Title: PINNOCHIO: A Physics-Informed Neural Network Framework for Coupled Hyperelastic Interface-Volume Simulation in Orthognathic Surgery

Abstract

Accurate prediction of patient-specific facial soft-tissue deformation is essential for refining orthognathic surgical plans. However, existing computational approaches are hindered by a rigid accuracy-efficiency trade-off. High-fidelity Finite Element Methods (FEM) are often too computationally expensive to be practical, while pure deep learning models frequently yield results that lack biomechanical consistency. Although Physics-Informed Neural Networks (PINNs) present a viable alternative, training them to capture the complex, heterogeneous mechanics of bone-soft-t tissue interactions proves highly unstable when relying solely on partial clinical supervision, such as data from outer facial surfaces.

To address these limitations, we introduce PINNOCHIO, a novel physics-informed framework designed for facial soft-tissue simulation. PINNOCHIO employs a hybrid sequential decomposition that explicitly separates the discontinuous movements at the bone-soft-tissue interface from the continuous volumetric hyperelastic deformation. This architectural distinction promotes stable training and supports a physics-enabled strategy for sim-to-real adaptation, thereby guaranteeing internal biomechanical consistency without the need for volumetric ground truth data.

In evaluations conducted on a clinical cohort of 40 patients, PINNOCHIO surpassed existing baseline methods in terms of both surface accuracy and physical validity. Additionally, the framework delivers a significant performance speedup compared to FEM, effectively resolving the longstanding accuracy-efficiency dilemma. By offering a highly reliable and practical solution, PINNOCHIO serves as an effective tool for interactive surgical planning.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC