Machine Learning to Predict Damage Mechanics in Trabecular Bone

点击翻译按钮获取中文摘要

Bone is a structurally complex biologic tissue, its mechanics are greatly affected by damage generated under physiologic and high impact loads. Characterizing bone fracture nucleation, propagation and stability of damaged bone tissue are challenging due to bone's complex ultrastructure, heterogeneous distribution of material, and the ubiquity of damage and flaws. Bone damage mechanics have been modelled deterministically with a variety of labour and computationally intensive formulations, but with only moderate agreement to experimental damage deposition. The effort to build such models has limited their optimization and application. The goal of this research program is to build hybrid machine learning (ML)-physics based models to better predict damage behaviour of structurally complex composite materials, with a focus on trabecular bone as a model system. Deep learning (DL) has many advantages for predicting damage and mechanical behaviour within tissues with complex ultrastructure: multi-scale effects can be integrated (through convolutional layers), generative deep models can represent the stochastic character of damage, the methods are tolerant of noise (typical of biological and damaging structures), and DL can incorporate non-linear relationships of material properties (often present in biological tissues). Once trained DL models can provide a computationally efficient and simple method for studying physical systems. Objectives: 1.Develop hybrid DL-physics models for the study of trabecular bone damage mechanics. 2.Quantify the non-linear interactions of material properties and multi-scale structure that affect normal damage mechanics of trabecular bone tissue. Interactions of tissue properties (mineralization, collagen crosslinking, hydrogen and covalent bonding within the collagen backbone) will be specifically investigated by ex vivo mechanical testing with damage labelling within µCT imaging. Hybrid DL-physics based models will be constructed combining neural networks with simplified linear-elastic finite element analysis and beam theory to predict apparent material behaviour and damage deposition. The structure will include networks for preprocessing (boundary conditions, material distribution, and geometry), post processing (damage deposition, and post-yield behaviour) and error correction neural networks to adjust simplified physics-based modelling results. This research will improve understanding of bone quality, allowing better prediction of tissue damage and mechanical stability. The tools developed will create a platform to study other composite engineered and biological materials with complex ultrastructure and a training platform for HQP in physics-based modeling and machine learning with applications in industry and academia. The methods for creating hybrid DL-physics models are also applicable to other areas of science where model formulation is challenging, noise complicates predictions, and large data sets exist.