Multi-Objective Evolutionary Neural Architecture Search for Material Microstructure Segmentation

Accurately characterizing microstructures is crucial for material design and discovery. However, the architecture design of existing deep learning models for microstructure segmentation still heavily relies on empirical knowledge or manual tuning, limiting their optimization efficiency and adaptability. To address this limitation, this paper proposed a self-adaptive modeling framework for material microstructure segmentation based on multi-objective evolutionary neural architecture search (MOENAS). Specifically, we defined different types of operation sets and two types of dynamic configuration units (Encoder cell and Decoder cell) to construct a UNet-like segmentation network tailored to microstructure characteristics. To address the irregularity and scale variations of microstructural morphology, we further proposed the deformable multi-scale aggregation (DMSA) module, which is embedded between Encoder cells to enhance the network ability to perceive deformations and extract multi-scale features effectively. Furthermore, to improve search space flexibility, we developed an adaptive population individual encoding scheme capable of representing the internal operations and structures of both Encoder and Decoder cells. During the search process, we employed a Non-dominated sorting differential evolution algorithm (NSDE) with knee domination to simultaneously optimize both cells, achieving a trade-off between segmentation accuracy and computational complexity. Experimental results on a material microstructure dataset demonstrate that MOENAS delivers highly competitive performance while significantly reducing the search cost.