Point Cloud Driven Strain Prediction in Precision Assemblies via Stress-Informed Sampling and Self-Supervised Learning

During the modeling of precision-assembled structures, compliant members frequently display markedly nonuniform stress fields and complex deformations, and conventional approaches are notably limited in feature representation and learning under small-data regimes. In response, we present a strain prediction framework for flexible structures that combines stress-informed point cloud downsampling with self-supervised pretraining. We begin by designing a weighted K-means downsampling scheme that fuses stress amplitude and gradient cues, achieving substantial compression while retaining salient structural characteristics. With the reduced point cloud, we pretrain using the unsupervised Point-UMAE under a masked autoencoding objective to learn multi-scale geometric-semantic representations. Finally, we perform regression finetuning with a small amount of labeled data to achieve high-accuracy prediction of strain distributions in complex structures.