Modern engineering systems require accurate prediction of rigid and flexible body interactions. While multi flexible body dynamics(MFBD) has improved system-level accuracy, traditional DAE solvers face computational limits in high-DOF nonlinear system...
Modern engineering systems require accurate prediction of rigid and flexible body interactions. While multi flexible body dynamics(MFBD) has improved system-level accuracy, traditional DAE solvers face computational limits in high-DOF nonlinear systems. This prevents real-time analysis needed for digital twins and HIL applications. To bridge this gap, a hybrid simulation paradigm that unites physics-based fidelity with data-driven efficiency is proposed. In this dissertation, I introduce the Data Integrated Model Driven Simulation (DIMDS) framework, which replaces computationally expensive flexible and nonlinear components with surrogate models trained on systematically generated design-of-experiments (DOE) data, thereby preserving physical consistency while achieving reduced runtime.
DIMDS unfolds through six stages: component identification and interface definition establish precise input–output bounds (displacement, velocity, orientation ↔ reaction forces, moments); DOE sampling via full-factorial design or Latin hypercube sampling captures global design trends under quasi-static analysis; surrogate modeling employs a toolbox of techniques—polynomial regression, radial basis functions, and multilayer perceptrons—selected and validated to balance interpretability, local smoothness, and nonlinear expressiveness; system integration couples surrogate evaluations into the MFBD solver through coordinate transformations, force and Jacobian assembly, proportional damping, and Newton–Raphson solution, with virtual-body constructs ensuring seamless compatibility with existing contact and control modules; fluid–structure co-simulation leverages the moving-particle semi-implicit method to extend DIMDS to multiphysics scenarios; and comprehensive validation verifies surrogate accuracy and overall performance.
Two representative case studies illustrate the framework’s capabilities. In the first, a laboratory spring is replaced by an interpolation-based surrogate trained on full-factorial grid data, reproducing benchmark flexible-body responses at grid nodes while providing smooth interpolation within the grid. In the second, an 18-dimensional input space of a vehicle suspension’s three interfaces (six degrees of freedom each) is sampled with 1,000 Latin hypercube points, and radial basis function and polynomial regression surrogates are trained and validated at the system level.
Performance evaluation demonstrates that the surrogate-augmented model reduces computation time significantly—while maintaining engineering-level accuracy in vertical, lateral, pitch, and roll responses. Under fluid-structure coupling, computation time also decreases without compromising correlation or tolerances in positional and orientational outputs. Furthermore, radial basis functions excel in local accuracy, whereas polynomial regression delivers superior global stability, guiding surrogate selection based on component characteristics.
This work’s primary contributions are threefold: establishment of the DIMDS six-stage procedure to overcome real-time MFBD limitations; development of a high-throughput data pipeline integrating DOE, quasi-static analysis, sequence optimization, warm-starting, distributed execution, and quality control; and formulation of a unified coupling algorithm—spanning interface definition, coordinate transformation, force/Jacobian integration, and damping—that is fully compatible with commercial MFBD platforms. Demonstration cases confirm two- to three-order-of-magnitude speedups and robust multiphysics interoperability, charting a practical path toward digital twins, design optimization, and interactive simulation. Limitations include surrogate reliability confined to training domains and challenges in high-frequency dynamics extrapolation, motivating adaptive monitoring and domain-expansion retraining strategies. Future work will address fluid surrogate development, multi-fidelity coupling, and improved constraint methods to further enhance metamodel reliability and applicability.