Mapping the Molecular Universe

Abstract

Abstract The success of large-language-models highlighted the importance of datasets, whose quantity, quality, and diversity dictate the accuracy and creativity of artificial intelligence (AI) systems in learning and inferring. However, in scientific domains, such datasets required for domain-specific AI/machine-learning (ML) models do not exist most of the time. In this perspective, by examining the logics and rationales behind the construction and mapping of a universe for small organic molecules (108–1012), we explore what new capabilities AI can bring to the materials discoveries. We identify the elusive structure–property relationship as a core challenge for new materials design and explore the possibility AI tools could enable us to design based on targeted properties instead of the structural derivation. In a broader context, we further discuss what challenges and limitations AI-driven material sciences still face and envision possible avenues into the future of AI4Materials.