failure modes simultaneously, without requiring historical data, real-time sensor input, or post-hoc interpretability justification. This is the design principle of AFRM®.
| Argument | Physical / statistical basis | Key source | AFRM® response |
|---|---|---|---|
| 1 — Latency | Δt > τmech in fast-fouling biofluids | Technology page · White Paper | Pre-computed schedule; feedback loop eliminated |
| 2 — Data structure | Closed-loop data is collinear; ANN extrapolates under deviation; MPC destabilises | Hassanpour et al. (2022) | No training region; parameters derived from blocking law physics |
| 3 — GMP interpretability | Nominal ANN models are not GMP-auditable; lack mechanistic understanding | Oetomo et al. (2025) | IF-THEN physical logic; fully documentable and reproducible |
The physical timing argument (Δt vs. τmech) and the kinetic foundations of the deterministic framework are developed in the Technology page and the White Paper.
The absence of standardised GMP validation protocols for digital control architectures, as documented by Ghasemzadeh et al. (2026), is discussed in the final section of the Perspective article.
Independent peer-reviewed literature supporting the physical foundations of the deterministic framework — including Duclos-Orsello et al. (2006), Dhingra et al. (2026), and supporting experimental sources — is catalogued on the Literature & Validation page.