Strokes2Deform: Physics-informed learning of deformation fields on 3D stroke clouds

Abstract

In the 3D architectural design process, deformation analysis of a designed object is a fundamental step. However, such a step is usually inaccessible to designers in the early sketching phase, limiting designers’ ability to reason about the shape’s structural stability. To bridge this gap, we propose Strokes2Deform, an end-to-end learning-based model that enables deformation-aware 3D sketching, requiring neither (surface) reconstruction nor (physical) simulation. Our physics-informed neural network takes as input a 3D sketch stroke cloud and user-specified boundary conditions, and directly predicts the induced deformation field on the 3D stroke cloud. Key to our method is: (i) a synthetic dataset of 40K 3D sketch–deformation pairs of architectural thin-shell structures and their corresponding deformation fields derived from Finite Element (FE) analysis, (ii) the use of fVDB to encode 3D sketch stroke clouds and their per-point features into sparse voxel grids with attributes, and leveraging its differentiable splatting and sampling operators for bidirectional point–voxel mappings, (iii) a dual-head neural network architecture that decouples deformation field prediction into unit-length displacement vectors and scalar displacement magnitudes, and (iv) our physics-informed loss functions derived from thin-shell deformation principles. We validate our method through extensive experiments, demonstrating accurate deformation prediction across sketches of varying complexity and strong agreement with reference deformations derived from FE simulations.