Enhancing cloth 3D simulation and object interaction via machine learning-based deformation models

Abstract

Realistic cloth simulation remains a challenging task in computer graphics, especially when simulating intricate interactions between deformable fabrics and complex three-dimensional geometries. Traditional physically based models, such as mass-spring or finite element methods, often suffer high computational cost or numerical instability, limiting their applicability in real-time environments or high-fidelity rendering. In this work, we propose a novel hybrid framework that enhances cloth simulation accuracy and computational efficiency through the integration of machine learning-based deformation prediction. Our method employs a supervised deep learning architecture to approximate the nonlinear dynamics of cloth behavior under varying boundary conditions and interactions with rigid or soft 3D bodies. By training on physically simulated datasets that include diverse cloth types, contact scenarios, and environmental forces, our approach generalizes to unseen geometries and interaction modes. The learned model is embedded within a coarse physical simulation loop to preserve global consistency while accelerating local deformation computations. We evaluate the proposed method using both synthetic benchmarks and real-world datasets. Quantitative results demonstrate a significant reduction in simulation time – up to 60% – while maintaining a high degree of physical plausibility compared to traditional solvers. Qualitative experiments show improved handling of high-frequency wrinkles, collision resolution, and dynamic contact response. This framework paves the way for practical applications in virtual try-on systems, real-time animation, and haptic feedback environments. The proposed method contributes to the growing field of differentiable simulation and demonstrates the potential of data-driven models in addressing the limitations of conventional cloth physics engines. Future work includes extending the model to support multi-layered garments, anisotropic materials, and adaptive resolution strategies.