Mapping the Science Race: Comparison of the US and Chinese Technology Portfolios using Artificial Intuition

• In Event: Infrastructure, Industrial Policy, and the Political Economy of AI

Abstract

The United States and China compete for global scientific and technological leadership. Federal research investment is a primary lever in this competition, yet policymakers lack systematic tools to compare portfolios across the two nations in real time. Existing approaches rely on self-reported classifications that are often inconsistent or overly broad, limiting their utility for rigorous comparative analysis. As the CHIPS and Science Act reorients federal efforts toward tracking outcomes of strategic technology investment, there is a timely need for scalable, automated, and reproducible methods for portfolio comparison.

Identifying the relevant topics of deep technology research are more challenging because many transformative technologies sit at the interfaces of multiple fields. Ordinary industrial classification codes are inaccurate as they don’t keep up with emergent technologies, and cannot handle multiple classifications. Funding data is a rich source of information but the technical content is typically dense and brief. Such data convey extensive information to experts because of their tremendous background, and thus they conduct this process accurately. But this approach is costly due to the human time required, motivating the use of Artificial Intelligence (AI) tools to replicate the expert’s thought process.

This paper uses artificial intuition, a computational method that replicates the expert classification process at scale, to systematically compare the research portfolios of the US National Science Foundation (NSF) and the National Natural Science Foundation of China (NSFC). Artificial intuition works by first extracting keywords from short scientific abstracts using unsupervised methods, then prompting a Large Language Model (LLM) to enrich those keywords with contextual metadata, mimicking the background knowledge an expert reader brings to the task. The enriched representations are then embedded and clustered to produce a non-overlapping set of topic labels, enabling clean portfolio-level accounting without the double-counting endemic to keyword-based schemes.

We apply this method to a corpus of publicly available award abstracts of NSF and NSFC awards. Our automatically generated topic labels are validated against manual classifications. We then use these labels to characterize and compare the two nations' investment distributions. Preliminary findings reveal meaningful divergences in portfolio emphasis. Chinese NSFC funding shows concentrated activity in materials and nanostructures, while US NSF investment is more evenly distributed across topic clusters. Thus while concentrated investment in strategic materials aligns with China's stated industrial policy goals, the US portfolio reflects a broader basic research mandate that may underweight applied technology areas targeted by current legislation.

This work contributes to the science policy literature by providing a reproducible, scalable infrastructure for comparative portfolio analysis that does not depend on self-reported taxonomies or labor-intensive manual review. It offers a new lens for monitoring international science competition, informing funding strategy, and evaluating whether investment aligns with national priorities. We demonstrate the efficacy and opportunities of an AI-driven system for industrial classification of summaries of federally funded projects that can be extended to research portfolio management, technology scouting, and other strategic activities.

Author

• Alexandra Graddy-Reed, University of Southern California
  Presenting Author