Hybrid Physics-ML Model for Forward Osmosis Flux with Complete Uncertainty Quantification

Forward Osmosis (FO) is a promising low-energy membrane separation technology, but challenges in accurately modelling its water flux (Jw) persist due to complex internal mass transfer phenomena. Traditional mechanistic models struggle with empirical parameter variability, while purely data-driven models lack physical consistency and rigorous uncertainty quantification (UQ). This study introduces a novel Robust Hybrid Physics-ML framework employing Gaussian Process Regression (GPR) for highly accurate, uncertainty-aware Jw prediction. The core innovation lies in training the GPR on the residual error between the detailed, non-linear FO physical model prediction (Jw_physical) and the experimental water flux (Jw_actual). Crucially, we implement a full UQ methodology by decomposing the total predictive variance (sigma2_total) into model uncertainty (epistemic, from GPR’s posterior variance) and input uncertainty (aleatoric, analytically propagated via the Delta method for multi-variate correlated inputs). Leveraging the inherent strength of GPR in low-data regimes, the model, trained on a meagre 120 data points, achieved a state-of-the-art Mean Absolute Percentage Error (MAPE) of 0.26% and an R2 of 0.999 on the independent test data, validating a truly robust and reliable surrogate model for advanced FO process optimization and digital twin development.

💡 Research Summary

The paper presents a robust hybrid physics‑machine‑learning framework for predicting water flux (Jw) in forward osmosis (FO) systems, addressing the long‑standing challenge of accurate flux modeling in the presence of complex internal mass‑transfer phenomena and limited experimental data. The authors first construct a detailed mechanistic FO model based on the Spiegler‑Kedem formulation, incorporating both external concentration polarization (ECP) and internal concentration polarization (ICP) through the structural parameter S = τ t_psl / ε_psl and solute diffusivity. This physical model yields a baseline flux prediction Jw_physical for any given set of ten input variables (membrane intrinsic properties A, ε_psl, τ, t_psl; solution concentrations cf_in and cd_in; feed and draw flow velocities uf_in and ud_in; channel length Lx; channel height tc).

Because the physical model cannot capture all real‑world nonlinearities and measurement variability, the authors define the residual error e = Jw_actual − Jw_physical and train a Gaussian Process Regression (GPR) model on this residual. The GPR uses a Matérn 5/2 kernel, and its hyper‑parameters are optimized via marginal likelihood maximization. The posterior mean g_GPR(z) provides a data‑driven correction to the physical prediction, while the posterior variance σ²_model(z) supplies epistemic (model) uncertainty that naturally grows with distance from the training domain. The final hybrid prediction is Jw_hybrid(z) = Jw_physical(z) + g_GPR(z).

A key contribution is the complete uncertainty quantification (UQ) strategy. Epistemic uncertainty is obtained directly from the GPR posterior. Aleatoric (input) uncertainty is propagated analytically using the first‑order Delta method: σ²_input ≈ Jᵀ Σ_z J, where Σ_z is the covariance matrix of the ten input features (constructed from experimentally estimated coefficients of variation and assumed correlations) and J is the Jacobian of the hybrid model with respect to the inputs (the sum of the physical model gradient and the GPR gradient). To validate the linear approximation, the authors perform Monte‑Carlo simulations (1,000 perturbed samples) for three representative test points, showing excellent agreement between the Delta‑method variance and the Monte‑Carlo variance.

The methodology is evaluated on a modest training set of 120 experimental points and an independent test set of 2,854 points. The hybrid model achieves a mean absolute percentage error (MAPE) of 0.26 % and an R² of 0.999 on the test data, dramatically outperforming pure mechanistic models and conventional data‑driven approaches such as artificial neural networks, especially in low‑data regimes. Moreover, the total predictive variance σ²_total = σ²_model + σ²_input provides a full probabilistic description of prediction confidence, enabling risk‑aware design, digital‑twin implementation, and real‑time process optimization.

In summary, the study demonstrates that (1) residual learning with GPR can effectively correct mechanistic FO models without sacrificing physical consistency; (2) Bayesian GPR furnishes a principled epistemic uncertainty estimate; (3) the Delta method, validated against Monte‑Carlo, offers a computationally cheap yet accurate propagation of input measurement errors; and (4) the resulting hybrid surrogate delivers state‑of‑the‑art accuracy and comprehensive uncertainty quantification, setting a new benchmark for forward osmosis modeling and its integration into advanced process‑control and design workflows.