Neural Networks Crack the Secrets of Dual-Non-Newtonian Fluid Flow Under Magnetic and Bioconvective Forces

Fluids that refuse to behave like water are everywhere in modern industry: blood coursing through capillaries, molten polymers being extruded into films, paints spreading across surfaces, and slurries moving through chemical reactors. These non-Newtonian materials defy the simple linear relationship between stress and deformation that governs ordinary fluids, and understanding how they flow when heated, magnetized, and laced with nanoparticles has long been one of the thorniest challenges in theoretical fluid mechanics. Now, a team of researchers led by Safeer Muhammad of the Islamic University of Madinah has published in Results in Engineering one of the most comprehensive attacks on the problem to date, combining classical numerical simulation with artificial neural networks to model a hybrid Casson-micropolar nanofluid under an extraordinary range of simultaneous physical effects.

The study’s central innovation lies in its sheer scope. Rather than examining one or two phenomena in isolation, the researchers constructed a single mathematical framework that couples magnetohydrodynamics, porous media, thermal radiation, viscous dissipation, Soret and Dufour cross-diffusion, the Buongiorno nanofluid model, multiple velocity and thermal slip conditions, and bioconvection driven by gyrotactic microorganisms. The working fluid itself is a dual material: a Casson fluid, which behaves like a plastic solid until a critical yield stress is exceeded, blended with micropolar characteristics that allow suspended particles to spin independently of the bulk motion. This micro-rotation matters enormously in real materials, because the internal structure of blood, polymer melts, and many slurries directly influences how efficiently they carry heat.

Mathematically, the team began with the full Casson rheological equation, in which the Cauchy stress tensor depends on whether the product of the deformation rate components exceeds a critical threshold. They then wrote the conservation equations for momentum, angular momentum, energy, nanoparticle concentration, and microorganism density for flow over a vertical linearly stretching sheet, where the wall velocity grows linearly with distance from the origin. Buoyancy enters the momentum balance through three separate Grashof numbers, one each for thermal, solutal, and microorganism density differences, meaning the fluid can be driven simultaneously by heat, by chemical concentration gradients, and by the swimming of bacteria. A transverse magnetic field of strength B0 imposes a Lorentz retarding force, while the porous matrix adds both Darcy resistance and Forchheimer inertial drag.

To render this formidable system tractable, the researchers applied classical similarity transformations that collapse the governing partial differential equations into a set of coupled nonlinear ordinary differential equations for five dimensionless profiles: the axial velocity, the micro-rotation or angular velocity, the temperature, the nanoparticle concentration, and the density of motile microorganisms. Each equation bristles with dimensionless groups, from the Casson parameter and magnetic parameter to the Eckert number measuring viscous heating, the Brownian motion and thermophoresis parameters from Buongiorno’s model, the Lewis and Schmidt numbers governing diffusion, and the Peclet number describing how strongly microorganisms swim relative to how fast they diffuse. Five distinct slip parameters modify the boundary conditions at the wall, allowing velocity, temperature, concentration, and microorganism density to jump discontinuously at the fluid-solid interface, a realistic feature at microscale that almost no previous study has treated all at once.

The numerical solution was carried out with MATLAB’s bvp5c solver, which implements a three-stage Lobatto IIIA collocation formula specially designed for boundary value problems. The higher-order equations were first reduced to a system of first-order equations, and a shooting method was used to satisfy the conditions at both the wall and the far field. To verify the code, the team compared their computed skin friction values against previously published results for the micropolar Casson problem, finding agreement to six or more significant figures across magnetic parameter values ranging from 1 to 1000. This benchmark, they argue, establishes the reliability of the framework before it was extended to the full multi-physics case.

The machine learning component is where the work takes a distinctly modern turn. The researchers generated training data from the bvp5c solutions and built a multilayer perceptron artificial neural network trained with the Levenberg-Marquardt backpropagation algorithm, a method prized for its rapid and stable convergence on nonlinear least-squares problems. Eight input physical parameters fed the network, which predicted all five flow profiles simultaneously. The data were split into training, validation, and testing subsets, and the team tracked mean squared error across epochs, observing steady decline until optimal validation performance was reached after as few as nine epochs in some scenarios. Error histograms showed errors tightly clustered around zero, and regression plots placed every data point from every subset on the equality line, with coefficients of determination equal to one. Absolute errors between the network predictions and the numerical reference solutions fell between ten to the minus four and ten to the minus five, a level of fidelity the authors say confirms the network as a trustworthy surrogate for the full solver.

The physics that emerges from the simulations is rich and often counterintuitive. Increasing the magnetic parameter slows the fluid dramatically, because the Lorentz force opposes motion and thins the momentum boundary layer, a result with direct implications for controlling cooling rates in metallurgical and manufacturing processes. A stronger porous matrix resistance similarly drains momentum from the flow. By contrast, all three buoyancy mechanisms accelerate the fluid: raising the thermal, solutal, or microorganism Grashof number curves the velocity profile upward as buoyant plumes near the heated, concentrated, or bioactively populated wall drive stronger convection. The spin gradient parameter sharpens angular velocity near the wall, while the microinertia parameter suppresses it, reflecting the resistance of suspended elements to changing their rotational state.

On the thermal side, the Eckert number raises fluid temperature by converting kinetic energy into heat through viscous dissipation, thickening the thermal boundary layer. Both Brownian motion, which randomizes nanoparticle trajectories and intensifies fluid-particle collisions, and thermophoresis, which drives nanoparticles from hot regions toward cold ones, elevate the temperature profile. Yet the two mechanisms split on concentration: Brownian motion flattens the nanoparticle concentration profile by spreading particles more uniformly, while thermophoresis piles particles up near the cold far field and steepens the concentration boundary layer. For the microorganisms, a higher Peclet number means swimming transport overwhelms diffusion, pushing cells away from the wall and lowering their density in the boundary layer. Slip effects act as universal suppressors: velocity, thermal, solutal, and microorganism slip each weaken the transfer of momentum, heat, mass, or cells from the stretching surface into the fluid, thinning the corresponding boundary layers.

The engineering quantities tell a similarly coherent story. The Forchheimer number, which quantifies nonlinear inertial drag in the porous medium, increases skin friction by steepening the velocity gradient at the wall. A heat source thickens the thermal boundary layer and reduces the Nusselt number, while a heat sink does the opposite, extracting heat and boosting heat transfer rates. Higher Schmidt numbers thin the concentration boundary layer and raise the Sherwood number, and higher Lewis numbers reduce microorganism density at the surface. Together, these trends offer designers of thermal systems a tunable menu: magnetic fields and porous matrices to tame runaway flows, heat sinks to maximize cooling, and slip engineering to modulate transport at microstructured walls.

The authors are candid about the limits of the work. The model is laminar, steady, and incompressible, and no experimental validation or real-time physical implementation was attempted. Future extensions, they suggest, could incorporate more complex geometries, variable material properties, and more advanced machine learning architectures to push predictive accuracy higher while cutting computational cost. Even so, the study stands as a striking demonstration that neural networks can now serve as near-perfect surrogates for some of the most forbidding equation systems in fluid mechanics, potentially allowing engineers to explore design spaces that brute-force numerical simulation would take orders of magnitude longer to map. For industries that live and die by heat transfer, from electronics cooling to biomedical devices, that capability may prove as consequential as the fluid physics itself.

Subject of Research: Numerical and machine learning analysis of MHD Casson-micropolar nanofluid flow with bioconvection over a stretching sheet

Article Title: Numerical and machine learning investigation of dual-material fluid flow behavior under thermal and solutal rheological mechanisms

Article References: Muhammad, S., Khan, M., Muhammad, Rehman, G., Atif, H. M., & Hussain, S. M. (2026). Numerical and machine learning investigation of dual-material fluid flow behavior under thermal and solutal rheological mechanisms. Results in Engineering, 32, Article 113022. https://doi.org/10.1016/j.rineng.2026.113022

Image Credits: AI Generated

DOI: 10.1016/j.rineng.2026.113022

Keywords: Casson fluid, micropolar fluid, magnetohydrodynamics, nanofluid, bioconvection, artificial neural network, Buongiorno model, thermal radiation, stretching sheet, boundary layer, Levenberg-Marquardt, porous medium