Physics-informed neural network for evaluating pressure drop in arterial stenoses based on simulation data

Gamilov T.M., Lange A., Osipova A.A., Liang F., Simakov S.S.
 pdf (652K)

This paper describes a method for generating a synthetic database of stenoses, consisting of 1620 entries. Each entry represents the results of a numerical experiment simulating the three-dimensional flow of a viscous incompressible fluid through a tube with a variable cross-section: pressure drop, mean flow rate, cross-sectionally averaged inlet blood flow velocity, maximum stenosis severity, stenosis length, stenosis asymmetry, tube radius, and Reynolds number. The database was validated by comparison with other models (with elastic walls) and bench experiments, showing a deviation in pressure drops of no more than 4%. The synthetic stenosis database was used to train a physics-informed neural network for the rapid estimation of pressure drop based on four key input parameters: Reynolds number, stenosis length, stenosis severity, and stenosis asymmetry coefficient. The physics-informed aspect was achieved by introducing penalties into the loss function for the absence of a positive pressure drop and for the lack of monotonicity of the pressure drop with respect to the input parameters. The physics-informed neural network demonstrated higher accuracy on hemodynamically significant stenoses when tested on a validation set and on new stenoses not represented in the database. The mean relative error for stenoses with a length of 8 healthy vessel radii was 6% for the physics-informed network and 13% for a classical neural network. The errors for short stenoses with a length of 4 radii were nearly identical: 9.5% for the physics-informed network and 10% for the classical neural network. The developed method for the functional assessment of the hemodynamic significance of stenoses can be used both as a standalone tool for clinical stenosis evaluation and as a component of network blood flow models. The approach becomes most relevant when modeling multi-vessel disease, which is predominant in clinical practice. The key advantage of the method lies in the physical correctness of the results and accuracy comparable to classical modeling, but with significantly lower computational costs.

Keywords: physics-informed neural network, synthetic database, stenosis, hemodynamics
Citation in English: Gamilov T.M., Lange A., Osipova A.A., Liang F., Simakov S.S. Physics-informed neural network for evaluating pressure drop in arterial stenoses based on simulation data // Computer Research and Modeling, 2026, vol. 18, no. 3, pp. 621-641
Citation in English: Gamilov T.M., Lange A., Osipova A.A., Liang F., Simakov S.S. Physics-informed neural network for evaluating pressure drop in arterial stenoses based on simulation data // Computer Research and Modeling, 2026, vol. 18, no. 3, pp. 621-641
DOI: 10.20537/2076-7633-2026-18-3-621-641
URL: http://crm-en.ics.org.ru/journal/article/3736/
Creative Commons License This work is licensed under a Creative Commons Attribution-NoDerivs 3.0 Unported License.

Copyright © 2026 Gamilov T.M., Lange A., Osipova A.A., Liang F., Simakov S.S.

Indexed in Scopus

Full-text version of the journal is also available on the web site of the scientific electronic library eLIBRARY.RU

The journal is included in the Russian Science Citation Index

The journal is included in the RSCI

International Interdisciplinary Conference "Mathematics. Computing. Education"

Website Copyright © 2009–2026 Institute of Computer Science