High-throughput adaptive physics refinement for tissue-scale adhesive dynamics

Explicitly resolving ligand–receptor interactions of circulating tumor cells (CTCs) across anatomically realistic vasculature remains computationally prohibitive at the submicrometer fidelity required for adhesive dynamics. In this work, we extend our previously introduced hybrid CPU–GPU adaptive physics refinement (APR) and single-window adaptive physics refinement adhesive dynamics (APR-AD) methods (Martin et al., 2025) into a scalable, high-throughput APR-AD platform. This extension introduces: (1) a wall-resolving multi-window formulation that enables thousands of concurrent adhesive transport simulations, (2) a one-way coupling strategy between bulk and fine domains that eliminates inter-window dependencies while preserving trajectory accuracy, and (3) a communication-free APR mode for steady-state flows that transforms the window phase into an embarrassingly parallel workload. We further present GPU-accelerated kernels for adhesive dynamics, including deterministic random number generation using linear feedback shift registers and octree-based receptor searches optimized for modern exascale systems. Using a double-bifurcating, tissue-scale vessel test case on the Aurora supercomputer, APR-AD tracks 3072 circulating tumor cells in parallel with an approximate 15× reduction in memory relative to a fully explicit model, while maintaining high-fidelity adhesive dynamics. These advances expand APR from a single-cell feasibility tool into a computational microscope for large-scale studies of cancer transport and other receptor-mediated transport phenomena.

Duke Scholars

Author Amanda Randles Biomedical Engineering