QCPINN — AI-Accelerated CFD
Physics-Informed Neural Networks for real-time intracranial haemodynamics (Imperial MSc Dissertation)
MSc Dissertation · Imperial College London · 2025–2026
Overview
My MSc dissertation at Imperial College London develops QCPINN — a Quasi-Cartesian Physics-Informed Neural Network — designed to replace computationally expensive Navier–Stokes finite-volume solvers with a real-time, mesh-free neural surrogate for patient-specific intracranial haemodynamics.
Traditional CFD pipelines for aneurysm haemodynamics require multi-hour simulations per patient geometry, making real-time clinical decision support infeasible. QCPINN embeds the governing PDE residuals directly into the network’s loss function, enforcing physical consistency without labelled simulation data.
Architecture
The QCPINN builds on the PINN framework (Raissi et al., 2019) with a domain-adapted coordinate transformation:
- Quasi-Cartesian embedding: Maps the curvilinear vessel geometry to a structured domain, improving gradient flow and convergence compared to vanilla collocation-point PINNs.
- Multi-task loss: Simultaneous minimisation of: PDE residual (continuity + momentum), boundary condition loss (no-slip walls, inlet Womersley profile), and sparse data loss (CFD snapshot supervision at key timesteps).
- Pulsatile inlet BC: Womersley flow profile with patient-specific cardiac frequency and stroke volume.
Input: (x, y, z, t) → [Encoder] → [Residual blocks × 6] → [Output head: u, v, w, p]
Loss = λ₁·L_pde + λ₂·L_bc + λ₃·L_data
Training runs on Imperial’s HPC cluster (CX3) using PyTorch + CUDA, with adaptive loss-weight scheduling (NTK-informed).
Key Results
| Metric | Value |
|---|---|
| Velocity MAE vs. CFD | < 4% |
| Pressure field RMSE | < 6% |
| Inference time (full 3D field) | < 80 ms |
| Training time (HPC, 4× A100) | ~14 hours |
The model achieves real-time inference over a patient-specific geometry — enabling a potential clinical tool where a surgeon inputs geometry from MRA imaging and receives haemodynamic risk maps within seconds rather than hours.
Clinical Relevance
Elevated Wall Shear Stress (WSS) and Oscillatory Shear Index (OSI) are established rupture-risk predictors for intracranial aneurysms. QCPINN reconstructs these fields with sub-clinical-grade accuracy, creating a pathway toward AI-assisted pre-operative planning.