Source Code and Datasets for: Transfer learning-based physics-informed neural networks for predicting flow and heat transfer in complex 3D structures

The reference benchmark solutions were generated via Computational Fluid Dynamics (CFD) simulations. The 3D irregular channel geometry was modeled using SpaceClaim and discretized using ANSYS Meshing with approximately 1.82 million mesh elements.

The underlying physical processes are governed by steady-state continuity, Navier-Stokes, and energy equations.The uploaded archive is organized into Python scripts (.py), PyTorch model weights (.pt), and data arrays (.npy and .mat formats). The "mat文件" folder contains the spatial coordinates (x, y, z) and the corresponding CFD reference field variables, including velocity components (u, v, w) in m/s, pressure (p) in Pa, and thermodynamic temperature (T) in K. The provided Python scripts handle data preprocessing, Latin Hypercube Sampling for spatial point generation, neural network training routines, and parameter initialization for transfer learning across adjacent operating conditions.To evaluate the robustness of the PINN model against data errors, artificial Gaussian noise with intensities scaling from 10% to 100% was introduced into the training labels.