Spatially-Informed Online Prediction of Milling Surface Deformation Using Multiphysics-Infused Graph Neural Network for Digital Twinning

Abstract Milling is essential for achieving precise surface flatness and durability, yet residual stress-induced deformation in high-strength alloys, coupled with the computational intensity of traditional finite element methods, poses significant challenges for rapid prediction within the digital twin framework. An efficient and accurate surrogate modeling technique is needed to facilitate process design. An emerging concept is to leverage graph neural networks, which map each mesh node to a graph vertex and reflect finite element connectivity as graph edges, to directly capture both local and global interactions. The limited, existing attempts, however, employ an entire mesh, yielding very large graphs and undermining iterative process assessment. In this work, we propose a physics-infused graph neural network and apply it to the machining process for the first time. Our approach uses a concise, dynamically evolving node graph composed exclusively of critical measurement points, ensuring an effective and accurate surrogate model. To enhance expressiveness, we introduce an attention-based dynamic graph mechanism that adaptively learns strong interactions between these selected nodes. By embedding physics-informed constraints derived from the Johnson–Cook damage model, the proposed approach accurately predicts post-milling deformation aligned with experimental measurement protocols. This will enable rapid surface quality prediction within a digital twin environment, supporting data-driven process planning. Extensive ablation studies on A2024 aluminum alloy demonstrate that integrating node graph features with a physics-informed loss function significantly enhances the model performance and stability, unleashing the full potential of physics-driven surrogate modeling in machining applications.