SE PhD Final Defense: Saeed Mohammadzadeh

SE PhD Final Defense: Saeed Mohammadzadeh

TITLE: Advances in Data-Driven Modeling for Computational Mechanics and Biomechanics

ADVISOR: Emma Lejeune ME

COMMITTEE: Chair: Hua Wang ME, SE; Pirooz Vakili, SE; Harold Park, ME; Abigail Plummer, ME

ABSTRACT: This dissertation investigates the challenges impeding the integration of deep learning into computational mechanics and biomechanics, and it offers innovative solutions to address these gaps. The work begins by identifying the scarcity of open-source benchmark datasets in mechanics—a limitation that has hindered progress compared to fields like computer vision. To remedy this, a novel dataset is developed using finite element-based phase-field fracture modeling to simulate complex crack propagation in heterogeneous materials, creating a unique challenge that drives the development of specialized neural network architectures. Beyond dataset creation, the dissertation addresses the prediction of full-field quantities (e.g., displacement, damage, and strain fields) across computational domains, a relatively unexplored area in mechanics. The research then shifts focus to the crucial issue of model calibration. Through extensive comparative analyses across seven mechanics-specific datasets, the study evaluates both post-training calibration methods like temperature scaling and training-time techniques such as ensemble model training. Findings indicate that ensemble averaging notably improves model calibration and predictive reliability, which is critical for uncertainty quantification in computational mechanics. The dissertation further extends its scope to biomechanics by tackling the complexities of analyzing human induced pluripotent stem cell-derived cardiomyocytes (hiPSC-CMs). Given the challenges posed by the structural immaturity and heterogeneity of these cells, an open-source Python toolkit named SarcGraph is introduced. Leveraging self-supervised deep learning methods, SarcGraph enables precise, scalable, and automated analysis of sarcomere organization, as demonstrated on an openly available hiPSC-CM imaging dataset. Collectively, the work not only overcomes key integration challenges but also paves the way for a promising future where deep learning fundamentally advances both computational mechanics and biomechanics through collaborative and open-source research practices.