万能AIによるサステナブル材料設計～分解性とタフさのトレードオフ解決に迫る新技術～

2025-12-23 理化学研究所

「万能AI」（複数モーダル・タスクを学習）によるサステナブル材料設計

＜関連情報＞

• https://www.riken.jp/press/2025/20251223_1/index.html
• https://www.sciencedirect.com/science/article/abs/pii/S2214993725005494

NMRデータサイエンスを活用した多様な生分解性ポリマーのための同時マルチモーダルおよびマルチタスク戦略 Simultaneous multimodal and multitask strategies for diverse biodegradable polymers powered by NMR data science

Xinyu Ni, Yoshifumi Amamoto, Jun Kikuchi
Sustainable Materials and Technologies  Available online: 19 November 2025
DOI:https://doi.org/10.1016/j.susmat.2025.e01781

Highlights

• A multimodal–multitask deep learning model was established to guide the balance between degradability and toughness.
• Multi-dimensional analysis combining TD-NMR characteristics and chemical descriptors.
• Focused on marine-relevant biodegradable polymers.

Abstract

As the world shifts toward a sustainable and circular economy, the demand is growing for materials that are both high-performing and environmentally responsible. Biodegradable polymers present a fundamental design challenge, as increased mechanical strength can impede breakdown, creating a trade-off that conventional strategies struggle to reconcile. Here we present a multimodal, multitask deep-learning framework that models mechanical performance and mass-loss behavior from molecular descriptors, thermal properties, and Nuclear Magnetic Resonance (NMR) signals. Within a panel of seven representative biodegradable polyesters relevant to marine/estuarine setting as 0.2-mm films and evaluated in vacuum-filtered estuary water, the framework jointly addresses multiple targets—strain at break, maximum stress, Young’s Modulus, and 30-day endpoint mass loss—and highlights hierarchical molecular and dynamic features underlying the toughness–degradability balance. SHAP analysis revealed that features governing segmental rigidity and thermal properties jointly drive the trade-off between mass loss and mechanical performance. In particular, descriptors from Time Domain Nuclear Magnetic Resonance (TD-NMR) relaxometry and ¹³C/¹H NMR principal components consistently emerged as top predictors linking degradability with strain at break, maximum stress, and Young’s Modulus. This framework establishes a generalizable route to integrate multimodal NMR-derived dynamics with machine learning for rational design of biodegradable polymers.