Rapid Flow-Field Data Generation Using Physics-Informed Neural Networks Combined with Interpolation Algorithms

High-resolution computational fluid dynamics (CFD) simulations often require prohibitive runtimes and memory. To alleviate these limitations, we propose a rapid flow-field datageneration framework that couples a physics-informed neural network (PINN) with interpolation. First, a PINN is trained on coarse-grid CFD data to synthesise additional coarse-grid samples. Second, the samples are mapped to a fine grid through interpolation. Third, the pretrained PINN refines the interpolated field, yielding a high-fidelity fine-grid solution. Experiments show that the proposed method raises the mean accuracy of three representative flow variables from <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$63\%, 75\%, 79\%$</tex>, and <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$82\%$</tex> to <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$96\%, 97\%, 97\%$</tex>, and <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$98\%$</tex> for linear, bilinear, cubic-spline and Kriging interpolation, respectively-an improvement of up to 33 percentage points. For the same fine-grid size, the four configurations complete in <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$12 ~\min, 15 ~\min, 21 ~\min$</tex> and 32 min with peak-memory utilisation of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$35\%, 39\%, 45\%$</tex> and <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$57\%$</tex>; all are markedly lower than the 42 min and <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$75\%$</tex> required by a full finegrid CFD run. These results demonstrate that the proposed PINNinterpolation approach delivers high-accuracy flow-field data while significantly reducing computational time and memory, providing an efficient tool for intelligent flow-field prediction in complex engineering applications.