Scientific discoveries often grow from links between ideas that have not been connected before. Researchers from the University of Chicago and the Santa Fe Institute have developed AI models that use the distribution of human expertise and collaboration to forecast where some of those links may emerge.
The approach is designed both to predict discoveries that follow from existing research and to suggest hypotheses beyond the areas scientists commonly explore. The team says its results point to a way for AI to complement research communities by accounting for their shared patterns of attention.
Using expertise as a signal
AI models trained on published findings are already being used to identify materials and targeted therapies. The researchers extend that approach by representing not only what scientific papers say, but also how researchers and topics are connected.
They built research hypergraphs from publication metadata. These structures captured relationships among materials, properties and the authors involved in the research. The model then generated random walk sequences across the hypergraphs to identify inferences that could be cognitively accessible to human experts, given their knowledge and collaborative networks.
In this way, the model treats the placement of expertise around a topic as useful information. The team found that those patterns offered a strong signal about the likelihood of future discoveries. Their explanation is that scientific advances often come from new connections between concepts that had seemed unrelated.
Forecasting familiar paths and less crowded ones
One role for the models is to anticipate discoveries that build directly on current knowledge. In the researchers’ account, this captures the routes that experts could plausibly follow through their existing expertise and connections.
A second role is to look beyond those accessible paths. The researchers tuned the models to avoid crowded areas of research, allowing them to propose plausible hypotheses that bridge more distant fields. Such suggestions could introduce concepts that might otherwise take years to be imagined, pursued or published, the team said.
The distinction matters because scientific attention is not distributed evenly. The researchers argue that field boundaries, established education and social connections shape which questions receive attention. A model that recognizes those patterns can also be adjusted to search outside them.
Reported gains across research problems
The team compared its method with approaches that analyze scientific content alone. In extreme cases, the human-aware models outperformed those methods by up to 400 percent. The reported results varied by application: prediction accuracy doubled for materials science problems, while accuracy in finding new drug applications improved by more than 40 percent.
The researchers attribute the gains to the models’ use of collective attention patterns among scientists. The results suggest that information about who works on which topics, and how those topics connect, can add predictive value beyond the scientific content itself.
These figures describe the team’s findings for the problems it studied. The article’s account does not establish that the same improvements will apply to every scientific field or task.
AI as a complement to scientific communities
The researchers describe scientific progress as leaning toward local exploration of familiar ideas rather than novel exploration of the unknown. Their system offers one way to surface opportunities that existing communities may overlook because of how expertise and attention are organized.
That does not remove the role of scientists: the model proposes and predicts connections, while its hypotheses still need to be considered and pursued by people. Its potential contribution is to widen the set of possibilities researchers can see, including ones that sit outside the busiest areas of work.
The paper frames human-aware AI as a way to move toward and beyond the contemporary scientific frontier. By learning from the structure of research and deliberately searching beyond common routes, these models could help researchers identify both likely next steps and less expected directions.