On the role of knowledge graphs in AI-based scientific discovery

Research and the scientific activity are widely seen as an area where the current trends in AI, namely the development of deep learning models (including large language models), are having an increasing impact. Indeed, the ability of such models to extrapolate from data, seemingly finding unknown patterns relating implicit features of the objects under study to their properties can, at the very least, help accelerate and scale up those studies as demonstrated in fields such as molecular biology and chemistry. Knowledge graphs, on the other hand, have more traditionally been used to organize information around the scientific activity, keeping track of existing knowledge, of conducted experiments, of interactions within the research community, etc. However, for machine learning models to be truly used as a tool for scientific advancement, we have to find ways for the knowledge implicitly gained by these models from their training to be integrated with the explicitly represented knowledge captured through knowledge graphs. Based on our experience in ongoing projects in the domain of material science, in this position paper, we discuss the role that knowledge graphs can play in new methodologies for scientific discovery. These methodologies are based on the creation of large and opaque neural models. We therefore focus on the research challenges we need to address to support aligning such neural models to knowledge graphs for them to become a knowledge-level interface to those neural models.