NeuralClothSim: Neural Deformation Fields Meet the Thin Shell Theory (2308.12970v3)

Abstract: Despite existing 3D cloth simulators producing realistic results, they predominantly operate on discrete surface representations (e.g. points and meshes) with a fixed spatial resolution, which often leads to large memory consumption and resolution-dependent simulations. Moreover, back-propagating gradients through the existing solvers is difficult, and they cannot be easily integrated into modern neural architectures. In response, this paper re-thinks physically plausible cloth simulation: We propose NeuralClothSim, i.e., a new quasistatic cloth simulator using thin shells, in which surface deformation is encoded in neural network weights in the form of a neural field. Our memory-efficient solver operates on a new continuous coordinate-based surface representation called neural deformation fields (NDFs); it supervises NDF equilibria with the laws of the non-linear Kirchhoff-Love shell theory with a non-linear anisotropic material model. NDFs are adaptive: They 1) allocate their capacity to the deformation details and 2) allow surface state queries at arbitrary spatial resolutions without re-training. We show how to train NeuralClothSim while imposing hard boundary conditions and demonstrate multiple applications, such as material interpolation and simulation editing. The experimental results highlight the effectiveness of our continuous neural formulation. See our project page: https://4dqv.mpi-inf.mpg.de/NeuralClothSim/.

• The paper introduces NeuralClothSim, a framework that encodes cloth deformations in a continuous neural network guided by Kirchhoff-Love theory.
• It integrates a displacement MLP with sine activation functions to accurately model high-frequency folds and wrinkles under prescribed boundary conditions.
• The method achieves memory-efficient, adaptive simulation consistency, paving the way for real-time applications in gaming and graphics.

Overview of Neural Deformation Fields for Cloth Simulation

This paper introduces an innovative approach to cloth simulation through the development of NeuralClothSim, a method proposing a novel integration of neural deformation fields (NDFs) within the context of the Kirchhoff-Love thin shell theory. This approach marks a shift in cloth simulation from the conventional explicit geometric representations and time-stepping methods to a continuous, differentiable, coordinate-based implicit neural representation of surface deformations.

Main Contributions

The key contributions of this research are:

1. Neural Representation of Surface Evolution: NeuralClothSim encodes the deformation of cloth surfaces in the weights of a neural network, using NDFs that allow for querying simulated states continuously in space and time.
2. Integration of Kirchhoff-Love Theory: The method supervises these NDFs using the nonlinear Kirchhoff-Love shell theory. This integration allows for the simulation of large-scale deformations and realistic cloth behaviors under various external forces and boundary conditions.
3. Adaptive and Consistent Simulation: The proposed method features an adaptive design, enabling consistent query capabilities across different spatial and temporal resolutions without the need for retraining the neural network.

Technical Approach

Formulation and Implementation

NeuralClothSim introduces a memory-efficient architecture employing a continuous neural network-based representation of dynamic surfaces supervised by thin shell theories. This approach involves intricate optimization of a displacement field MLP, capturing the surface evolution of cloth under varied conditions such as material properties and external disturbances. The method employs periodic sine activation functions to represent high-frequency deformations like folds and wrinkles effectively.

Handling Boundary and Initial Conditions

The paper proposes embedding boundary conditions directly into the neural network through the definition of spatial and temporal distance functions. This ensures that boundary conditions are enforced rigorously, mitigating issues related to unbalanced gradient descent often encountered in neural models driving simulation conditions.

Differentiability and Simulation Editing

NeuralClothSim is inherently differentiable, a feature leveraged for optimizing the dynamic system's energy and ensuring consistent simulations across different discretization levels. Furthermore, the model supports flexible editing post-training, making it versatile for rapid adjustments in simulation scenarios—a significant advantage for applications in gaming and real-time graphics.

Evaluation and Implications

Validation and Comparisons

NeuralClothSim was tested using established benchmarks like the Belytschko obstacle course, demonstrating close alignment with analytical solutions. Comparisons with existing state-of-the-art methods highlight NeuralClothSim's advantage in memory efficiency and simulation consistency, particularly in terms of maintaining continuity across varying resolutions.

Potential Applications and Future Directions

While the method currently doesn't address contacts, friction, and collisions, its differentiability and adaptive structure present a robust stepping stone for future research. Potential application areas include enhanced garment simulation in movie production, real-time deformation modeling in gaming, and integration as an inverse model in machine vision tasks. The proposal of continuously modeling material properties further widens its applicability, allowing seamless interpolation and optimization in graphic simulations.

Conclusion

NeuralClothSim presents a compelling development in cloth simulation, effectively utilizing neural networks to refine and enhance the simulation process. By marking a departure from traditional mesh-based methods, NeuralClothSim introduces a new paradigm where simulations are adaptive, consistent, and seamlessly integrated with machine learning frameworks. Although future work is necessary to extend functionality, especially regarding collision and friction modeling, the framework lays a foundational path for developing more adaptable and integrated simulation technologies in computer graphics and beyond.