AI and Supercomputing are Powering the Next Wave of Breakthrough Science - But at What Cost?

TL;DR

This study quantifies how the convergence of AI and high-performance computing (HPC) shapes scientific discovery at scale, using metadata from over 5 million publications across 27 fields (2000–2024). It combines AI keyword searches, HPC-acknowledgment mining, and large-language model (LLM) classifications to identify AI, HPC, and AI+HPC papers, then assesses novelty and top-cicatation impact within a rigorous regression framework that reveals a synergistic boost when both technologies are used together. The findings show that AI+HPC papers are disproportionately represented among top-cited and novel works, indicating a substantial, yet uneven, acceleration of breakthrough science. The study also documents growing global inequality in access to compute resources and AI expertise, calling for policy interventions to ensure broader, equitable diffusion of these transformative tools. $AI \times HPC$ synergy emerges as a key driver of frontier science, but only if structural barriers to access are addressed to prevent a two-tier scientific ecosystem.

Abstract

Artificial intelligence (AI) and high-performance computing (HPC) are rapidly becoming the engines of modern science. However, their joint effect on discovery has yet to be quantified at scale. Drawing on metadata from over five million scientific publications (2000-2024), we identify how AI and HPC interact to shape research outcomes across 27 fields. Papers combining the two technologies are up to three times more likely to introduce novel concepts and five times more likely to reach top-cited status than conventional work. This convergence of AI and HPC is redefining the frontier of scientific creativity but also deepening global inequalities in access to computational power and expertise. Our findings suggest that the future of discovery will depend not only on algorithms and compute, but also on how equitably the world shares these transformative tools.

Figures (6)

• Figure 1: Growth and convergence of AI and HPC in scientific research. (A) Number of AI and HPC publications over time (log scale). (B) Share of all scientific papers involving AI (orange, left axis) and HPC (blue, right axis), based on Scopus data. (C) Co-occurrence of AI and HPC publications: share of HPC papers that use AI (orange, left axis) and share of AI papers that use HPC (blue, right axis). (D) Among AI+HPC papers, classification by contribution type: “AI use” (green) vs. “AI development” (red). (E) Disciplinary distribution of AI, HPC, and combined AI+HPC work. For each field, bars show the share of total output involving AI (top, orange), HPC (middle, blue), or both (bottom, purple).
• Figure 2: Breakthrough potential in AI and HPC-powered science. (A) Share of papers in the top 1% of citations by field and computational approach: AI only (orange), HPC only (blue), and both AI and HPC (purple). The dashed line marks the 1% baseline expected by random chance. (B) Share of novel papers (i.e., those introducing a new term subsequently reused) among the top 1% most cited, across fields and computational approaches. Panels A and B display the ten scientific fields with the largest volume of publications combining AI and HPC. (C) Predicted probability of novelty (from regression models controlling for confounders such as affiliation, prior citations, and year) for four groups: papers using neither AI nor HPC (gray), AI only (orange), HPC only (blue), and both (purple).
• Figure 3: Global inequality in computational resources and AI/HPC-powered science. (A) Inequality in the global distribution of high-performance computing resources, measured using the Mean Log Deviation (MLD) for HPC system counts and FLOPs (1995–2024), based on the Top500 dataset (own elaboration, see \ref{['sm:computing_capacity']}]). (B) Top 10 countries by cumulative supercomputing capacity (FLOPs), 1995–2024. (C) Inequality in scientific publications for AI, HPC, and all research fields, measured using the Mean Log Deviation (MLD), 2000–2024. (D) Top 10 countries by cumulative AI and HPC publications, 2000–2024. The four panels show that inequality is rising sharply in access to computational power and in AI- and HPC-intensive research output, while science as a whole shows stable -- or even slightly decreasing -- levels of concentration.
• Figure S1: Convergence of AI and HPC in scientific research (all fields). Disciplinary distribution of AI, HPC, and combined AI+HPC work. For each field, bars show the share of total output involving AI.
• Figure S2: Breakthrough potential in AI and HPC-powered science. (A) Share of papers in the top 1% of citations by field and computational approach. The dashed line marks the 1% baseline expected by random chance. (B) Share of novel papers (i.e., those introducing a new term subsequently reused) among the top 1% most cited, across fields and computational approaches. In fields where AI+HPC papers remain relatively few, estimates should be interpreted with caution. Results for fields with larger sample sizes are shown in the main text.