Advancing deep learning in computational mechanics and biomechanics: overcoming challenges and paving a promising future
OA Version
Citation
Abstract
Deep learning has revolutionized numerous scientific fields, yet its integration within computational mechanics and biomechanics remains limited. This is primarily due to unique challenges including but not limited to the scarcity of suitable benchmark datasets, uncertainties regarding model reliability, and complexities inherent to biological systems. This dissertation systematically identifies and investigates these challenges, adopts methodologies from the broader deep learning and computer vision communities, develops robust computational solutions, and openly shares datasets, tools, and analytical results, enabling and inspiring continued advancements within the research community. The first critical challenge addressed is the notable absence of open-source benchmark datasets within mechanics. In computer vision, benchmark datasets have been fundamental to driving deep learning innovation, fueling the development of increasingly effective algorithms and methods. However, such datasets have historically been lacking within mechanics, significantly limiting progress in applying deep learning methods to this field. To bridge this gap, a novel and challenging dataset was developed using finite element-based phase-field fracture modeling to simulate complex crack propagation in heterogeneous materials. This dataset introduces a unique, rigorous challenge for deep learning—distinct from conventional computer vision tasks such as object recognition or semantic segmentation—thus encouraging innovative model development within the mechanics community. Subsequently, this dissertation tackles the prediction of full-field quantities of interest, such as displacement, damage, and strain fields throughout the entire computational domain, an area comparatively underexplored in the mechanics community. Through this effort, robust neural network architectures are benchmarked, and both the dataset and baseline performance scores are openly shared to facilitate further improvements by other researchers. Although deep learning provides powerful tools capable of uncovering complex patterns and achieving impressive predictive accuracy, these models frequently suffer from inadequate calibration, i.e., the alignment between predicted outcome probabilities and observed occurrences. Model calibration is especially crucial in computational mechanics, where substantial attention is traditionally placed on uncertainty quantification and model reliability. To systematically address this critical issue, extensive investigations were conducted into deep learning model calibration. Rigorous comparative analyses were performed on seven distinct mechanics-specific datasets, evaluating both post-training calibration approaches, such as temperature scaling, and training-time calibration techniques, specifically ensemble model training. Empirical findings from these studies clearly demonstrated that ensemble averaging significantly enhances calibration performance and predictive reliability, directly benefiting applications where accuracy and uncertainty quantification are paramount. Transitioning from simulated mechanical systems to living biological tissues, this dissertation then addresses computational challenges associated with analyzing human induced pluripotent stem cell-derived cardiomyocytes (hiPSC-CMs). These cells hold substantial promise for advancing cardiac research, disease modeling, and regenerative medicine, yet analyzing their structural organization poses significant difficulties due to their inherent biological complexity and structural immaturity. To enable accurate, scalable, and automated analysis, the dissertation introduces SarcGraph, an open-source Python toolkit explicitly designed for quantitative structural analysis at the sarcomere scale. Leveraging recent advancements in state-of-the-art self-supervised deep learning methods, SarcGraph was enhanced to robustly handle structural heterogeneity and accurately identify sarcomere structures, even in highly disordered or immature cellular samples. Furthermore, the improved capabilities of SarcGraph were demonstrated by analyzing an openly available hiPSC-CM imaging dataset, showcasing its effectiveness in quantifying cellular structural organization and providing insights essential for further development of reliable analysis methods. Taken together, the dissertation addresses critical gaps and challenges at the intersection of deep learning, computational mechanics, and biomechanics, offering practical solutions supported by open datasets, tools, and analyses. This work aims to serve as a foundation, fostering broader adoption of deep learning methods within these scientific communities and promoting continuous innovation through collaborative and open-source research practices.
Description
2025
License
Attribution 4.0 International