Technology landscape · Machine learning, German applicants · 2014–2024
German machine-learning patent families grew 6.7-fold from 2014 to 2021, and 6,361 more sit outside the core codes.
9,479 families in the core IPC groups for machine learning and pattern recognition. Half carry a neural-network code. Robert Bosch holds 22.9% of all families.
- Data
- EPO PATSTAT Global, Spring 2026 (patstat_2026a)
- Unit
- DOCDB family
- Scope
- German applicants · IPC G06N and G06F18/00 with predecessors, 115 symbols
- Queries
- Query 0a to 10, all published
G06N3: 19 → 747 · G06N20: 16 → 509 · G06N10: 1 → 152 families
- 6.7×
Families per year grew from 211 (2014) to 1,417 (2021).
Growth - 49.6%
of the 9,479 families carry a neural-network code (G06N3).
Sub-areas - 22.9%
of the families belong to Robert Bosch (2,166), more than the next two applicants together.
Leaders - 50.0%
of the 2021–2024 families have a DPMA member, level with the US at 50.2%.
Offices - 6,361
further families use machine learning inside another field and carry no core code.
Hidden
01 · Growth
Filings grew for seven years in a row, then levelled off after 2021
2014–2016
211, 221 and 287 families. Counting only the codes of today's IPC edition gives 78 for 2014; the gap is the predecessor symbols, so it reflects reclassification history.
2017–2019
The steepest single-year increases: +68.6% in 2017 and +66.5% in 2018.
2021
The peak at 1,417 families, 6.7 times the 2014 level. Measured 2014 to 2023, the compound annual growth rate is 22.9%.
2022–2024
1,155 families in 2022 and 1,347 in 2023; the two code-set readings converge. 2024 (979) is still being published, and no growth claim rests on it.
Patent families by earliest filing year, two code-set readings
Data table: filing activity 2014–2024
| Year | Families (completed code set) | Year on year | Families (current-edition codes only) |
|---|---|---|---|
| 2014 | 211 | — | 78 |
| 2015 | 221 | +4.7% | 97 |
| 2016 | 287 | +29.9% | 142 |
| 2017 | 484 | +68.6% | 321 |
| 2018 | 806 | +66.5% | 599 |
| 2019 | 1,266 | +57.1% | 1,059 |
| 2020 | 1,306 | +3.2% | 1,129 |
| 2021 | 1,417 | +8.5% | 1,333 |
| 2022 | 1,155 | −18.5% | 1,146 |
| 2023 | 1,347 | +16.6% | 1,348 |
| 2024* | 979 | −27.3%* | 979 |
| Total | 9,479 | — | 8,231 |
Query 2 · 29.9 GB passEach family counted once, in its earliest in-scope filing year; the years sum to 9,479
02 · Sub-areas
Half the families carry a neural-network code
4,699 of 9,479 families (49.6%) carry G06N3, the group for computing arrangements based on biological models that holds neural networks and deep learning. General machine learning (G06N20) reaches 3,127 families (33.0%) and pattern recognition (G06F18) 2,466 (26.0%). Quantum computing (G06N10) and probabilistic or fuzzy models (G06N7) are level at 493 each, on opposite trajectories.
Patent families per machine-learning sub-area
Data table: machine-learning sub-areas
| Sub-area | Families | Share of 9,479 |
|---|---|---|
| Neural networks & deep learning (G06N3) | 4,699 | 49.6% |
| Machine learning, general (G06N20) | 3,127 | 33.0% |
| Pattern recognition (G06F18) | 2,466 | 26.0% |
| Knowledge-based models (G06N5) | 1,096 | 11.6% |
| Quantum computing (G06N10) | 493 | 5.2% |
| Probabilistic & fuzzy models (G06N7) | 493 | 5.2% |
| Other computational models (G06N1, G06N99) | 146 | 1.5% |
| Sum of sub-areas | 12,520 | 132.1% |
Query 3 · 21.7 GB passNon-exclusive: the seven values sum to 12,520
A family can carry codes from several sub-areas, so shares are read against the family total and not against each other; no pie chart is used. Two predecessor symbols are folded into their successors: G06F15/18 counts towards G06N20 and G06K9/62 towards G06F18.
03 · Evolution
Quantum computing peaked last, with 152 families in 2023
Neural networks
G06N3 goes from 19 families in 2014 to 747 in 2021 and then holds a plateau near 700. General machine learning follows one step lower, 16 to 509, then near 450.
Quantum computing
1 family in 2014, 44 in 2020, 152 in 2023. Still small at 493 in total.
Probabilistic and fuzzy
G06N7 peaks at 85 families in 2019 and falls to 30 in 2023, consistent with a classical modelling branch being crowded out by learned models.
Read with care
Pattern recognition drops from 302 (2021) to 94 (2022). G06F18 came with the 2023 IPC edition and took over from G06K9/62; the 2022 cohort sits on that boundary. The dip is a classification transition.
Patent families per sub-area and earliest filing year
Data table: sub-areas by year
| Sub-area | 2014 | 2015 | 2016 | 2017 | 2018 | 2019 | 2020 | 2021 | 2022 | 2023 | 2024* |
|---|---|---|---|---|---|---|---|---|---|---|---|
| Neural networks (G06N3) | 19 | 21 | 50 | 171 | 364 | 686 | 723 | 747 | 679 | 723 | 516 |
| Machine learning, general (G06N20) | 16 | 36 | 57 | 135 | 226 | 405 | 460 | 509 | 435 | 479 | 369 |
| Pattern recognition (G06F18) | 138 | 130 | 160 | 216 | 299 | 453 | 400 | 302 | 94 | 155 | 119 |
| Knowledge-based models (G06N5) | 34 | 40 | 37 | 57 | 111 | 160 | 163 | 159 | 118 | 130 | 87 |
| Quantum computing (G06N10) | 1 | 0 | 2 | 1 | 7 | 21 | 44 | 90 | 78 | 152 | 97 |
| Probabilistic & fuzzy (G06N7) | 21 | 21 | 41 | 38 | 65 | 85 | 59 | 72 | 40 | 30 | 21 |
| Other (G06N1, G06N99) | 12 | 10 | 17 | 47 | 50 | 3 | 1 | 2 | 1 | 3 | 0 |
Query 3 · 21.7 GB passNon-exclusive; 2024 provisional
04 · Neural nets
Learning methods outweigh architecture, 2,443 families to 1,401
Inside G06N3, German filings are more often about how a model is trained, validated or adapted than about the structure of the network. Among named architectures the order is combinations of networks (477), auto-encoder and encoder-decoder networks (168), convolutional networks (167) and recurrent networks (86). Among learning regimes: backpropagation (208), supervised (187), non-supervised (119), reinforcement (103), transfer (80), adversarial (69) and distributed or federated learning (67). 186 families sit in G06N3/063, the hardware realisation of neural networks.
Patent families per G06N3 subgroup, top 15
Data table: G06N3 subgroups
| IPC symbol | Title | Introduced | Families |
|---|---|---|---|
| G06N3/08 | Learning methods | 2000 | 2,443 |
| G06N3/04 | Architecture, e.g. interconnection topology | 2000 | 1,401 |
| G06N3/02 | Neural networks | 2000 | 1,177 |
| G06N3/045 | Combinations of networks | 2023 | 477 |
| G06N3/00 | Computing arrangements based on biological models | 2000 | 253 |
| G06N3/084 | Backpropagation, e.g. using gradient descent | 2023 | 208 |
| G06N3/09 | Supervised learning | 2023 | 187 |
| G06N3/063 | using electronic means (hardware realisation) | 2000 | 186 |
| G06N3/0455 | Auto-encoder networks; Encoder-decoder networks | 2023 | 168 |
| G06N3/0464 | Convolutional networks [CNN, ConvNet] | 2023 | 167 |
| G06N3/0475 | Generative networks | 2023 | 127 |
| G06N3/088 | Non-supervised learning, e.g. competitive learning | 2023 | 119 |
| G06N3/092 | Reinforcement learning | 2023 | 103 |
| G06N3/044 | Recurrent networks, e.g. Hopfield networks | 2023 | 86 |
| G06N3/047 | Probabilistic or stochastic networks | 2023 | 86 |
| G06N3/082 | Learning methods modifying the architecture | 2023 | 80 |
| G06N3/096 | Transfer learning | 2023 | 80 |
| G06N3/042 | Knowledge-based neural networks | 2023 | 75 |
| G06N3/094 | Adversarial learning | 2023 | 69 |
| G06N3/098 | Distributed learning, e.g. federated learning | 2023 | 67 |
Query 4 · 21.7 GB passNon-exclusive. Titles and introduction dates from IPC edition 20260101
These subgroups are young. Every symbol below G06N3/045 in the table, the named architectures and learning regimes, came with the 2023 IPC edition. Their counts describe how far offices have re-tagged the branch, not when the techniques were invented, which is also why the parent groups still hold most of the volume. Read this chapter as a relative profile, not as a time series.
05 · Leaders
Robert Bosch holds 22.9% of the landscape
Bosch's 2,166 families are more than the next two applicants together, Siemens AG (900) and SAP (746). Twelve of the twenty-five largest applicants are vehicle manufacturers or automotive suppliers: BMW, Volkswagen, ZF, Daimler, Audi, Porsche, Mercedes-Benz Group, Valeo, Conti Temic and three Continental entities. Outside that block the list holds one research organisation (Fraunhofer, 161), one chemicals company (BASF, 69), one semiconductor company (Infineon, 62) and one telecommunications operator (Deutsche Telekom, 51).
Patent families of the fifteen largest German applicants
Data table: twenty-five largest applicants
| # | Applicant | Type | Families | Share |
|---|---|---|---|---|
| 1 | Robert Bosch GmbH | Company | 2,166 | 22.9% |
| 2 | Siemens AG | Company | 900 | 9.5% |
| 3 | SAP SE | Company | 746 | 7.9% |
| 4 | Siemens Healthcare GmbH | Company | 410 | 4.3% |
| 5 | Bayerische Motoren Werke AG | Company | 358 | 3.8% |
| 6 | Volkswagen AG | Company | 333 | 3.5% |
| 7 | IBM Deutschland GmbH | Company | 217 | 2.3% |
| 8 | ZF Friedrichshafen AG | Company | 204 | 2.2% |
| 9 | NEC Europe Ltd | Company | 179 | 1.9% |
| 10 | Daimler AG | Company | 168 | 1.8% |
| 11 | Siemens Healthineers AG | Company | 166 | 1.8% |
| 12 | Fraunhofer-Gesellschaft | Research | 161 | 1.7% |
| 13 | Audi AG | Company | 146 | 1.5% |
| 14 | Dr. Ing. h.c. F. Porsche AG | Company | 129 | 1.4% |
| 15 | Mercedes-Benz Group AG | Company | 118 | 1.2% |
| 16 | Continental Automotive Technologies GmbH | Company | 102 | 1.1% |
| 17 | Valeo Schalter und Sensoren GmbH | Company | 95 | 1.0% |
| 18 | Continental Automotive GmbH | Company | 75 | 0.8% |
| 19 | Siemens Mobility GmbH | Company | 71 | 0.7% |
| 20 | BASF SE | Company | 69 | 0.7% |
| 21 | Continental Autonomous Mobility Germany GmbH | Company | 65 | 0.7% |
| 22 | Rohde & Schwarz GmbH & Co. KG | Company | 64 | 0.7% |
| 23 | Infineon Technologies AG | Company | 62 | 0.7% |
| 24 | Conti Temic microelectronic GmbH | Company | 60 | 0.6% |
| 25 | Deutsche Telekom AG | Company | 51 | 0.5% |
Query 5 · 24.3 GB passGrouped by harmonised applicant name; share of the 9,479 families
Nine of the twelve largest portfolios are led by neural networks. SAP files more in general machine learning (413) and knowledge-based models (232) than in neural networks (230). IBM Deutschland puts 149 of its 217 families into quantum computing, 30.2% of the entire German quantum sub-area. Daimler concentrates on pattern recognition (134 of 168 families).
Where each of the twelve largest portfolios sits, families per sub-area
| Applicant | G06N3 Neural nets | G06N20 ML general | G06F18 Pattern rec. | G06N5 Knowledge | G06N7 Probabilistic | G06N10 Quantum | Other |
|---|---|---|---|---|---|---|---|
| Robert Bosch | 1,367 | 772 | 528 | 217 | 140 | 20 | 8 |
| Siemens AG | 504 | 302 | 136 | 227 | 80 | 15 | 15 |
| SAP | 230 | 413 | 127 | 232 | 57 | 3 | 47 |
| Siemens Healthcare | 238 | 162 | 118 | 23 | 26 | 4 | 22 |
| BMW | 143 | 106 | 129 | 24 | 24 | 0 | 0 |
| Volkswagen | 163 | 99 | 127 | 20 | 10 | 12 | 0 |
| IBM Deutschland | 51 | 24 | 0 | 17 | 6 | 149 | 3 |
| ZF Friedrichshafen | 134 | 45 | 58 | 3 | 3 | 5 | 1 |
| NEC Europe | 104 | 74 | 19 | 58 | 13 | 0 | 8 |
| Daimler | 31 | 12 | 134 | 1 | 2 | 0 | 0 |
| Siemens Healthineers | 103 | 61 | 39 | 11 | 5 | 2 | 1 |
| Fraunhofer-Gesellschaft | 90 | 30 | 38 | 9 | 2 | 20 | 0 |
Query 5 · 24.3 GB passRows do not sum to the portfolio: a family can sit in several sub-areas
Corporate groups are not consolidated. The harmonised name merges spelling variants but not corporate structures: Siemens appears as four entries and Continental as three, and Daimler AG and Mercedes-Benz Group AG are the same group under its former and current name. Entries overlap where two of them co-file (52 families carry both Siemens Healthcare and Siemens Healthineers), so rows cannot be added. Sector labels come from psn_sector; across the landscape 7,907 families have an applicant recorded as a company, 283 a public research organisation, 251 a university, 261 an individual, and 1,258 an applicant with no sector recorded. These counts are non-exclusive.
06 · Domains
Road vehicles are the largest field of application
900 families carry B60W (conjoint control of vehicle sub-units), and five further vehicle-related subclasses follow in the top sixteen: traffic control (G08G, 468), vehicle fittings (B60R, 379), radar and radio navigation (G01S, 362), navigation (G01C, 290) and vehicle registering (G07C, 205). Healthcare is the second cluster, split between diagnosis and surgery (A61B, 427) and healthcare informatics (G16H, 423). Industrial automation (G05B 661, G05D 228), enterprise software (G06Q, 569), communications and measurement follow.
Applied IPC subclasses co-occurring with the core machine-learning codes
Data table: co-occurring IPC subclasses
| Subclass | Title | Families |
|---|---|---|
| G06F | Electric digital data processing | 2,287 |
| G06K | Graphical data reading; presentation of data; record carriers | 2,117 |
| G06V | Image or video recognition or understanding | 1,272 |
| G06T | Image data processing or generation, in general | 1,141 |
| B60W | Conjoint control of vehicle sub-units of different type | 900 |
| G05B | Control or regulating systems in general | 661 |
| G06Q | ICT for administrative, commercial, financial purposes | 569 |
| G08G | Traffic control systems | 468 |
| A61B | Diagnosis; Surgery; Identification | 427 |
| G16H | Healthcare informatics | 423 |
| B60R | Vehicles, vehicle fittings or vehicle parts | 379 |
| G01S | Radio direction-finding; radio navigation; radar | 362 |
| H04L | Transmission of digital information | 350 |
| G01C | Measuring distances, levels or bearings; navigation | 290 |
| G01R | Measuring electric variables; measuring magnetic variables | 242 |
| G01N | Investigating or analysing materials | 235 |
| H04W | Wireless communication networks | 232 |
| G05D | Systems for controlling or regulating non-electric variables | 228 |
| G07C | Time or attendance registers; registering vehicles | 205 |
| G01M | Testing static or dynamic balance of machines or structures | 194 |
Query 6 · 21.7 GB passFirst colour: vehicle-related subclasses. Non-exclusive
G06F (2,287), G06K (2,117), G06V (1,272) and G06T (1,141) are computing and image-processing neighbours of the core codes and do not describe a field of use; G06K carries the historic pattern-recognition groups that preceded G06F18. They are in the table and left out of the chart. One family usually carries several subclasses, so the figures cannot be added and do not partition the 9,479 families.
07 · Offices
The US is the most-used office, and the DPMA has drawn level in recent filings
5,463 of 9,479 families (57.6%) have a US member, ahead of the DPMA (4,757, 50.2%), the EPO (3,934, 41.5%), the PCT route (3,579, 37.8%) and China (3,535, 37.3%).
Families with a member at each office
Data table: filing offices
| Office | Families with a member | Share of 9,479 |
|---|---|---|
| US — United States | 5,463 | 57.6% |
| DE — Germany (DPMA) | 4,757 | 50.2% |
| EP — European Patent Office | 3,934 | 41.5% |
| WO — PCT route | 3,579 | 37.8% |
| CN — China | 3,535 | 37.3% |
| JP — Japan | 1,025 | 10.8% |
| KR — Korea | 605 | 6.4% |
| CA — Canada | 175 | 1.8% |
| AU — Australia | 173 | 1.8% |
| ES — Spain | 135 | 1.4% |
| GB — United Kingdom | 125 | 1.3% |
| BR — Brazil | 106 | 1.1% |
Query 7 · 21.7 GB passCounted across all members of an in-scope family; PCT rate 3,579 / 9,479 = 37.8%
In 2014–2017 filings the US was used by 63.3% of families against 55.3% for the DPMA. In 2018–2020 the gap widened to almost 18 points (66.4% against 48.6%). In 2021–2024 the two are level at 50.2% and 50.0%. China rose from 29.3% to 41.9% and settled at 36.1%; Japan's share rose from 9.5% to 12.0%.
Share of families in each filing period with a member at the office (%)
Data table: geographic shift by filing period
| Office | 2014–2017 (1,203 families) | 2018–2020 (3,378 families) | 2021–2024 (4,898 families) |
|---|---|---|---|
| US | 762 — 63.3% | 2,242 — 66.4% | 2,459 — 50.2% |
| DE | 665 — 55.3% | 1,643 — 48.6% | 2,449 — 50.0% |
| EP | 518 — 43.1% | 1,507 — 44.6% | 1,909 — 39.0% |
| WO | 453 — 37.7% | 1,196 — 35.4% | 1,930 — 39.4% |
| CN | 352 — 29.3% | 1,414 — 41.9% | 1,769 — 36.1% |
| JP | 114 — 9.5% | 324 — 9.6% | 587 — 12.0% |
| KR | 72 — 6.0% | 197 — 5.8% | 336 — 6.9% |
Query 7 · 21.7 GB passFamilies per period: 1,203, 3,378 and 4,898
Offices are non-exclusive: a family filed in Germany, at the EPO and in the US counts at all three, so the figures sum well above 9,479. The 2021–2024 period contains 2024 filings still being published, and national-phase entries from PCT applications filed in 2022–2024 are in many cases not yet due. Shares for the US, CN, JP and KR in that period are lower bounds.
08 · Partners
Most foreign co-applicants are arms of groups already on the German side
IBM Corp (221 families), Siemens, NEC and the Bosch and IBM subsidiaries in India and China are foreign arms of groups on the German side of the same filings. The external partners are mostly research institutions, plus F. Hoffmann-La Roche (15), Toyota Motor Europe (8) and Intel (6). The three universities partner with industry: Tsinghua and Carnegie Mellon appear alongside Robert Bosch on every joint family, and 20 of Nanyang's 28 are shared with Continental Automotive Technologies.
The fifteen most frequent non-German co-applicants
Data table: non-German co-applicants
| Co-applicant | Country | Joint families |
|---|---|---|
| IBM Corp | US | 221 |
| Huawei Technologies Co. Ltd | CN | 154 |
| Siemens Corp | US | 64 |
| Tsinghua University | CN | 53 |
| Siemens Ltd China | CN | 51 |
| NEC Corp | JP | 49 |
| Nanyang Technological University | SG | 28 |
| Bosch Global Software Technologies Private Ltd | IN | 26 |
| IBM China Investment Co. Ltd | CN | 25 |
| Carnegie Mellon University | US | 24 |
| Robert Bosch Engineering & Business Solutions Private Ltd | IN | 20 |
| F. Hoffmann-La Roche AG | CH | 15 |
| Huawei Cloud Computing Technologies Co. Ltd | CN | 10 |
| Continental Automotive France | FR | 10 |
| Dentsply Sirona Inc | US | 8 |
| Toyota Motor Europe | BE | 8 |
| Intel Corp | US | 6 |
| Toyota Jidosha KK | JP | 6 |
| Roche Molecular Systems Inc | US | 6 |
| Siemens Energy Inc | US | 5 |
| FEV Türkiye | TR | 5 |
| Yujin Robot Co. Ltd | KR | 4 |
| Technische Universiteit | NL | 4 |
| The General Hospital Corp | US | 4 |
| Kyndryl Inc | US | 4 |
Query 8 · 24.3 GB passGrey: foreign arm of a group on the German side. Non-exclusive
Huawei Technologies is the second-largest name at 154 families, with a different profile: no single Huawei-plus-German pair reaches ten families, the largest being nine. Its German-linked filings are shared with many individually named applicants resident in Germany, not with one German corporate entity. A family enters the table when at least one applicant is resident in Germany and at least one further applicant elsewhere.
09 · Pairs
1,546 families have more than one applicant, and the largest pairs are within one group
16.3% of the families carry more than one applicant. The largest pairs are internal, led by IBM Corp with IBM Deutschland (217). The largest industry–academia pairs all involve Robert Bosch: Tsinghua University (53), Carnegie Mellon (24) and the University of Freiburg (20). Outside Bosch: Continental with Nanyang (20), Eleqtron with the University of Siegen (12), Forschungszentrum Jülich with RWTH Aachen (9). Cross-brand automotive pairs are small, Audi with Volkswagen at most (18).
The fifteen most frequent applicant pairs
Data table: applicant pairs
| Applicant pair | Countries | Joint families |
|---|---|---|
| IBM Corp + IBM Deutschland | US / DE | 217 |
| Siemens AG + Siemens Corp | DE / US | 62 |
| Robert Bosch + Tsinghua University | DE / CN | 53 |
| Siemens Healthcare + Siemens Healthineers | DE / DE | 52 |
| Siemens AG + Siemens Ltd China | DE / CN | 51 |
| NEC Corp + NEC Europe | JP / DE | 49 |
| Bosch Global Software Technologies + Robert Bosch | IN / DE | 26 |
| IBM China Investment + IBM Deutschland | CN / DE | 25 |
| IBM China Investment + IBM Corp | CN / US | 25 |
| Carnegie Mellon University + Robert Bosch | US / DE | 24 |
| Robert Bosch Engineering & Business Solutions + Robert Bosch | IN / DE | 20 |
| Continental Automotive Technologies + Nanyang Technological University | DE / SG | 20 |
| Albert-Ludwigs-Universität Freiburg + Robert Bosch | DE / DE | 20 |
| Audi + Volkswagen | DE / DE | 18 |
| F. Hoffmann-La Roche + Roche Diagnostics | CH / DE | 12 |
| Eleqtron + Universität Siegen | — / DE | 12 |
| Porsche + Volkswagen | DE / DE | 12 |
| Audi + Porsche | DE / DE | 12 |
| Cariad + Robert Bosch | DE / DE | 11 |
| Continental Automotive France + Continental Automotive | FR / DE | 9 |
| Siemens AG + Siemens Mobility | DE / DE | 9 |
| Continental Automotive Technologies + Continental Autonomous Mobility | DE / DE | 9 |
| Huawei Technologies + an individual applicant resident in Germany | CN / DE | 9 |
| Forschungszentrum Jülich + RWTH Aachen | DE / DE | 9 |
Query 8 · 24.3 GB passFirst colour: industry–academia pairs; grey: pairs within one group or brand family
Pairs come from harmonised names, so a partner recorded under two variants appears in two rows. Albert-Ludwigs-Universität Freiburg is present under a short and a public-law-body variant, and a further pair with 17 families sits under the second one; the ranking is a lower bound for such partners. A pair is counted once per family, and a family with three applicants yields three pairs, so pair counts cannot be added to the 1,546 multi-applicant families.
10 · Grants
Japan grants 82.0% of its 2014–2020 families, the DPMA 23.9%
In the 2014–2020 cohort (4,581 families) examination is largely complete. The US grants 73.8% and the EPO 50.6%. One possible explanation for the German figure is that a DE filing often serves as a priority anchor for a later EP or PCT filing, so examination is never requested; the data cannot distinguish that from refusal or withdrawal. 4,757 families have a DE member and 3,934 an EP member, so the two routes run in parallel.
Granted families as a share of families at each office, 2014–2020 cohort (%)
Data table: grant rates by office
| Office | Families at office | Granted | Grant rate |
|---|---|---|---|
| JP — Japan | 438 | 359 | 82.0% |
| US — United States | 3,004 | 2,218 | 73.8% |
| AU — Australia | 78 | 48 | 61.5% |
| KR — Korea | 269 | 161 | 59.9% |
| CN — China | 1,766 | 1,004 | 56.9% |
| EP — European Patent Office | 2,025 | 1,025 | 50.6% |
| CA — Canada | 105 | 29 | 27.6% |
| DE — Germany (DPMA) | 2,308 | 552 | 23.9% |
| ES — Spain | 110 | 110 | 100.0% |
Query 7 · 21.7 GB passSpain in the table only: all 110 ES families are recorded as granted, which reflects national procedure. PCT excluded
Rates for filings from 2021 onwards are not shown: examination pendency would make them artificially low. Grant procedures differ between offices (deferred examination, utility-model conversion, national validation), so read each office against itself over time, not against the others.
11 · Citations
Siemens Healthcare draws 23.4 citations per family, five times Robert Bosch
Siemens Healthcare's 410 families draw twice the next value; Robert Bosch's 2,166 families attract 4.7 each. The two medical-imaging entities, Siemens Healthcare and Siemens Healthineers, account for 11,446 of the citations counted. Among larger portfolios SAP reaches 10.5 over 746 families and Siemens AG 6.8 over 900. The automotive suppliers ZF (2.8) and Valeo (3.9) and the manufacturers Daimler (3.0) and BMW (4.2) sit below the software and medical portfolios.
Forward citations per family, applicants with at least 60 families
Data table: citation impact
| Applicant | Families | Citations received | Per family |
|---|---|---|---|
| Siemens Healthcare | 410 | 9,589 | 23.4 |
| NEC Europe | 179 | 2,058 | 11.5 |
| Siemens Healthineers | 166 | 1,857 | 11.2 |
| SAP | 746 | 7,797 | 10.5 |
| Infineon Technologies | 62 | 630 | 10.2 |
| BASF | 69 | 666 | 9.7 |
| Conti Temic microelectronic | 60 | 506 | 8.4 |
| Fraunhofer-Gesellschaft | 161 | 1,347 | 8.4 |
| Audi | 146 | 1,214 | 8.3 |
| Siemens AG | 900 | 6,120 | 6.8 |
| Volkswagen | 333 | 2,151 | 6.5 |
| IBM Deutschland | 217 | 1,323 | 6.1 |
| Continental Automotive | 75 | 384 | 5.1 |
| Siemens Mobility | 71 | 354 | 5.0 |
| Robert Bosch | 2,166 | 10,227 | 4.7 |
| BMW | 358 | 1,508 | 4.2 |
| Valeo Schalter und Sensoren | 95 | 371 | 3.9 |
| Daimler | 168 | 497 | 3.0 |
| ZF Friedrichshafen | 204 | 575 | 2.8 |
| Rohde & Schwarz | 64 | 171 | 2.7 |
Query 9 · 32.6 GBFamily counts match the Leaders chapter exactly
Citations accumulate with age. A family filed in 2015 has had nine more years to be cited than one filed in 2024, and this chapter does not normalise for that, so portfolios weighted towards early filings score higher by construction. The measure counts distinct citing publications of any family member across all offices and does not separate examiner from applicant citations.
13 · How we counted
How the numbers were counted
- One family, one countA DOCDB family counts once, in the earliest filing year among its in-scope applications. The years sum to 9,479 families across 13,376 applications.
- German applicantsAt least one applicant with
person_ctry_code = 'DE'. This measures ownership: a German subsidiary of a foreign group is in scope, a foreign holding filing for a German team is not. - Completed code setIPC G06N and G06F18/00, subtree expanded by a recursive walk (112 symbols), plus the predecessors G06K9/62, G06F15/18 and G06N1/00: 115 symbols. Withdrawn symbols are kept.
- IPC as primary scheme, measuredA CPC-primary scope would miss 1,488 families (15.7%) that carry the codes only in IPC; IPC is kept because its reclassification chain makes the time series readable.
Checks and limits
- Two readingsCurrent-edition codes only give 8,231 families; with predecessors 9,479. The union buys recall at some cost in precision. The true figure lies between the two; this report uses the upper one.
- G06K9/00 excludedThe residual group feeds 188 successor symbols, only 24% of them in scope. A 24-family title sample of the 1,689 families it would add was off-topic in 18 cases. Predecessors were not themselves subtree-expanded, because the subgroups of G06K9/62 went to G06V.
- CPC-only families1,824 families carry the equivalent CPC codes but no IPC Layer 1 code and are outside the scope.
- Title-only familiesA title-keyword search finds 606 families (6.4% of the scope) that the classification does not reach. They are recorded as a gap, because absorbing them would cost more precision than it buys recall.
- Layer 2 lower bound213 CPC symbols found through the title breadcrumb and definitions, singular and plural searched separately. Offices that do not assign CPC are under-represented.
- Groups not consolidatedSiemens, Continental, Bosch and Daimler/Mercedes-Benz appear under several overlapping entries that cannot be added.
- 2024 incompleteAbout 18 months from filing to full publication. No growth claim rests on 2024.
- Grants and reclassificationGrant rates cover the 2014–2020 cohort only. The G06N3 subgroups and the G06F18 series are affected by ongoing reclassification and are not time series of invention.
- Precision24 random families read by title; two were not identifiable as machine learning, roughly 8%. No external series measures the same thing, so no external benchmark is quoted.
- ExclusionsUtility models and designs are excluded (
appln_kind <> 'D'), as are pseudo-families and families whose earliest in-scope filing falls outside 2014–2024. Breakdowns other than the annual series and the layer split are non-exclusive. - EditionsComputed on 15 August 2026 against IPC edition 20260101 and CPC edition 202608.
Code-set readings, layer coverage and the annual series
Query 1 · 29.9 GB passqueries.sql · Queries 0a to 0e build the code sets · Query 10 holds the precision samples and the negative test (28.3 GB)
Release protocol
- Agentsearch + draftBuilt the code sets and queried PATSTAT read-only. Every number comes from SQL.
- Checks13 rulesEvery claim must trace back to a query; chart and table must agree.
- Reviewsubject expertChecks scope, counts and wording before publication.
- Approvalsecond personA second person not involved in the report approves publication.
Next step
Which technology or region should we count for you?
A report like this one starts at 1,500 € plus VAT. Or learn to build it yourself on the EPO's Technology Intelligence Platform.