MG-LDM: A multimodal guided latent diffusion model with Mamba-based temporal encoding for inverse topological design of tissue engineering skin substitutes

Tissue engineering full-thickness skin substitutes (FSS) with clinically relevant mechanical, structural, and biological characteristics remain a major challenge in regenerative medicine. To address this, we propose a multimodal guided latent diffusion model with Mamba-based temporal encoding (MG-LDM) as a unified inverse design framework for the topology optimization and additive manufacturing of 3D bio-printed FSS. A high-resolution multimodal dataset was constructed, consisting of stress–strain sequences, seven-channel three-dimensional stress field distributions, and extrusion-based 3D printing parameters annotated with cell viability metrics. MG-LDM integrates a U-shaped Mamba encoder for temporal sequence modeling, a densely connected graph convolutional network (DC-GCN) for spatial feature extraction, and a multi-layer perceptron (MLP) encoder for processing manufacturing parameters. These heterogeneous representations are fused via a guided cross-attention mechanism into a unified latent condition, which drives a diffusion-based structure generator. A dual-path topology generation strategy incorporating a neural signed distance function (SDF) ensures geometric continuity and mechanical fidelity. Experimental evaluations indicate that MG-LDM consistently outperforms representative baselines in terms of geometric accuracy, structure–function consistency, and robustness across multiple modalities. Physical validations confirm that MG-LDM enables the generation of biocompatible and mechanically controllable FSS with enhanced structural integrity and regenerative performance, supporting its applicability in personalized regenerative medicine.