Deep learning acceleration of Total Lagrangian Explicit Dynamics for soft tissue mechanics

• Published in: Computer methods in applied mechanics and engineering Vol. 358; p. 112628
• Main Authors: Meister, Felix, Passerini, Tiziano, Mihalef, Viorel, Tuysuzoglu, Ahmet, Maier, Andreas, Mansi, Tommaso
• Format: Journal Article
• Language: English
• Published: Amsterdam Elsevier B.V 01.01.2020; Elsevier BV
• Subjects: Acceleration; Accuracy; Biomechanics; Computer graphics; Computer simulation; Data-driven computational modeling; Deep learning; Finite element method; Neural networks; Real time; Soft tissue biomechanics; Soft tissues; Time integration; Total Lagrangian Explicit Dynamics; Deep learning; Finite element method; Data-driven computational modeling; Soft tissue biomechanics; Total Lagrangian Explicit Dynamics
• ISSN: 0045-7825; 1879-2138
• DOI: 10.1016/j.cma.2019.112628

Simulating complex soft tissue deformations has been an intense research area in the fields of computer graphics or computational physiology for instance. A desired property is the ability to perform fast, if not real-time, simulations while being physically accurate. Numerical schemes have been explored to speed up finite element methods, like the Total Lagrangian Explicit Dynamics (TLED). However, real-time applications still come at the price of accuracy and fidelity. In this work, we explore the use of neural networks as function approximators to accelerate the time integration of TLED, while being generic enough to handle various geometries, motion and materials without having to retrain the neural network model. The method is evaluated on a set of experiment, showing promising accuracy at time steps up to 20 times larger than the “breaking” time step, as well as in a simple medical application. Such an approach could pave the way to very fast but accurate acceleration strategies for computational biomechanics. •Deep learning acceleration of the Total Lagrangian Explicit Dynamics (TLED).•Acceleration learned from the kinematic and dynamic state of the system.•Our method enables to use time steps beyond the time step limit of the explicit solver.•Generalization to motions, geometries and tissue materials unseen in training.•Simulation of liver motion is illustrated as a simple example of human anatomy.