Multi-fidelity learning-based latent diffusion model for three-dimensional inverse microstructure design of dual phase steels

• Multi-fidelity training of generative models for inverse microstructure design was investigated. • The latent diffusion model was pretrained using low-fidelity data. • Low-rank adaptation (LoRA) drastically increased the model performance with number of limited data. • The LoRA fine-tuned model reproduced target properties with an average accuracy above 98 %. • The model was capable of generating alternative microstructures that achieved target properties. Microstructure plays a critical role in determining the mechanical performance of structural materials, which is mainly why microstructure optimization is a key objective in the design of metallic materials. Data-driven approaches have emerged as efficient alternatives to traditional trial-and-error methods. In particular, generative models have been an attractive solution for inverse microstructure design due to their one-step prediction capability. However, generative models, particularly those that are relatively large, require large volumes of data. This data requirement increases drastically when dealing with 3D microstructure data. This study presents an approach that adopts multi-fidelity training a diffusion-based generative model using low-rank adaptation (LoRA) to design 3D microstructures with targeted mechanical properties. Low-fidelity finite element simulation data are used to pretrain the model, which is later fine-tuned using LoRA with a small set of high-fidelity simulations whose parameters are calibrated with experimental data. The results demonstrate that this approach significantly reduces data and computational costs while maintaining high performance. Evaluations using high-fidelity simulation confirm that the fine-tuned model accurately reproduces target stress–strain behavior. This work provides an efficient and scalable solution to train generative models for inverse microstructure design in 3D.