Content area
This dissertation investigates the synergistic integration of deep learning with physics-based simulations —not as a replacement for physics solvers, but as an augmentation that enhances efficiency, controllability, and scalability. Physics-based simulations are indispensable in science, engineering, and entertainment, enabling the virtual reproduction of complex physical phenomena ranging from fluid dynamics and solid mechanics to human motion and facial animation. Despite their accuracy, these simulations often incur high computational costs, limiting their use in interactive applications, rapid prototyping, and real-time engineering workflows.
Recent advances in deep learning present opportunities to accelerate simulations, while preserving physical fidelity. Across multiple domains, including controlled modeling of active quasistatic soft bodies (such as the human face), accelerated simulation via super-resolution in both soft-bodies (specifically, human faces) and splashy particle-based fluid simulations, and support for fast numerical solvers in fluidic topology optimization, this work develops hybrid methods in which simulations inform learning and learning enriches simulations, fostering a mutually beneficial relationship between the two. Because physics simulators inherently model the laws of nature, integrating them into learning frameworks improves generalization, even when data is sparse or irregularly sampled. By embedding differentiable physics layers into neural architectures, enforcing physically motivated modeling and constraints, and using simulation outputs as structured priors, we demonstrate that physics-guided deep learning can achieve both accuracy and speed in data-limited regimes. The approaches presented here advance the frontiers of simulation-assisted learning, improving fidelity and opening new pathways toward real-time, controllable, and physically grounded applications.