Machine Learning Patents from Germany

The headline reading of this landscape depends on how the code set is built. Counting only the codes that exist in the current IPC edition yields 8,231 families and an apparent 17-fold rise from 78 families in 2014. Adding the predecessor symbols those codes were carved out of yields 9,479 families, 211 in 2014, and a 6.7-fold rise. The second reading is the one used throughout this report; the difference is reclassification history, not invention. See the Methodology section.

Filing activity grew for seven consecutive years, from 211 families in 2014 to 1,417 in 2021. The steepest single-year increases fall in 2017 (+68.6%) and 2018 (+66.5%). After 2021 the series turns: 1,155 families in 2022, 1,347 in 2023. Measured 2014 to 2023, the compound annual growth rate is 22.9%.

The two lines are the same landscape counted with two code sets. The dashed line uses only the symbols that exist in the IPC edition in force today; the solid line adds the symbols those were carved out of. Where the two diverge, the gap is the pace at which offices have re-tagged older documents, not a change in filing behaviour. This report uses the solid line throughout.

Half the landscape sits in one branch: 4,699 of 9,479 families (49.6%) carry a G06N3 code for computing arrangements based on biological models, the group 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 families each, but they are not comparable in trajectory: G06N7 peaked at 85 families in 2019 and has declined since, while G06N10 rose from 1 family in 2014 to 152 in 2023. The Technology Evolution section separates the two.

Three shapes sit in one chart. Neural networks (G06N3) go from 19 families in 2014 to 747 in 2021 and then hold a plateau near 700 — the branch stopped accelerating but did not shrink. General machine learning (G06N20) follows the same curve one step lower, 16 to 509 and then a plateau near 450.

Quantum computing is the one series that is still climbing at the end of the window: 1 family in 2014, 44 in 2020, 152 in 2023. Probabilistic and fuzzy models (G06N7) move the other way — a peak of 85 families in 2019 and 30 in 2023 — consistent with a classical modelling branch being crowded out by learned models rather than growing alongside them.

Training beats topology. 2,443 families carry G06N3/08 (learning methods) against 1,401 for G06N3/04 (architecture) — German filings in this branch are more often about how a model is trained, validated or adapted than about the structure of the network itself.

Among the named architectures the order is combinations of networks (477), auto-encoder and encoder-decoder networks (168), convolutional networks (167) and recurrent networks (86). Among the named learning regimes it is supervised learning (187), backpropagation (208), non-supervised learning (119), reinforcement learning (103), transfer learning (80), adversarial learning (69) and distributed or federated learning (67). 186 families sit in G06N3/063, the hardware realisation of neural networks — a small but distinct hardware-side presence next to the predominantly method-side portfolio.

One portfolio carries almost a quarter of the landscape: Robert Bosch holds 2,166 of the 9,479 families (22.9%), more than the next two applicants together. Siemens AG follows with 900 and SAP with 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 is thin: one research organisation (Fraunhofer, 161), one chemicals company (BASF, 69), one semiconductor company (Infineon, 62) and one telecommunications operator (Deutsche Telekom, 51).

Nine of the twelve portfolios are led by neural networks. Three are not, and each in a different way. SAP files more in general machine learning (413) and knowledge-based models (232) than in neural networks (230) — the only portfolio here where symbolic and statistical methods outweigh connectionist ones. IBM Deutschland puts 149 of its 217 families into quantum computing, which is 30.2% of the entire German quantum sub-area from a single applicant. Daimler is concentrated in pattern recognition (134 of 168 families) rather than in the network branch.

Road vehicles are the largest field of application by a clear margin. 900 families carry B60W (conjoint control of vehicle sub-units), and five further vehicle-related subclasses follow within the top sixteen: traffic control (G08G, 468), vehicle fittings (B60R, 379), radar and radio navigation (G01S, 362), navigation and surveying (G01C, 290) and vehicle registering (G07C, 205).

Healthcare is the second cluster, split across two subclasses of comparable size: diagnosis and surgery (A61B, 427) and healthcare informatics (G16H, 423). Industrial automation follows — control and regulating systems (G05B, 661) plus control of non-electric variables (G05D, 228) — then enterprise and financial software (G06Q, 569), communications (H04L 350, H04W 232) and measurement and testing (G01R 242, G01N 235, G01M 194).

The United States is the most-used office overall: 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%).

The period split shows the distance closing. Across 2014–2017 filings the US was used by 63.3% of families against 55.3% for the DPMA, an 8-point gap. Across 2018–2020 filings the gap widened to almost 18 points (66.4% against 48.6%). Across 2021–2024 filings the two are level at 50.2% and 50.0%. China rose from 29.3% to 41.9% and then settled at 36.1%; Japan is the only office in the group whose share is higher in the last period (12.0%) than in the first (9.5%).

Most of the international network is intra-group. IBM Corp (221 families), Siemens Corp (64), Siemens Ltd China (51), NEC Corp (49), IBM China Investment (25), Bosch Global Software Technologies (26) and Robert Bosch Engineering & Business Solutions (20) are all foreign arms of groups that already appear on the German side of the same filings — the pairs table in the next section makes that explicit.

The genuinely external partners are research institutions and a small number of companies: Tsinghua University (53 joint families), Nanyang Technological University (28) and Carnegie Mellon University (24), plus F. Hoffmann-La Roche (15), Toyota Motor Europe (8) and Intel (6). All three universities partner with industry rather than with German academia: Tsinghua and Carnegie Mellon appear alongside Robert Bosch on every one of their joint families, and 20 of Nanyang’s 28 are shared with Continental Automotive Technologies.

Huawei Technologies is the second-largest name at 154 families, but its profile differs from IBM’s: no single Huawei-plus-German pair reaches ten families, the largest being nine. In this dataset Huawei’s German-linked filings are shared with many individually named applicants resident in Germany rather than with one German corporate entity.

1,546 of the 9,479 families (16.3%) carry more than one applicant. The largest pairs are internal: IBM Corp with IBM Deutschland (217 families), Siemens AG with Siemens Corp (62), Siemens Healthcare with Siemens Healthineers (52), Siemens AG with Siemens Ltd China (51) and NEC Corp with NEC Europe (49). These are group filings rather than collaborations between independent parties.

The largest genuine industry–academia pairs all involve Robert Bosch: Tsinghua University (53), Carnegie Mellon University (24) and the University of Freiburg (20). Outside Bosch the pattern is thinner — Continental Automotive Technologies with Nanyang Technological University (20), Forschungszentrum Jülich with RWTH Aachen (9) and Eleqtron with the University of Siegen (12) — the only pair in the top group in which neither partner appears elsewhere in the applicant ranking.

Cross-brand cooperation inside the German automotive sector is visible but small: Audi with Volkswagen (18), Porsche with Volkswagen (12), Audi with Porsche (12) and Cariad with Robert Bosch (11).

Restricted to the 2014–2020 filing cohort, where examination is largely complete, the outcome differs sharply by office. Japan grants 82.0% (359 of 438 families) and the United States 73.8% (2,218 of 3,004), while the European Patent Office reaches 50.6% (1,025 of 2,025) and the DPMA 23.9% (552 of 2,308).

The German figure is the outlier, and it is the office German applicants use most after the US. One possible explanation is that a DE national filing frequently serves as a priority anchor for a later EP or PCT filing, so examination is never requested; the data here cannot distinguish that from refusal or withdrawal after examination. What the data does show is that 4,757 families have a DE member while 3,934 have an EP member — the two routes are used in parallel rather than as alternatives.

Citation density and portfolio size point in different directions. Siemens Healthcare records 23.4 citations per family across 410 families — twice the next value and five times that of Robert Bosch, whose 2,166 families attract 4.7 each. The two medical-imaging entities together (Siemens Healthcare and Siemens Healthineers) account for 11,446 of the citations counted here.

Among the larger portfolios, SAP stands out at 10.5 citations per family over 746 families, and Siemens AG reaches 6.8 over 900. At the other end, the automotive suppliers ZF (2.8), Valeo (3.9) and the manufacturers Daimler (3.0) and BMW (4.2) sit below the level of the software and medical portfolios.

The domain tags find 6,361 German families that carry no core machine-learning code. Against the 9,479 families of the core search that is an increase of 67%, and it makes the hidden set 40.1% of the combined total of 15,867 families.

The share is not a recent phenomenon and it is not shrinking: domain-tagged-only families were 36.0% of the combined total in 2014 and 46.4% in 2023, with the lowest point at 30.9% in 2019. A landscape built on G06N and G06F18 alone would have missed at least three families in every ten in every single year, and close to five in the last three.

The overlap works in the other direction too: 2,916 of the 9,479 core families (30.8%) also carry at least one domain tag. For those, the tag says which field the method was applied in — which is what makes the Application Domains section readable at all.