Investigation of machine learning and deep learning methodologies for applications in biomechanics, additive manufacturing, and microstructure fingerprinting: A Biomaterialomics approach

Abstract

Traditional approaches to developing biomaterials and implants require intuitive tuning of process variables, long innovation cycle, and high costs. Accelerating the production of personalized, implantable biomaterials and biomedical devices is critical to meet the unmet clinical needs. In this context, it will be demonstrated in this thesis colloquium presentation that a data-science approach can effectively integrate computational tools, databases, experimental methods, machine learning (ML), and advanced manufacturing on a single platform to develop fourth-generation biomaterial implants, whose microstructure or properties or performance can be predicted suing digital twins. In biomedical engineering, additive manufacturing (AM) has attracted significant attention in the scientific community, because of its perceived potential to fabricate patient-specific implants. Directed Energy Deposition (DED) and Laser-Powder Bed Fusion (L-PBF) are two of the many variants of AM techniques that have piqued the interest of biomaterials researchers. Furthermore, machine learning models were developed for a better holistic understanding of AM processes, wherein they are used to find the relationship between the process parameters and geometrical features of the 3D-printed single tracks. In particular, the process maps for the L-PBF of SS316 using single tracks are constructed. We conducted multi-parameter optimization to optimize the density and roughness of 3D printed coupons and to create a process parameter recommendation system, which was used to fabricate femoral stem prototypes. The design of musculoskeletal implants, together with the human subject weight and bone conditions are known to influence the biomechanical response. A significant part of this thesis was focused on to develop machine learning-based approaches in tandem with parametric Finite Element Analysis (FEA) to predict the periprosthetic biomechanical response in the acetabulum after Total Hip joint replacement, while establishing the potential of such approach as a fast surrogate of FEA for implant biomechanics analysis for strain prediction in a few seconds. In addition, we utilized deep learning models, specifically generative adversarial networks (GANs), to generate the synthetic microstructural images of titanium alloy. Different quantitative metrics and morphological parameters were used to assess the model's performance and the distribution of microstructural features between real and synthetic images. The adaptive augmentation approach helps reduce overfitting and leaking of training data into generated images and, most importantly, performs well in the limited data regime (~ 1000 images or less), which is the case in metallurgical process design studies. This work has shown the increasing adaptability and integrated formulation of Artificial Intelligence (AI) tools in biomaterials science will enable the adoption of interdisciplinary approaches to establish links between processing, structure, and properties (PSPs) and illustrate the formulation and relevance of the 'biomaterialomics' approach to new research topics, patient-specific implants, and additive manufacturing.