Microstructure generation of Ni-based single-crystal superalloys across continuous composition spaces via a latent diffusion transformer

In Ni-based single-crystal superalloys, the pronounced microstructural sensitivity to subtle compositional variations within a vast multicomponent space renders experimental establishment of composition–microstructure mappings infeasible. In this work, using phase-field simulation data of the Ni–Al–Mo system, we develop a Composition-Conditioned Latent Diffusion Transformer (CCLDT) for the generalized generation of microstructure images within the continuous composition space. Built on the Latent Diffusion Model (LDM) framework, CCLDT employs a Diffusion Transformer (DiT) as its denoising backbone. Through the optimization of its compositional conditioning mechanism, CCLDT yields mean Intra-FID scores of 11.31 and 18.70 under seen and unseen compositions, respectively, demonstrating superior visual fidelity in the learned generative mapping. Furthermore, for each physical statistical descriptor, over 75% of the points show low deviations: specifically, absolute errors below 0.02 for γ′ volume fraction and relative errors below 6% for the autocorrelation function (ACF), thereby validating the model’s strong physical consistency within the continuous composition space. Importantly, CCLDT highlights Mo’s dominant role and its synergistic interaction with Al, consistent with physical analyses based on lattice misfit and energy balance considerations. In addition, the alignment of Intra-FID with physical metrics demonstrates its effectiveness in microstructural assessment. To the best of our knowledge, this work presents the first application of a DiT-based latent diffusion model for Ni-based single-crystal superalloys, enabling generative mapping from continuous composition space to microstructures.