Since 2016, the Global Innovation Index (GII) has sought to identify innovation clusters using a bottom-up approach (
Three innovation metrics are employed in the compilation of the top 100 GII innovation clusters worldwide: the location of inventors listed in published patent applications; the location of authors listed in published scientific articles; and the location of firms in receipt of venture capital (VC) investment.
For patents, this method relies on applications under WIPO’s Patent Cooperation Treaty (PCT). PCT patents offer a useful basis for analyzing patents globally. The PCT system applies a single set of procedural rules and collects information based on uniform filing standards. This reduces any potential biases that could arise from using data collected from multiple national sources. The patents selected were published between 2021 and 2025, the most recent five-year period available. This is to minimize the volatility that can occur between given years.
As a second step, scientific publications from the Web of Science’s Science Citation Index Expanded (SCIE) were incorporated. The SCIE provides detailed coverage of the world’s most impactful academic journals. For the purposes of the analysis, science and technology fields were the focus, whereas articles from the fields of social sciences and humanities were disregarded. In addition, scientific publications were limited solely to articles of original research. This is to exclude other published items, such as meeting abstracts, conference summaries or paper briefs. As with PCT filings, the most recent five-year period according to data availability was used for the SCIE – publication years 2020 to 2024.
For a second consecutive year, VC deal data were incorporated in order to further enrich the understanding of innovation activity at the cluster level. By integrating information on firms that have been in receipt of VC funding, the analysis is able to capture entrepreneurial and early-stage innovation activity, reflecting not only scientific and inventive outputs, but also innovation finance and startup dynamics.
For VC data, PitchBook’s Venture Capital Database was utilized. The database offers detailed deal-level VC count and value information, and is used to identify the precise geographic location of firms receiving VC investments. As we wanted to capture VC activity, the total number of deals was counted, regardless of the total value. A VC deal is defined as any firm-funding event. Thus a firm that had multiple funding events may appear multiple times in the deal counts. VC data were sourced for the same five-year period as were scientific articles, that is 2020–2024, based on the year the funding event took place.
The WIPO PCT patent data set consists of approximately 1.3 million patent applications published between 2021 and 2025, containing 4.4 million inventor addresses. For the SCIE, the data set comprises 8.6 million articles published between 2020 and 2024, containing 30.4 million listed author addresses. For PitchBook’s Venture Capital Database, the data set consists of 51,576 locations between 2020 and 2024, containing 237,796 deals.
The geocoding process for addresses used in this report is as follows. PCT inventor addresses were geocoded using the Environmental Systems Research Institute (ESRI) ArcGIS World Geocoder service. In cases where the ESRI results were ambiguous or insufficiently accurate, the city name was extracted from the address string and matched against entries in a dataset derived from the GeoNames Gazetteer database[i]—a global dataset of approximately 766,000 geocoded city names. If the extracted city did not match any record in GeoNames, an attempt was made to geocode the city name directly using the World Geocoder service.
This same city-matching approach was applied to SCIE author addresses and VC deal locations. In both datasets, the addresses were provided in a pre-parsed format, which significantly improved the ability to match them using the GeoNames database. For SCIE and VC city names that could not be matched using GeoNames, again an attempt was made to geocode the city name using ESRI’s World Geocoder.
Overall, 99.9 percent of inventor addresses were geocoded either at the city level or a more accurate level, while 99.8 percent of scientific author addresses were geocoded at the city level. For VC data, 99.5 percent of VC deals were geocoded at the city level or better. Appendix IV Table 1 provides a summary of the geocoding results for the top 20 countries, which together account for the majority of inventor, scientific author and VC deals addresses. As shown in the table, the coverage of geocoded PCT inventor addresses across all 20 countries was above 97 percent. Similarly, coverage of scientific author addresses and VC deal addresses was high at above 96 percent.
Addresses were clustered by applying the density-based spatial clustering of applications with the noise (DBSCAN) algorithm. This algorithm requires predefined radius and density parameters. As in previous years, a radius of 15 km and a density of 4,500 listed inventors/authors were applied. Equal weight was given to inventors and authors by expressing data points as a share of total inventor and author addresses, respectively. Given that the number of scientific articles far exceeds that of patents, cluster identification based on the raw data points would have resulted in clusters shaped predominantly by the scientific author landscape.
The location of VC-backed firms was excluded from the initial cluster formation process, because of the relatively high degree of geographic dispersion compared to PCT inventors and scientific article authors. The sparser distribution of VC data points risked introducing noise and distorting cluster boundaries if it had been included during the DBSCAN algorithm execution. To address this limitation, VC locations were assigned to clusters post hoc using a modified nearest-neighbor approach. Specifically, the DBSCAN prediction function was adapted so as to assign each VC point to a cluster if it fell within 15 km of any core point within that cluster. Critically, we restricted distance calculations to core points only, excluding boundary points from consideration. This approach ensures that VC firms are only assigned to clusters having a sufficiently dense level of entrepreneurial activity within their immediate vicinity, thereby remaining consistent with the density-based clustering logic, while at the same time accommodating the distinctive spatial distribution of VC investments.
The clustering step resulted in an initial list of 244 clusters. After review, neighboring clusters were merged if the edge of one cluster was within 3–5 km of another and where the co-author/co-inventor relationships were higher than for any other relationship with any other cluster or non-cluster points. A total of 18 clusters met these criteria, with mergers reducing the overall number of clusters identified to 235.
The remaining 235 clusters were then ranked by counting the number of patents, scientific articles and VC deals within a given cluster. Numbers were aggregated using fractional counting, in which counts reflect the share of a patent’s inventors and an article’s authors present within a given cluster. In addition, mirroring the equal weighting approach described above, fractional counts are relative to the total numbers of patents, scientific articles and VC deals (Appendix IV Table 2).
To produce an intensity ranking, the European Commission’s Global Human Settlement Layer (GHSL) population distribution data (
References
Bergquist, K. and C. Fink (2020). The top 100 science and technology clusters. In Dutta, S., B. Lanvin and S. Wunsch-Vincent (eds), The Global Innovation Index 2020: Who Will Finance Innovation? Ithaca, NY, Fontainebleau and Geneva: Cornell University, INSEAD and World Intellectual Property Organization. Available at: www.wipo.int/edocs/pubdocs/en/wipo_pub_gii_2020.pdf.
PitchBook (2024). Global VC Ecosystem Rankings: An update on our location-based VC Ecosystem Rankings. September 23, 2024. Available at: https://pitchbook.com/news/reports/q3-2024-pitchbook-analyst-note-global-vc-ecosystem-rankings.
Schiavina, M., S. Freire, A. Carioli and K. MacManus (2023). GHS-POP R2023A – GHS population grid multitemporal (1975–2030). Brussels: European Commission, Joint Research Centre (JRC). Available at: http://data.europa.eu/89h/2ff68a52-5b5b-4a22-8f40-c41da8332cfe.