New Publication in Nature Communications

Congratulations to Prof Ivor Tsang, Director and Dr Qian Hangwei, Scientist and Mr Li Xinyao, PhD Student from A*STAR Centre for Frontier AI Research (A*STAR CFAR) on the publication of their latest research in the prestigious journal, Nature Communications.

In their paper titled ‘Learning ORDER-Aware Multimodal Representations for Composite Materials Design’, the team introduces ORDER (ORDinal-aware imagE-tabulaR alignment), a novel multimodal pretraining framework that jointly learns from tabular descriptors and microstructure images, enabling physically meaningful representations of composite materials. By preserving property continuity in the latent space, ORDER supports reliable interpolation across sparsely observed designs and improves data-efficient materials understanding, prediction, retrieval, and generation.

Order Pipeline Fig 1. The pretraining pipeline of ORDER

ORDER is built on two key components:

Ordinal-Aware Contrastive Learning: Organises multimodal representations according to material property ordinality, ensuring that materials with similar mechanical behaviour occupy nearby regions in latent space.
Physics-Based Ordinal Surrogates: Leverages domain knowledge to construct surrogate signals without requiring full property annotations, enabling effective pretraining under extreme data scarcity.

By learning robust multimodal representations, ORDER enables more accurate material property prediction while supporting cross-modal retrieval and microstructure generation in limited-data settings.

The research is a key outcome of the A*STAR Career Development Fund (CDF) project and AI for Materials Discovery Grand Challenge, led by Dr Qian and Prof Tsang respectively.

Collectively, these initiatives are advancing AI-driven approaches for next-generation scientific discovery, underscoring A*STAR CFAR’s AI for Science research pillar and its commitment to addressing real-world scientific challenges.

Read the full paper here.

Order Pipeline Fig 1. The pretraining pipeline of ORDER

ORDER is built on two key components:

Ordinal-Aware Contrastive Learning: Organises multimodal representations according to material property ordinality, ensuring that materials with similar mechanical behaviour occupy nearby regions in latent space.
Physics-Based Ordinal Surrogates: Leverages domain knowledge to construct surrogate signals without requiring full property annotations, enabling effective pretraining under extreme data scarcity.

By learning robust multimodal representations, ORDER enables more accurate material property prediction while supporting cross-modal retrieval and microstructure generation in limited-data settings.

The research is a key outcome of the A*STAR Career Development Fund (CDF) project and AI for Materials Discovery Grand Challenge, led by Dr Qian and Prof Tsang respectively.