Speaker: Aishwarya Pawar, Iowa State University assistant professor mechanical engineering, William March Scholar in Mechanical Engineering

Title: From differentiable splines to digital twins in biomedical applications

Abstract: Differentiable spline modeling is transforming how CAD geometry is integrated with machine learning and simulation. We introduce THB-Diff, a GPU-accelerated framework that embeds truncated hierarchical B-splines directly into differentiable programming environments such as PyTorch. By treating CAD geometry as a native, differentiable layer, THB-Diff enables end-to-end gradient-based shape optimization, local adaptive refinement, and seamless integration of spline evaluation into neural networks. Custom CUDA kernels provide efficient forward and backward passes, delivering significant gains in accuracy and performance for surface fitting, reconstruction, and inverse design. These advances bridge CAD and AI, enabling analysis-suitable geometry to be optimized directly, without mesh approximations, in a unified, learning-enabled pipeline.

Building on this foundation, we demonstrate a high-impact application in biomechanics and digital healthcare. I present ValveFit, a differentiable and GPU-accelerated B-spline framework for patient-specific reconstruction of tricuspid heart valves from 4D echocardiography. Using PDE-constrained, gradient-based optimization, ValveFit rapidly deforms an idealized template to fit sparse and noisy clinical data while enforcing smoothness and physical plausibility. This approach addresses a critical need in managing tricuspid regurgitation risk in children with Hypoplastic Left Heart Syndrome and enables near-real-time generation of simulation-ready cardiac models. Together, these results illustrate how differentiable spline modeling elevates CAD from static design to a powerful engine for AI-driven optimization and predictive digital twins.

Host: Jeff Essner, GDCB professor