Stable Long-Horizon Neural ODE Reduced-Order Models via Learned Feedback for Biological Growth and Remodeling

Abstract

Reduced-order models (ROMs) are essential for rapid simulation of complex biomechanical systems and for bridging the gap between high fidelity models and clinical application. However, ROMs for tissue growth and remodeling (G&R) remain largely unexplored. Here, we present a Neural Ordinary Differential Equation (NODE) ROM framework that learns latent dynamics of coupled mechanical deformation and tissue growth, demonstrated in the context of skin growth during tissue expansion (TE). TE is a challenging problem involving nonlinear contact, history-dependent material behavior, and mechanobiology driven growth. The displacement field is compressed via Proper Orthogonal Decomposition (POD) into a low-dimensional latent space, and a NODE learns the resulting dynamics conditioned on patient-specific parameters. To address long-horizon error accumulation, a key challenge in autoregressive latent dynamical models, we propose a closed-loop architecture in which encoded features of the evolving growth field are fed back into the dynamics at each step. We compare feedback representations of increasing expressiveness: scalar, linear POD-based, and nonlinear CNN-based. The CNN-based growth feature feedback substantially stabilizes long-horizon rollouts. The best model captures 90.3% of validation cases within clinical tolerance based on the final skin area gain, compared to 43.7% for the open-loop baseline. Moreover, the NODE ROM achieves over 20000x the speed of full finite element simulations. More broadly, these results suggest that selectively retaining inexpensive physics of the state evolution and feeding features from these fields back into the latent dynamical system is a promising strategy for stable and accurate ROMs of G&R in biological tissues.