Abstract
Cardiovascular disease remains the leading cause of mortality worldwide, and timely diagnosis depends on imaging tools that can reveal the mechanical state of the heart in real time. Echocardiography is the primary bedside imaging modality for cardiac assessment, but it only provides sparse, one-dimensional Doppler velocity measurements whose accuracy is highly sensitive to probe position and beam angle. Reconstructing full three-dimensional blood flow fields for the left ventricle from sparse 1D velocity measurements is an ill-posed problem, where current traditional clinical methods are insufficient.
This report documents contributions made during the Spring 2026 semester in the Kong Lab at Washington University in St. Louis. The work focuses on a physics-informed latent diffusion model (LDM) that reconstructs 3D left ventricular (LV) velocity and pressure fields conditioned on sparse velocity measurements.
The research contributions made during this semester are as follows. First, a systematic parameter sweep of 31 experiments across 6 categories is conducted, which investigates how different scaling strategies for the three physics loss components affect generation quality during DDIM sampling with 200 denoising steps. The key finding is that applying boundary condition loss alone with a linear ramp-down schedule from 0.1 to 0.01 achieves the lowest total loss across all tested configurations.
Additionally, several engineering and infrastructure contributions were completed. Two automation scripts were developed: map_fortran_to_json.py, which translates Fortran genBC parameters and initial conditions into svZeroDSolver JSON format with full M/L/T unit scaling, and a four-script pipeline that converts legacy svFSI .inp input files to the svMultiPhysics XML format. Furthermore, svMultiPhysics and svZeroDSolver were built from source on WashU's compute1 HPC cluster, and an LV simulation environment was established with batch runs for 100+ cases. This semester's work lays the foundation for the Summer 2026 SURGE research project, which will extend LVGen by integrating Doppler conditioning into the LDM generation pipeline.
Document Type
Final Report
Class Name
Mechanical Engineering and Material Sciences Independent Study
Date of Submission
5-11-2026
Recommended Citation
Xu, Yuxuan, "Physics-Informed Diffusion Models for Left Ventricular Blood Flow Reconstruction" (2026). Mechanical Engineering and Materials Science Independent Study. 327.
https://openscholarship.wustl.edu/mems500/327