AI Learns the Physics of Magnetized Nanofluid Flow for Solar Thermal Design

Simulating how a magnetized nanofluid behaves as it flows over a porous, stretching cylinder is a computationally punishing business. The governing equations are nonlinear, coupled, and must be solved repeatedly across wide ranges of magnetic strength, curvature, radiation, and porosity before any real engineering insight emerges. A study published in Discover Nano on 28 October 2026 proposes a way around that bottleneck: a physics-informed neural network, or PINN, trained against a high-fidelity numerical solution, that reproduces the velocity, temperature, and concentration fields of a Williamson-model ternary nanofluid in a fraction of the time, and then pairs that surrogate with a global sensitivity analysis to reveal which parameters actually matter.

The working fluid is deliberately exotic. The researchers modeled a ternary, or tri-hybrid, nanofluid in which multi-walled carbon nanotubes, A7075 alloy nanoparticles, and graphene oxide are all suspended simultaneously in a base liquid of 60:40 ethylene glycol and water. Each nanoparticle contributes something distinct: the carbon nanotubes bring high thermal conductivity, the A7075 a beneficial metallic phase, and the graphene oxide an extensive surface area for thermal interaction. The ethylene glycol–water blend was chosen for its balance of thermal conductivity and flow properties, a combination the authors note functions efficiently as a heat-transfer agent in contexts relevant to solar thermal energy systems.

The physical setup is a steady, incompressible Williamson fluid flowing over a cylinder that stretches axially, with an inclined magnetic field applied at an angle to the axial direction. The model incorporates Darcy–Forchheimer porous-medium resistance, thermal radiation, activation energy in the concentration equation, and convective boundary conditions at the cylinder wall. Using similarity transformations, the governing partial differential equations collapse into a set of nonlinear ordinary differential equations for the dimensionless velocity, temperature, and concentration profiles, parameterized by quantities such as the curvature parameter, magnetic parameter, Williamson fluid parameter, Forchheimer number, Prandtl number, Schmidt number, Biot number, and radiation parameter.

To generate trustworthy training data, the team first solved these equations with a modified finite difference implementation of the Keller-Box method, a scheme long favored for boundary-layer problems because it is unconditionally stable and second-order accurate. The method transforms the higher-order equations into first-order form, discretizes the domain with central differences, linearizes the nonlinear terms with Newton’s method, and solves the resulting system with block tridiagonal elimination. The authors report that the solution proved fairly stable with respect to the initial guesses within the examined range, and these high-fidelity results then served as the benchmark data for training and validating the neural surrogate.

The machine learning architecture itself is a hybrid. A graph neural network stage represents the computational domain as roughly 1000 discrete nodes spanning the dimensionless coordinate from 0 to 10, each connected to its nearest neighbors, with node states updated through a tanh activation and random weights. A feed-forward prediction network with two hidden layers of 11 and 14 sigmoid-activated neurons then maps the final node states to the three physical fields, using a linear output layer to accommodate the continuous range of predicted variables. Physics compliance enters through residual terms derived from the momentum, energy, and concentration equations, evaluated with central difference approximations, and folded into a combined loss function weighting both data fidelity and equation residuals.

Training used the Levenberg–Marquardt algorithm with L2 weight decay, and the convergence behavior was striking. As epochs accumulated, the ordinary differential equation residuals approached zero and the loss fell on a logarithmic scale, with the total loss reaching roughly ten to the minus nine during training; the abstract reports PINN loss metrics persistently below ten to the minus eleven, which the authors interpret as evidence of exceptional computational precision. Regression comparisons between predicted and numerical profiles yielded coefficients of determination exceeding 0.99, and in the parameter-by-parameter validation studies the reported R-squared values equaled 1, with root-mean-square errors between roughly ten to the minus four and ten to the minus six across the velocity, temperature, and concentration predictions.

The sensitivity analysis is where the surrogate earns its keep. Using Sobol variance decomposition over parameter ranges spanning plus or minus twenty percent of baseline values, the team computed first-order and total-order indices for each response. For velocity, the curvature parameter dominated with a total-order index of 0.487, followed by the magnetic parameter at 0.244, the Williamson parameter at 0.181, and the Forchheimer number at 0.093, together accounting for about ninety-seven percent of the response variance. Temperature was governed primarily by the Prandtl number at 0.717, with the Biot number at 0.189 and the radiation parameter at 0.101. Concentration was overwhelmingly controlled by the Schmidt number at 0.931, with minor contributions from curvature at 0.054 and activation energy at 0.017.

The parametric sweeps themselves follow physically intuitive patterns. Increasing the curvature parameter raises all three profiles, attributed to enhanced centrifugal forces, improved convective heat transfer, and more efficient nanoparticle mixing. A stronger magnetic field suppresses velocity through the Lorentz force while elevating temperature and concentration, a pattern the authors note is consistent with prior experimental work on MHD nanofluid flow. Higher Williamson parameter values increase velocity through augmented viscoelastic momentum transfer, the Biot number thickens the thermal boundary layer, thermal radiation thickens it further, and rising activation energy slows the chemical reaction and raises the concentration profile. Streamline visualizations show that with the magnetic field switched off the boundary layer grows uniformly downstream, whereas at a magnetic parameter of three the Lorentz force decelerates the fluid and lowers the peak velocity to around 7.17.

The practical payoff, the authors argue, is speed: once trained, the PINN surrogate delivers rapid predictions that alleviate the computational cost of repeated high-fidelity simulations, making parametric analysis and global sensitivity evaluation far more tractable for real-time assessment, optimization, and decision support in solar thermal and industrial cooling contexts. The team also reports that the deep network components achieved a mean deviation of 0.8 milliseconds in operational timing, supporting efficient sensitivity analysis and rapid thermal control. Comparison of the skin-friction results against earlier published studies at limiting parameter values showed consistency with prior findings.

The authors are careful about scope. The framework is constrained to the specified configuration, parameter range, and underlying physical assumptions, and the computational model embeds simplified assumptions about the fluid flow, porous material, and thermophysical properties. Experimental validation would be required to demonstrate predictive capability under actual operating conditions, and future work is expected to target broader parameter ranges, more intricate geometries, transient phenomena, more realistic porous structures, and uncertainty-aware learning with adaptive sampling. Within those limits, the study offers a concrete demonstration that embedding physical laws directly into a neural network’s loss function can turn a slow, equation-bound simulation problem into a fast, physics-constrained surrogate, without losing sight of which physical levers actually drive the system.

Subject of Research: Physics-informed neural network surrogate modeling and global sensitivity analysis of magnetized ternary nanofluid flow over a porous stretching cylinder

Article Title: Physics informed neural network surrogate modeling and global sensitivity analysis of magnetized nanofluid flow over a porous cylinder for solar thermal energy applications

Article References: Hussain, M., & Alhushaybari, A. (2026). Physics informed neural network surrogate modeling and global sensitivity analysis of magnetized nanofluid flow over a porous cylinder for solar thermal energy applications. Discover Nano, 21(1), Article 569. https://doi.org/10.1186/s11671-026-05023-5

Image Credits: AI Generated

DOI: 10.1186/s11671-026-05023-5

Keywords: physics-informed neural networks, nanofluid, magnetohydrodynamics, solar thermal energy, global sensitivity analysis, Sobol indices, heat transfer, surrogate modeling, ternary nanofluid, porous cylinder, machine learning, computational fluid dynamics