<p>Accurate prediction of stress state is of great importance for evaluating the plastic deformation and ductile fracture of metallic materials. This study aims at developing a deep learning framework for predicting stress field and triaxiality in dual-phase (DP) advanced high-strength steels (AHSSs). The framework is introduced via crystal plasticity-based convolutional neural network (CP-CNN) that is capable of rapidly and accurately predicting von Mises equivalent stress and stress triaxiality distributions in the generated microstructure of DP steels with varying martensite fractions. This work introduces two approaches involving unique-input and multi-input features with encoder-decoder architecture. It is found that the former approach possesses excellent agreement of stress field with CPFEM simulations, meanwhile the latter approach demonstrates a significant improvement in prediction of stress triaxiality. This study discloses that an integrated squeeze-excitation residual network (SE-ResNet) and self-attention modules together within the network enable the CP-CNN to effectively describe both local feature refinement and global feature dependencies, which are crucial in capturing stress heterogeneity in the DP steels. In particular, the stress partitioning among the constituent phases of ferrite and martensite is accurately predicted. Moreover, within the investigated synthetic DP microstructure, the predictive capability and generalizability of the developed CP-CNN are well demonstrated on various unseen conditions, such as varying grain orientations and martensite fractions, leave-one-martensite-fraction-out validation, and additional strain levels. The present findings elucidate an obvious merit of CP-CNN that the computational time is significantly reduced by ~ 2.5 orders of magnitude compared to the CPFEM simulation. The present CP-CNN framework offers significant potentials for materials design and development of AHSS.</p>