Electrospinning Machine | A versatile multimodal learning framework bridging multiscale knowledge for material design

Recent years have witnessed the widespread application of Artificial Intelligence (AI) in materials science, greatly accelerating the discovery process of new materials. Nonetheless, real-world material systems exhibit multiscale complexity, spanning composition, processing, structure, and properties. Moreover, the data is often fragmented, with structural characterization data being particularly scarce due to its prohibitively high cost of acquisition. This absence of critical information poses a formidable barrier to the modeling and generalization of multimodal AI.

Recently, a research team led by Professor Ji Jian and Researcher Zhang Peng from Zhejiang University have published their latest findings in the journal npj Computational Materials, entitled "A versatile multimodal learning framework bridging multiscale knowledge for material design." The researchers proposed a multimodal learning framework, MatMCL, which employs a "structure-guided contrastive learning" strategy to achieve highly robust performance prediction even with incomplete modalities, while also enabling cross-modal generation and retrieval.

Figure 1: Overall workflow of the MatMCL framework.

The research team fabricated a multimodal dataset comprising processing conditions, micro-nano structures, and mechanical properties based on electrospun nanofiber materials. They then mapped the microstructures and processing conditions into a shared latent space. Through a contrastive learning mechanism, the model was enabled to identify the underlying relationships between processing and structure.

Figure 2: Mechanical property prediction.

Even when structural information was absent, MatMCL demonstrated compelling predictive power and retained an ability to perceive structural features of the fibers. Its performance significantly surpassed that of conventional methods across multiple mechanical property metrics.

Figure 3: Cross-modal retrieval.

The results confirm that the model can effectively retrieve corresponding structures based on processing conditions, and vice versa. In summary, MatMCL has demonstrated a robust capability in elucidating the complex correlations within multimodal material data, underscoring its significant potential for advancing materials understanding and design.

Figure 4: Structure generation.

MatMCL incorporates a conditional generation module based on a diffusion model, enabling it to reliably generate high-quality microstructures from given processing conditions. The generated results are virtually indistinguishable from real samples, demonstrating powerful cross-modal understanding and structural generation capabilities. This provides a new pathway to overcome database limitations and realize intelligent material design.

Figure 5: Nanofiber-reinforced composite design.

The research team further extended the framework to the design of nanofiber-reinforced composites. By employing a multi-stage learning strategy, they achieved high-precision prediction even with extremely limited samples. This success in addressing the dual challenges of "few-shot learning and missing modalities" demonstrates the model's exceptional adaptability and generalizability for real-world material design. 

This study provides a novel and universal AI solution to the long-standing challenges in materials science concerning multiscale, multimodal, and incomplete data, exhibiting high flexibility and scalability.

Paper link: https://www.nature.com/articles/s41524-025-01767-3