Skip to content

A Hybrid Physics-Informed Neural Network (PINN) for the Electro-Oxidation of 2-Chlorophenol on BDD Electrodes in a Flow-By Reactor Under Batch Recirculation

FieldValue
JournalProcesses
Published2026-06-09
AuthorsAlejandro Regalado-Méndez, Damayrí M. Salinas-Camacho, Reyna Natividad, Mario E. Cordero, Luis Gonzaga Alonso Zárate
DOI10.3390/pr14121862

The hybrid physics-informed neural network reported R² = 0.9927 for 2-chlorophenol concentration at x = 0.2 m and t = 4 h. The corresponding MSE was 0.0009 and RMSE was 0.0294. Training took 7.36 h, while inference averaged 8.0 ms over 1,000 evaluations. The CFD model required 42 days. The six experimental points used for the accuracy figures were all included in training.

The earlier parametric model completed a simulation in 8 s, making it fastest for a single run. The PINN required 7.36 h before its 8.0 ms inference became available. It was therefore faster for repeated evaluations after training.

The authors caution that the CFD and PINN comparison is not strictly equivalent. CFD was a first-principles model, while the PINN combined simulated data, experimental data, and physical constraints.

The system used a 2.5 L continuous stirred tank as the reservoir. A magnetic pump sent solution through a flow meter and a 0.2 m flow-by reactor. Conductive particles were fluidized upward, forming the high-surface-area working electrode. Treated solution returned to the tank, so the liquid volume stayed constant during the batch.

The reactor model coupled convection, axial dispersion, and a pseudo-first-order reaction. Danckwerts boundary conditions fixed inlet curvature and set outlet curvature to zero. The tank used a well-mixed lumped ODE with equal inflow and outflow.

Reference conditions were C0 = 1 mol/m³, u = 0.0947 m/s, Dax = 0.0005 m²/s, and kapp = 1.224 1/h. The batch ran for 4 h at Q = 1 L/min.

The FBR network learned the spatial and temporal concentration field. The CST network learned the tank concentration over time. Automatic differentiation computed the derivatives needed to penalize deviations from each governing equation.

Both networks used four hidden layers, the tanh activation function, and a linear output. The FBR network had 60 neurons per layer and took x and t as inputs. The CST network had 50 neurons per layer and took t as its only input.

The loss combined concentration error with PDE and ODE residuals. It also penalized boundary and initial condition deviations. Automatic differentiation computed the required derivatives. The PDE grid contained 200,000 collocation points, while the ODE grid contained 400. Boundary points numbered 400. The two initial-condition grids contained 2,000 and 50,000 points.

The fixed weights were 1 for the PDE, 10⁴ for the ODE and boundary terms, 10⁶ for the FBR initial condition, and 10⁷ for data and the CST initial condition.

Training used Adam for 20,000 epochs. The listed learning rates were 1 × 10⁻³ initially and 5 × 10⁻⁵. A grid search selected the final hyperparameters. Simulated training data came from the previously validated parametric model. All six experimental points were added to training, with no held-out set. The authors reported stable training with no observed numerical instabilities after using soft boundary constraints, tanh activations, and balanced loss weights.

The reported metrics covered 2-CPh concentration at x = 0.2 m and t = 4 h under batch recirculation.

ModelR²MSERMSEReported computation time
PINN0.99270.00090.02947.36 h training
CFD0.99170.16330.404142 days
Parametric0.98310.03070.17548 s per simulation

At 4 h, the experimental concentration was 0.0001 mol/m³. The PINN gave 0.00268 mol/m³, the CFD model gave 0.00130 mol/m³, and the parametric model gave 0.00943 mol/m³. CFD was closest at the final tail. PINN had the highest reported R² and the lowest MSE and RMSE across the model comparison. The parametric model was fastest but had the largest final-point deviation.

The PINN output was smooth and followed the six experimental points closely. However, the fit does not establish predictive accuracy on unseen conditions because those points were part of training.

The one-at-a-time analysis varied one parameter at a time and ran up to 8,000 epochs. It assessed final 2-CPh concentration at x = 0.2 m and t = 4 h.

FactorTested rangeFinal concentration variationMaximum normalized sensitivity
kapp0.9790 to 1.958 1/h−556.34%14.20
Dax1 × 10⁻⁴ to 1 × 10⁻³ m²/s−85.67%3.32
u0.05 to 0.15 m/s−290.13%6.20
Q0.5 to 1.5 L/min−32.92%6.55

The apparent kinetic constant ranked first in sensitivity. Linear velocity and volumetric flow rate were similar, while axial dispersion ranked last. The authors state that kapp depends on current intensity, so its estimation and control are central. Accurate measurement of kapp matters more for prediction than changes to Dax within the tested range.

A central limitation is the validation design. No independent experimental test set was reported. The authors call for data at other operating conditions or time points to assess generalization. Since simulated training data came from a validated parametric model, the PINN can also inherit that model’s assumptions and errors. The one-at-a-time analysis does not quantify interactions between parameters.

The electrochemical model uses pseudo-first-order kinetics with an apparent kapp. It does not resolve Butler-Volmer kinetics or the current distribution. The authors propose a Butler-Volmer mass balance and current distribution model for future work. They also suggest expressing the pseudo-kinetics in terms of current intensity.

Training used fixed loss weights selected through trial and error. Adaptive weighting was not tested. Loss reached its minimum within about 10,000 to 18,000 epochs, then increased toward 20,000 epochs. Early stopping is therefore required for stable results.

The reported stability applies to the tested parameter ranges. The paper says a 1% change in u leads to a 6.20% output change, yet it reports ±0.62% for a ±10% flow fluctuation. It assigns Q a separate sensitivity of 6.55. These statements do not align, so the disturbance estimates need clarification before control use.

Implications for BDD electrode engineering

Section titled “Implications for BDD electrode engineering”

The 8.0 ms inference time makes repeated evaluation practical for process monitoring. The authors suggest use in digital twins and real-time control, but no controller or digital twin was tested. The model therefore supports surrogate calculation, not a claim of closed-loop performance.

For the reactor, current intensity and the resulting kapp deserve priority measurement. Flow rate, linear velocity, and residence time also affect the decay curve. The low-concentration tail needs particular attention because the 4 h PINN value was above the experimental value. A control target near that concentration should therefore use a calibrated effluent measurement rather than the PINN value alone.

The model does not compare BDD material attributes. It reports no results for boron concentration, crystallinity, crystal orientation, surface roughness, electrode area, substrate, thickness, or metallization. Its R² cannot rank BDD builds. The paper cites high hydroxyl radical generation as the basis for using BDD in water treatment. An electrode build should therefore be qualified with its own effluent decay data under the intended current and flow conditions.

The studied reactor uses boron-doped diamond electrodes for 2-CPh electro-oxidation, with conductive particles fluidized as the working bed. 6CCVD’s diamond materials catalog lists boron-doped diamond, which matches the paper’s electrode material. The paper gives no diamond grade, thickness, substrate, or metallization, so those details need separate definition through engineering support.

View Original Abstract

The electro-oxidation of persistent organic pollutants such as 2-chlorophenol (2-CPh) using boron-doped diamond (BDD) electrodes offers a promising wastewater treatment route, yet conventional mechanistic models (e.g., CFD) suffer from prohibitive computational costs. This study develops a hybrid physics-informed neural network (PINN) to model the electro-oxidation of 2-CPh in a flow-by reactor coupled with a continuous stirred tank under batch recirculation mode. The PINN integrates a diffusion-convection partial differential equation with a lumped-parameter ordinary differential equation for the tank, embedding physical constraints directly into the loss function. The model was trained on simulated data generated from a previously validated parametric model and optimized using a systematic hyperparameter grid search. The PINN achieved excellent agreement with experimental data, yielding a coefficient of determination (R2) of 0.9927, a mean square error of 0.0009, and a root mean square error of 0.0294—outperforming both the CFD and parametric models in accuracy. Sensitivity analysis revealed that the apparent kinetic constant is the most influential parameter (normalized sensitivity of 14.20). While the CFD model required 42 days and the parametric model 8 s, the PINN achieved a balanced trade-off with a runtime of 7.36 h. We conclude that the PINN provides a highly accurate, computationally feasible surrogate model suitable for integration into digital twins and real-time control frameworks for electrochemical wastewater treatment.

  1. 2003 - Electrooxidation of 2-Chlorophenol on PolyNiTSPc-Modified Glassy Carbon Electrodes [Crossref]
  2. 1979 - ES&T Special Report: Priority Pollutants: I-a Perspective View [Crossref]
  3. 2012 - Electrochemical Degradation of Phenol and 2-Chlorophenol Using Pt/Ti and Boron-Doped Diamond Electrodes [Crossref]
  4. 2020 - Electro-Oxidation of 2-Chlorophenol with BDD Electrodes in a Continuous Flow Electrochemical Reactor [Crossref]
  5. 2005 - Anodic Degradation of Toxic Aromatic Compound in the Flow Through Cell with Carbon Fiber Electrode [Crossref]
  6. 1999 - On the Performance of Ti/SnO2 and Ti/PbO2 Anodes in Electrochemical Degradation of 2-Chlorophenol for Wastewater Treatment [Crossref]
  7. 2018 - Preparation of the CNTs/AG/ITO Electrode with High Electro-Catalytic Activity for 2-Chlorophenol Degradation and the Potential Risks from Intermediates [Crossref]
  8. 2021 - Electrochemical Reactors Equipped with BDD Electrodes: Geometrical Aspects and Applications in Water Treatment [Crossref]
  9. 2020 - Characterization of the Reaction Environment in Flow Reactors Fitted with BDD Electrodes for Use in Electrochemical Advanced Oxidation Processes: A Critical Review [Crossref]
  10. 2024 - Recent Advances in Electrochemical Flow Reactors Used in Advanced Oxidation Processes: A Critical Review [Crossref]