The research explores the connection between surprising scientific breakthroughs and their subsequent impact. Charles Sanders Peirce's concept of abduction, where unexpected findings disrupt existing expectations, is used as a theoretical framework. The authors challenge the notion that surprise alone explains success, arguing that a balance between sufficient knowledge within a field to recognize anomalies and sufficient knowledge outside the field to offer alternative explanations is crucial. The study aims to empirically demonstrate the predictive power of surprising discoveries and inventions and to pinpoint the origins of these breakthroughs. Existing models of scientific novelty, which often focus on simple novelty scores or institutional structures, are inadequate for capturing the complexity of this phenomenon. Therefore, the authors propose a novel generative model to more accurately predict and assess the impact of surprising research outputs.

Prior literature successfully modeled discoveries and inventions using combinatorial processes, focusing on simple combinations of components. This research builds upon this foundation, but emphasizes the importance of higher-order structures in complex networks, drawing inspiration from work on transportation networks, neural networks, and food webs. The authors' approach distinguishes itself by separately considering research content (e.g., concepts and methods) and context (e.g., scientific disciplines), providing a more precise characterization of novelty and surprise. This distinction allows for a clearer understanding of how different types of novelty contribute to scientific and technological progress and also allows them to contrast scientific discovery with technological search.

To predict new innovations, a generative hypergraph model is developed. This model extends the mixed-membership stochastic block model into higher dimensions, characterizing complete combinations of research content and context. The model first constructs a continuous embedding for nodes from the hypergraph of contents or contexts, then allows the embedding to evolve stochastically, and finally draws a new hypergraph from the updated embedding, predicting next year's combinations. This approach allows the identification of surprising combinations—those least likely according to the model—and their subsequent impact. The model is applied to three datasets: 19,916,562 biomedical articles from MEDLINE, 541,448 physical science articles from the American Physical Society (APS), and 6,488,262 US patents. Research content is operationalized using keywords from curated ontologies (MeSH terms, PACS codes, USPC codes), while context is represented by cited journals or technology classes. Hypergraphs are constructed for both content and context, with nodes representing keywords/classes and hyperedges representing papers/patents combining those elements. The model's accuracy in predicting future combinations is assessed using AUC (Area Under the Curve) scores. The novelty of a combination is quantified as its improbability or surprisal. The relationship between surprise (content and context novelty) and impact (citations, awards) is analyzed by dividing papers into citation deciles and examining the distribution of novelty scores within each decile. Finally, the authors assess the sources of surprising advances by analyzing the novelty of researchers, research teams, and research expeditions (measured by distance between team members' publishing backgrounds and the publication venue).

The model accurately predicts future combinations of research content and context (AUC scores > 0.95 for most cases). Surprise, measured as the improbability of content and context combinations, is strongly associated with high citation impact. Papers in the top 10% of citations (hit papers) exhibit significantly higher content and context novelty than average. Nobel Prize-winning papers show lower context novelty but higher content novelty, likely due to awards being conferred by disciplinary communities. The probability of a paper being a hit increases monotonically with novelty percentile; papers with surprising context combinations are more likely to be hits than those with surprising content combinations alone. Content and context novelty provide nearly independent information, with low correlation between them. Expert classifications of biomedical papers from Faculty Opinions reinforce this distinction, with content novelty linked to “New findings” and “Drug targets”, and context novelty associated with “Controversial”, “Interesting hypothesis”, and “Technical advance”. Patents exhibit less of a predictive power of context novelty due to weaker disciplinary boundaries in technology. Scientists tend to cite contexts similar to their own publication venues, while inventors cite distant sources with similar likelihood, indicating differences in how science and technology communities operate. Analysis of career, team, and expedition novelty reveals that surprising research expeditions, where scientists from one field address a problem in a distant field, are particularly strongly associated with high-impact papers. These expeditions are more predictive of success than interdisciplinary careers or teams alone.

The findings strongly support Peirce's claim that abduction is a key driver of scientific progress. The high predictive power of the model shows that a substantial portion of scientific impact can be anticipated by assessing the improbability of research outputs. The emphasis on collective abduction highlights that breakthroughs frequently arise from collaborations across disciplines, demonstrating the significance of cross-disciplinary 'expeditions'. The distinction between content and context novelty clarifies two distinct dimensions of novelty, offering a more nuanced understanding of innovation. While existing biases might discourage novelty in funding or publication, the study demonstrates a clear impact bias favoring novelty among successfully published papers. The observed differences between science and technology highlight the unique characteristics of each domain and how those characteristics may influence innovation strategies.

This research provides a robust model for predicting scientific and technological breakthroughs based on the concept of surprise. The findings underscore the importance of interdisciplinary collaboration and 'knowledge expeditions' in driving impactful discoveries. The model's high predictive power and the detailed analysis of content and context novelty offer valuable insights for evaluating scientific institutions and fostering innovation. Future research should investigate the causal relationship between cross-disciplinary expeditions and scientific progress, and explore how to systematically incentivize such ventures to enhance the rate of breakthroughs.

The study evaluates the surprise measure only on papers that passed peer review; analysis of rejected papers would provide a more comprehensive understanding. The coarse-grained operationalization of content and context, using keywords and cited journals, may not fully capture the nuances of scientific contributions. The model distinguishes realized papers from random ones, not predicting all published papers. Finally, the representation of papers as unstructured bags of keywords ignores the inherent structure of scientific contributions. Despite these limitations, the findings offer a valuable model for predicting and understanding scientific progress.