Mapping science and revealing disciplinary communication modalities via pre-trained graph neural networks

Current studies predominantly highlight the growing intersections among disciplines but lack insights into more nuanced aspects of science communication. This work investigates disciplinary communication through two metrics: interactivity, defined as the product of knowledge absorption and diffusion, capturing the overall breadth of knowledge interaction; and radiation, the ratio of outward diffusion to absorption, reflecting the relative tendency to export knowledge. To achieve this, we encode the disciplinary information of each paper as a continuous vector by pre-trained graph neural networks on extensive academic data. The metrics are derived from the distances computed using the paper vectors. We categorize the disciplines into four quadrants: “exposed,” “absorptive,” “service,” and “hermetic”, based on the two metrics. Our findings indicate that life-related sciences (medicine, neuroscience) are “exposed,” with open characteristics. Formal sciences (mathematics, physics and astronomy) are “hermetic,” with limited interaction breadth and radiation capacity. Chemistry, business and management are “absorptive,” focusing on knowledge absorption with limited dissemination. Engineering and Energy are “service-oriented,” centered on transformation and connecting. Our findings and computational methods could contribute to a better understanding of scientific communication systems.