SAGE-PINN: A Singularity-free Axisymmetric Geometry-Encoded Physics-Informed Neural Network for Axisymmetric Multiphysics Flow in Cylindrical Coordinates

Abstract

Physics-informed neural networks (PINNs) provide a mesh-free framework for multiphysics modeling, yet fluid flow in axisymmetric cylindrical coordinates remains largely unexplored because the governing operators are singular at the axis and momentum and energy are strongly coupled through advection and buoyancy. We introduce SAGE-PINN, a Singularity-free Axisymmetric Geometry-Encoded Physics-Informed Neural Network for this class of problems. SAGE-PINN removes the coordinate singularity by construction: the parity-preserving variable $s=r^2$ and a modified stream-function representation eliminate explicit $1/r$ terms from the trained residuals and satisfy mass conservation identically. Boundary-fitted analytical ansätze enforce the boundary conditions exactly, while staged momentum-thermal-joint training, residual normalization, Péclet- and Richardson-number continuation, and a physics-scaled drift guard stabilize the coupled optimization. The framework is trained without labelled data, using only physics and boundary constraints, and its predictions are validated against finite-element reference solutions for steady nanofluid flow and heat transfer through a stenosed artery over a wide range of flow and thermal regimes, without case-specific tuning. Mass conservation remains at single-precision round-off, and over the central operating range the axial velocity and temperature are reproduced within $1.6\%$ and $9.1\%$ relative $L_2$ error, respectively. The main deterioration occurs only in the limiting regimes of strongly buoyancy-influenced slow flow and highly advection-dominated heat transfer. Embedding cylindrical geometry, conservation structure, and boundary physics directly into the network provides a robust route to PINN simulation of axisymmetric multiphysics flows.