Framework of Copolymer gel Dynamics and Ligand Interaction Factors, Objective-free exploration, and CVT-MAPElites

Discovery-oriented research is a fundamental pursuit in chemical and materials science, especially when objective-free or purpose-ambiguous exploration can yield unexpected novel compounds or materials. Recently, data-driven objective-free exploration methods have emerged to support such discovery in materials science. However, these methods have mostly been limited to relatively simple systems (e.g., chemical composition), and extending them to real-world multiscale, multi-component materials like soft matter remains a critical challenge. In this study, we present a novel framework for objective-free exploration of a complex gel–ligand interaction system that integrates multiscale data. We embed monomer physicochemical information using four different large-language models (LLMs), and combine these embeddings with polymer dynamics descriptors obtained via time-domain NMR measurements. Using this combined representation, we perform novelty-driven search with BLOX (BoundLess Objective-free eXploration) and utilize CVT-MAP-Elites to map candidates in an interpretable behavioral characteristic space. Unlike conventional molecular descriptors, the LLM-based embedding supports a much broader and richer exploration of gel-ligand interaction space evaluated by 13C NMR signal intensities. The results demonstrate that our framework uncovers a wider diversity of candidate materials than descriptor-based approaches, thereby establishing a new pathway for discovery-oriented design of functional materials that integrates multiscale and multi-feature information.