Integrating multimodal dot product fusion with conditional generative adversarial network to build a C-P-S-P chain in Mg alloys

Establishing quantitative relationships among composition, process, structure, and property (C-P-S-P) is a fundamental challenge in materials science, yet conventional models relying solely on numerical data fail to capture microstructural spatial heterogeneity, limiting prediction accuracy. Here, we propose an integrated framework combining dot product fusion multimodal strategy with multidimensional continuous conditional generative adversarial network (MCCGAN) to construct the complete C-P-S-P chain for Mg-Al-Sn alloys. A multimodal convolutional neural network (CNN) employing dot product fusion of composition (Al, Sn), process parameters (solid-solution temperature and time), and Electron Backscatter Diffraction (EBSD) phase maps generated by MCCGAN achieves high ultimate tensile strength (UTS) prediction accuracy with a coefficient of determination ( R 2 ) of 0.928 and root mean square error (RMSE) of 3.241 MPa, substantially outperforming numerical models. To overcome experimental data scarcity, MCCGAN generates physically consistent virtual microstructures conditioned on continuous composition-process variables, and the morphological fidelity of these generated structures is evaluated by low Fréchet Inception Distance (FID) scores, while their physical consistency is validated through accurate phase fraction reproduction. By integrating MCCGAN-generated microstructures with the multimodal predictor, we construct comprehensive response surfaces across the entire design space revealing the synergistic competition between solid-solution and secondary phase strengthening mechanisms across the design space. This framework provides a scalable pathway for accelerated alloy design by bridging sparse experimental data with continuous optimization demands.