Breakthrough innovation
Breakthrough patents by European country
Trend over time
Novelty by technology field
Comparison by country
Top 10 European regions
Map of Europe
2024 incomplete: patent applications are published about 18 months after filing, and the year still contains only about 83% of the patents of a complete year. Shares are little affected; the 2024 figures per million inhabitants will be revised upward.
About this tracker
This dashboard analyzes 4.6 million patents filed with the European Patent Office (EPO) between 1983 and 2024, covering 24 European countries. It identifies breakthrough patents and measures their output by country, by region (NUTS2) and by technology field.
A patent is considered a breakthrough patent if it is in the top 1% of the annual distribution according to one of the following two measures:
Novelty (backward-looking): measures how different a patent is from prior art at the time of filing. Formally, it is the sum of the cosine similarities between the embedding vector of the abstract and those of all patents filed in the previous 5 years. A patent is classified as highly novel if it has a low score (bottom 1%), indicating that it is distant from existing technologies. This measure is available for all years, including the most recent.
Breakthrough : ratio between future influence and proximity to prior art (forward_sum / backward_sum). A patent is classified as a breakthrough patent if it has a high ratio (top 1%), which identifies patents that are both very different from existing technologies and very influential on future innovations. Note: this measure is not available after 2019 because it requires observing the patents filed in the following 5 years.
The country charts show the number of breakthrough patents per million inhabitants.
Authors
This tracker is developed by:
- Antonin Bergeaud
- Dario Marino
Methodology
Computing the scores (KPST)
The methodology is inspired by Kelly, Papanikolaou, Seru and Taddy (2021, AER Insights) :
- Text embeddings : the Englishabstract of each patent is turned into a high-dimensional numerical vector with the PatentSBERTa_V2 model, a Sentence-BERT model pre-trained on patent texts. When a patent has an abstract only in German or French (about 16% of patents), the machine English translation from Google Patents is used: all patents are thus compared in the same language.
- Novelty score (
backward_sum): sum of the cosine similarities between the vector of the patent’s abstract and those of all patents filed in the previous 5 years. A low score means that the patent is far from the state of the art. - Future proximity (
forward_sum): sum of the cosine similarities between the vector of the abstract and those of all patents filed in the following 5 years. Used as the numerator of the breakthrough score. - Breakthrough score (
importance): ratioforward_sum / backward_sum. A high ratio identifies patents that are both far from existing technologies and close to future innovations.
Definition of breakthrough patents
A patent is classified in the top 1% within each filing year:
- Highly novel : bottom 1% of the distribution of
backward_sum(the most distant from prior art) - Breakthrough : top 1% of the distribution of
importance(the highest forward/backward ratio)
Thresholds are computed separately for each year to control for changes in the distributions over time.
Novelty index by technology field
The index shown in the chart by technology field measures the relative gap between the average novelty score of a technology field and the overall average:
\[\text{Index} = \frac{\overline{backward\_sum}_{\text{global}} - \overline{backward\_sum}_{\text{field}}}{\overline{backward\_sum}_{\text{global}}}\]
A positive index means that the field’s patents are on average more novel than the average. The index is computed each year.
Technology classification (CPC)
The chart of novelty by technology field uses 4 fields of the Cooperative Patent Classification (CPC, the harmonized EPO/USPTO system):
| CPC code | Field |
|---|---|
| G06N | AI / Machine learning |
| C12N | Biotechnology |
| A61K | Pharmaceuticals |
| H01L and H10 | Semiconductors (H10 takes over H01L since the CPC reforms; current codes apply to all years) |
Geographic attribution
By country: patents are attributed based on the location of the inventors (source: REGPAT, OECD). When a patent has several inventors from different countries, it is counted fractionally (1/N per inventor).
By region (NUTS2): patents are attributed to NUTS2 regions according to the regional code of the inventor’s residence. The regional data cover about 380 European NUTS2 regions.
Data sources
- Patents : abstracts and CPC codes of EPO publications, Google Patents Public Data (BigQuery) — 4,649,815 EPO patents filed from 1983 to 2024
- Location : REGPAT (OECD, August 2026 edition) — country and region (NUTS2) of origin of the inventors
- Classification : CPC codes (Cooperative Patent Classification) — 4 targeted fields
- Population : World Bank (indicator SP.POP.TOTL) — for normalization per million inhabitants
Update
Data as of October 6, 2026: EPO publications up to September 16, 2026 (Google Patents) and REGPAT (OECD) of August 2026. The tracker is updated after each new REGPAT edition. 2024 is incomplete: patent applications are published about 18 months after filing, and the year still contains only about 83% of the patents of a complete year. Shares are little affected; the 2024 figures per million inhabitants will be revised upward.