Surrogate Modelling for Patient-Specific Cerebral Aneurysm CFD Reconstruction

Abstract

Patient-specific computational fluid dynamics (CFD) shows in detail how blood moves inside a cerebral aneurysm, but it is slow to run. For a single patient, preparing the geometry, generating a mesh, solving, and post-processing may take anywhere from hours to days, and this becomes a real obstacle when many cases need to be reviewed in a short time. In this work we ask a more limited question: once a CFD simulation has been run, can a neural network reconstruct its velocity and pressure fields accurately enough to be useful, and which type of network does this best? We built an automated OpenFOAM pipeline that converted 58 aneurysm geometries from the AneuriskWeb database into 10,210,671 labelled CFD points, and we trained three surrogate models on the same data. A single global model that uses Fourier coordinate encoding and a per-case embedding (CaseFourierField) reaches 74.08% velocity reconstruction accuracy from one checkpoint. Training a separate Sinusoidal Representation Network (SIREN) for each patient does better, with 85.05% mean accuracy and 97.67% on the best case, but needs one model per geometry. A graph neural operator (AneurysmGNO) is the most physically grounded of the three, although under the memory-limited evaluation we were able to run it is not yet competitive, and we explain why. We also report an early approach that did not work: a Poiseuille physics prior performed well on a small controlled dataset but failed badly across all 58 cases. We include this because it explains the design decisions that followed. All results are measured within known geometries and should be read as field reconstruction, not as prediction on unseen patients.