Designing mechanically efficient geometry for architectural structures like shells, towers, and bridges is an expensive iterative process. Existing techniques for solving such inverse mechanical problems rely on traditional direct optimization methods, which are slow and computationally expensive, limiting iteration speed and design exploration. Neural networks would seem to offer an alternative, via data-driven amortized optimization for specific design tasks, but they often require extensive regularization and cannot ensure that important design criteria, such as mechanical integrity, are met. In this work, we combine neural networks with a differentiable mechanics simulator and develop a model that accelerates the solution of shape approximation problems for architectural structures. This approach allows a neural network to capture the physics of the task directly from the simulation during training, instead of having to discern it from input data and penalty terms in a physics-informed loss function. As a result, we can generate feasible designs on a variety of structural types that satisfy mechanical and geometric constraints a priori, with better accuracy than fully neural alternatives trained with handcrafted losses, while achieving comparable performance to direct optimization, but in real time. We validate our method in two distinct structural shape-matching tasks, the design of masonry shells and cable-net towers, and showcase its real-world potential for design exploration by deploying it as a plugin in commercial 3D modeling software. Our work opens up new opportunities for real-time design enhanced by neural networks of mechanically sound and efficient architectural structures in the built environment.
翻译:暂无翻译