{"status":"ok","message-type":"work","message-version":"1.0.0","message":{"institution":[{"name":"bioRxiv"}],"indexed":{"date-parts":[[2026,10,1]],"date-time":"2026-10-01T12:57:16Z","timestamp":1790859436040,"version":"4.1.0"},"posted":{"date-parts":[[2026,9,28]]},"group-title":"Bioengineering","reference-count":36,"publisher":"openRxiv","license":[{"start":{"date-parts":[[2026,9,28]],"date-time":"2026-09-28T00:00:00Z","timestamp":1790553600000},"content-version":"vor","delay-in-days":0,"URL":"https:\/\/www.biorxiv.org\/about\/FAQ#license"}],"funder":[{"award":["P30CA093373"],"award-info":[{"award-number":["P30CA093373"]}],"id":[{"id":"https:\/\/ror.org\/040gcmg81","id-type":"ROR","asserted-by":"publisher"}]},{"award":["U01CA289068"],"award-info":[{"award-number":["U01CA289068"]}],"id":[{"id":"https:\/\/ror.org\/040gcmg81","id-type":"ROR","asserted-by":"publisher"}]}],"content-domain":{"domain":[],"crossmark-restriction":false},"short-container-title":[],"accepted":{"date-parts":[[2026,9,28]]},"abstract":"<jats:title>Abstract<\/jats:title>\n                <jats:p>\n                  Trans-arterial radioembolization using yttrium-90 (Y-90) microspheres treats unresectable liver cancer with radiation. The tumor-to-parenchyma dose ratio is mainly governed by the patient-specific hepatic arterial flow distribution, which transports the microspheres. Even though computational fluid dynamics (CFD) simulation can predict this flow distribution, its high computational cost and time limit intraprocedural application. To overcome this computational limitation, a physics-informed neural network (PINN) surrogate was developed to provide rapid predictions of flow distribution in a patient-specific hepatic artery model. The surrogate maps spatial coordinates to the velocity vector\n                  <jats:inline-formula>\n                    <jats:inline-graphic xmlns:xlink=\"http:\/\/www.w3.org\/1999\/xlink\" xlink:href=\"754555v1_inline1.gif\"\/>\n                  <\/jats:inline-formula>\n                  with Cartesian components\n                  <jats:italic>u, v, w<\/jats:italic>\n                  and the static pressure\n                  <jats:italic>p<\/jats:italic>\n                  . Reference velocity and pressure fields came from a converged finite-element CFD solve of the same geometry at about 1.05 \u00d7 10\n                  <jats:sup>5<\/jats:sup>\n                  nodes; a controlled fraction was used to supervise the training while the remainder was held out for validation. Using the NVIDIA PhysicsNeMo framework, we compared four coordinate networks (MLP, Modified Fourier, Multi-Scale Fourier and SIREN) under soft and hard inlet enforcement across a supervised data sweep (data fraction\n                  <jats:italic>df<\/jats:italic>\n                  = 0 to 0.20) and repeated\n                  <jats:italic>df<\/jats:italic>\n                  = 0.20 at a matched 300,000-step budget. Predicted flow splits were propagated through an established Y-90 dose-kernel pipeline. Accuracy is governed primarily by the supervised-data fraction, not the architecture: at\n                  <jats:italic>df<\/jats:italic>\n                  \u2265 0.10 the backbones agree to within one to two\n                  <jats:italic>R<\/jats:italic>\n                  <jats:sup>2<\/jats:sup>\n                  points, and\n                  <jats:italic>R<\/jats:italic>\n                  <jats:sup>2<\/jats:sup>\n                  (\n                  <jats:italic>u<\/jats:italic>\n                  ) rises from \u2248 0 under pure physics (\n                  <jats:italic>df<\/jats:italic>\n                  = 0) to \u2248 0.95 by\n                  <jats:italic>df<\/jats:italic>\n                  = 0.01 for the best backbones, saturating near 0.97\u20130.99 by\n                  <jats:italic>df<\/jats:italic>\n                  = 0.20. At a matched 300,000-step budget the Multi-Scale Fourier network leads (\n                  <jats:italic>R<\/jats:italic>\n                  <jats:sup>2<\/jats:sup>\n                  (\n                  <jats:italic>u<\/jats:italic>\n                  ) = 0.978) ahead of SIREN (\n                  <jats:italic>R<\/jats:italic>\n                  <jats:sup>2<\/jats:sup>\n                  (\n                  <jats:italic>u<\/jats:italic>\n                  ) = 0.976) and the MLP (\n                  <jats:italic>R<\/jats:italic>\n                  <jats:sup>2<\/jats:sup>\n                  (\n                  <jats:italic>u<\/jats:italic>\n                  ) = 0.968); the best value seen (\n                  <jats:italic>R<\/jats:italic>\n                  <jats:sup>2<\/jats:sup>\n                  (\n                  <jats:italic>u<\/jats:italic>\n                  ) = 0.985, Modified Fourier) needed 1.8 times that budget. On the\n                  <jats:italic>df<\/jats:italic>\n                  = 0.20 checkpoint the per-outlet flow split matches CFD (\n                  <jats:italic>r<\/jats:italic>\n                  = 0.999,\n                  <jats:italic>L<\/jats:italic>\n                  <jats:sup>1<\/jats:sup>\n                  = 4.4 percentage points) and the derived dosimetry agrees to within 2.4% for perfused volume, mean dose and every isodose volume. At\n                  <jats:italic>df<\/jats:italic>\n                  = 0 training collapses toward the trivial near-zero state, which satisfies continuity, momentum and the zero-pressure outlets exactly and is ruled out only by the inlet and data terms; this, not architecture choice, is the limiting failure mode. Once trained, the network evaluates the full 105,284-node field in about 2 s on a four-core CPU. The supervised-data fraction, with a practical floor between\n                  <jats:italic>df<\/jats:italic>\n                  = 0.05 and 0.10, is the primary design lever for such hepatic-arterial PINN surrogates.\n                <\/jats:p>","DOI":"10.64898\/2026.09.26.754555","type":"posted-content","created":{"date-parts":[[2026,9,28]],"date-time":"2026-09-28T15:00:29Z","timestamp":1790607629000},"source":"Crossref","is-referenced-by-count":0,"title":["A Physics-Informed Neural Network Surrogate for Patient-Specific Hepatic Arterial Hemodynamics in Yttrium-90 Radioembolization: Network Architecture, Boundary-Condition Enforcement, and Data Efficiency"],"prefix":"10.64898","author":[{"given":"Erfan","family":"Taatizadeh","sequence":"first","affiliation":[{"name":"Department of Biomedical Engineering, University of California, Davis, CA, USA"}],"role":[{"vocabulary":"crossref","role":"author"}]},{"given":"Nisanth Kumar","family":"Panneer Selvam","sequence":"additional","affiliation":[{"name":"Department of Biomedical Engineering, University of California, Davis, CA, USA"}],"role":[{"vocabulary":"crossref","role":"author"}]},{"given":"Emilie","family":"Roncali","sequence":"additional","affiliation":[{"name":"Department of Biomedical Engineering, University of California, Davis, CA, USA"},{"name":"Department of Radiology, University of California, Davis Health, Sacramento, CA, USA"}],"role":[{"vocabulary":"crossref","role":"author"}]}],"member":"54368","reference":[{"key":"2026100105000867000_2026.09.26.754555v1.1","doi-asserted-by":"publisher","DOI":"10.1146\/annurev-bioeng-110824-021423"},{"key":"2026100105000867000_2026.09.26.754555v1.2","doi-asserted-by":"publisher","DOI":"10.3390\/bioengineering7030064"},{"key":"2026100105000867000_2026.09.26.754555v1.3","doi-asserted-by":"publisher","DOI":"10.1002\/cnm.2764"},{"key":"2026100105000867000_2026.09.26.754555v1.4","doi-asserted-by":"publisher","DOI":"10.1002\/cnm.2791"},{"key":"2026100105000867000_2026.09.26.754555v1.5","doi-asserted-by":"publisher","DOI":"10.1016\/B978-0-323-85740-6.00012-1"},{"key":"2026100105000867000_2026.09.26.754555v1.6","doi-asserted-by":"publisher","DOI":"10.1007\/s10439-020-02469-1"},{"key":"2026100105000867000_2026.09.26.754555v1.7","doi-asserted-by":"publisher","DOI":"10.1007\/s10915-022-01939-z"},{"key":"2026100105000867000_2026.09.26.754555v1.8","doi-asserted-by":"publisher","DOI":"10.1007\/s10409-021-01148-1"},{"key":"2026100105000867000_2026.09.26.754555v1.9","doi-asserted-by":"publisher","DOI":"10.1038\/s42256-021-00302-5"},{"key":"2026100105000867000_2026.09.26.754555v1.10","doi-asserted-by":"publisher","DOI":"10.1016\/j.cma.2019.112623"},{"key":"2026100105000867000_2026.09.26.754555v1.11","doi-asserted-by":"publisher","DOI":"10.1109\/TMI.2022.3161653"},{"key":"2026100105000867000_2026.09.26.754555v1.12","doi-asserted-by":"publisher","DOI":"10.1016\/j.compbiomed.2024.108706"},{"key":"2026100105000867000_2026.09.26.754555v1.13","doi-asserted-by":"publisher","DOI":"10.1016\/j.ynexs.2024.100016"},{"key":"2026100105000867000_2026.09.26.754555v1.14","doi-asserted-by":"publisher","DOI":"10.1016\/j.compbiomed.2023.107287"},{"key":"2026100105000867000_2026.09.26.754555v1.15","doi-asserted-by":"publisher","DOI":"10.1146\/annurev-bioeng-110824-124907"},{"key":"2026100105000867000_2026.09.26.754555v1.16","unstructured":"A. S. Krishnapriyan , A. Gholami , S. Zhe , R. M. Kirby , and M. W. Mahoney , \u201cCharacterizing Possible Failure Modes in Physics-Informed Neural Networks,\u201d in Advances in Neural Information Processing Systems, 2021, pp. 26548\u201326560."},{"key":"2026100105000867000_2026.09.26.754555v1.17","unstructured":"R. Leiteritz and D. Pfl\u00fcger , \u201cHow to Avoid Trivial Solutions in Physics-Informed Neural Networks,\u201d 2021. [Online]. Available: https:\/\/arxiv.org\/abs\/2112.05620"},{"key":"2026100105000867000_2026.09.26.754555v1.18","unstructured":"M. Tancik et al., \u201cFourier Features Let Networks Learn High Frequency Functions in Low Dimensional Domains,\u201d in Advances in Neural Information Processing Systems, 2020, pp. 7537\u20137547."},{"key":"2026100105000867000_2026.09.26.754555v1.19","unstructured":"V. Sitzmann , J. N. P. Martel , A. W. Bergman , D. B. Lindell , and G. Wetzstein , \u201cImplicit Neural Representations with Periodic Activation Functions,\u201d in Advances in Neural Information Processing Systems, 2020, pp. 7462\u20137473."},{"key":"2026100105000867000_2026.09.26.754555v1.20","unstructured":"R. Fathony , A. K. Sahu , D. Willmott , and J. Z. Kolter , \u201cMultiplicative Filter Networks,\u201d International Conference on Learning Representations (ICLR), 2021."},{"key":"2026100105000867000_2026.09.26.754555v1.21","doi-asserted-by":"publisher","DOI":"10.3390\/fluids8020046"},{"key":"2026100105000867000_2026.09.26.754555v1.22","doi-asserted-by":"publisher","DOI":"10.1002\/mp.16379"},{"key":"2026100105000867000_2026.09.26.754555v1.23","unstructured":"T. Salimans and D. P. Kingma , \u201cWeight Normalization: A Simple Reparameterization to Accelerate Training of Deep Neural Networks,\u201d in Advances in Neural Information Processing Systems, 2016, pp. 901\u2013909."},{"key":"2026100105000867000_2026.09.26.754555v1.24","doi-asserted-by":"publisher","DOI":"10.1186\/s13104-021-05631-7"},{"key":"2026100105000867000_2026.09.26.754555v1.25","unstructured":"R. Pal , S. Mukherjee , U. Dutta , and A. Choudhury , \u201cSolving Navier\u2013Stokes Equations Using Data-Free Physics-Informed Neural Networks with Hard Boundary Conditions,\u201d 2025. [Online]. Available: https:\/\/arxiv.org\/abs\/2511.14497"},{"key":"2026100105000867000_2026.09.26.754555v1.26","doi-asserted-by":"publisher","DOI":"10.1016\/j.cma.2025.117914"},{"key":"2026100105000867000_2026.09.26.754555v1.27","unstructured":"PhysicsNeMo Contributors , NVIDIA PhysicsNeMo: An open-source framework for physics-based deep learning in science and engineering. (Feb. 24, 2023). [Online]. Available: https:\/\/github.com\/NVIDIA\/physicsnemo"},{"key":"2026100105000867000_2026.09.26.754555v1.28","doi-asserted-by":"publisher","DOI":"10.1016\/j.enganabound.2024.105933"},{"key":"2026100105000867000_2026.09.26.754555v1.29","doi-asserted-by":"publisher","DOI":"10.1016\/j.jcp.2022.111722"},{"key":"2026100105000867000_2026.09.26.754555v1.30","doi-asserted-by":"publisher","DOI":"10.3934\/mbe.2023512"},{"key":"2026100105000867000_2026.09.26.754555v1.31","doi-asserted-by":"publisher","DOI":"10.46298\/mbj.16176"},{"key":"2026100105000867000_2026.09.26.754555v1.32","doi-asserted-by":"publisher","DOI":"10.3390\/fluids10120306"},{"key":"2026100105000867000_2026.09.26.754555v1.33","doi-asserted-by":"publisher","DOI":"10.1016\/j.jcp.2020.109913"},{"key":"2026100105000867000_2026.09.26.754555v1.34","unstructured":"S. Wang , S. Koohy , Y. Lu , and P. Perdikaris , \u201cWhen PINNs Go Wrong: Pseudo-Time Stepping Against Spurious Solutions,\u201d 2026. [Online]. Available: https:\/\/arxiv.org\/abs\/2604.23528"},{"key":"2026100105000867000_2026.09.26.754555v1.35","doi-asserted-by":"publisher","DOI":"10.1063\/5.0188830"},{"key":"2026100105000867000_2026.09.26.754555v1.36","unstructured":"S. Wang , S. Sankaran , P. Stinis , and P. Perdikaris , \u201cSimulating Three-Dimensional Turbulence with Physics-Informed Neural Networks,\u201d 2025. [Online]. Available: https:\/\/arxiv.org\/abs\/2507.08972"}],"container-title":[],"original-title":[],"link":[{"URL":"https:\/\/syndication.highwire.org\/content\/doi\/10.64898\/2026.09.26.754555","content-type":"unspecified","content-version":"vor","intended-application":"similarity-checking"}],"deposited":{"date-parts":[[2026,10,1]],"date-time":"2026-10-01T12:00:30Z","timestamp":1790856030000},"score":1,"resource":{"primary":{"URL":"http:\/\/biorxiv.org\/lookup\/doi\/10.64898\/2026.09.26.754555"}},"subtitle":[],"short-title":[],"issued":{"date-parts":[[2026,9,28]]},"references-count":36,"URL":"https:\/\/doi.org\/10.64898\/2026.09.26.754555","relation":{},"subject":[],"published":{"date-parts":[[2026,9,28]]},"subtype":"preprint"}}