M3PI-DeepONet predicts aneurysm blood flow 36x faster with <4% error
New physics-informed AI models unsteady 3D blood flow using only 0.3% labeled data.
A research team led by Oscar L. Cruz-Gonzalez, Valérie Deplano, and Badih Ghattas has introduced M3PI-DeepONet, a physics-informed deep operator network designed to simulate unsteady hemodynamics in abdominal aortic aneurysms (AAA). The model addresses a persistent challenge in computational medicine: predicting complex 3D flow patterns, including wall shear stress, vortex structures, and pressure distributions, quickly enough for clinical use. Traditional computational fluid dynamics (CFD) is accurate but computationally expensive, limiting its adoption in time-sensitive diagnostic workflows.
M3PI-DeepONet introduces an Aggregated Injection strategy that fuses latent representations from multiple input branches before trunk injection, allowing the coordinate basis to adapt to multiple physical constraints. It is the first architecture to combine layer-wise gating with a multi-branch operator-network topology, creating an input-adaptive trunk basis. The model is trained using the 3D Navier-Stokes equations as governing physical laws, alongside initial and boundary conditions, and only 0.3% of labeled internal data. In tests on idealized AAA geometries, it predicts unsteady 3D velocity fields with an average relative L2 error below 4% and pressure errors around 5%, while achieving an approximate 36x inference speedup over reference CFD simulations. This work marks a significant step toward real-time, non-invasive clinical diagnostics for aneurysm risk assessment.
- M3PI-DeepONet integrates 3D Navier-Stokes equations as physics constraints, requiring only 0.3% of labeled internal data
- Average relative L2 velocity error <4% and pressure error ~5% on unsteady 3D AAA flow predictions
- Conservative inference speedup of ~36x vs. reference CFD once branch conditioning inputs are available
Why It Matters
Enables near-real-time, non-invasive aneurysm hemodynamic assessment, potentially shifting cardiovascular diagnostics from hours-long CFD simulations to instant AI predictions.