Introduction
Understanding hemodynamics—the study of blood flow—within the heart’s left ventricle is vital for the early detection of cardiac diseases. However, data from simulations or medical imaging is often high-dimensional and noisy. To extract meaningful patterns, researchers employ Reduced-Order Modeling (ROM). For decades, Proper Orthogonal Decomposition (POD) has been the benchmark, providing linear, energy-ranked, and perfectly independent (orthogonal) modes. Recently, Autoencoders (AEs) have been proposed as nonlinear successors. This study critically assesses whether these AI-generated modes represent physical reality or become mathematical abstractions that lose interpretability as complexity increases.
Methodology
The researchers conducted a comparative analysis between POD and four specific Autoencoder architectures: Linear (LIN), Non-linear (non-LIN), Convolutional (CAE), and Variational (VAE). The models were evaluated using a high-fidelity Computational Fluid Dynamics (CFD) dataset of an idealized left ventricle, capturing 10 cardiac cycles with Reynolds numbers up to 5500.
Performance was measured using two primary metrics: Relative Root Mean Square Error (RRMSE) for reconstruction accuracy and a custom Orthogonality measure ($O$) to determine the independence of the extracted modes.
Key Insights and Results
The Orthogonality Trade-off
A significant finding is the degradation of mode independence in standard Autoencoders. While POD modes are orthogonal by design, AEs exhibit "spatial redundancy" as the latent space size increases. For example, in the Non-linear AE, the orthogonality measure plummeted to 0.08, indicating a near-total loss of mode independence. This redundancy makes it difficult for clinicians to interpret individual modes as unique physical mechanisms.
The VAE Advantage
The Variational Autoencoder (VAE) emerged as the most robust deep-learning alternative. By regularizing the latent space to follow a Gaussian distribution, the VAE acts as a "physics-friendly" constraint. At a latent dimension of $r=5$, VAE modes qualitatively resembled POD modes, successfully identifying the mean flow and primary oscillations while preserving orthogonality much better than standard CAEs.
Reconstruction Accuracy
Surprisingly, for this specific intraventricular flow, the linear simplicity of POD proved difficult to surpass. POD achieved a near-perfect reconstruction error of 0.5% as more modes were added. In contrast, nonlinear AEs hit a performance plateau; the Convolutional AE, despite its spatial awareness, remained stuck at a 22.3% RRMSE. This suggests that more layers do not inherently translate to better physical reconstruction in fluid dynamics.
Conclusion
While Autoencoders offer powerful nonlinear compression, they often sacrifice the clean orthogonality and high accuracy provided by classical POD in cardiac flow applications. The VAE shows promise in balancing compression with physical interpretability. Future work will likely focus on Physics-Informed Neural Networks (PINNs) to embed the laws of fluid mechanics directly into the model's loss function, ensuring that extracted modes remain physically valid even in complex, asymmetrical patient geometries.
Summary updated