A.1 Methodology for patent analysis
Patent identification and mapping to a definition
Transforming a technology definition into a patent data collection is one of the crucial steps in laying the foundation for a patent analysis. The most relevant aspect of a patent analysis is therefore the method of finding patents that fit the definition. Guidelines, such as those provided by WIPO (2015), can generally be followed with relevant keywords collected as well as patent classes and search concepts built that gather the patents to a definition within an optimal relation of precision and recall.
GenAI, however, is a modern concept without clear technical definition and where patent classes are not fully established yet. The most relevant patent classifications for GenAI are found in the Co-operative Classification Scheme(1)https://www.epo.org/en/searching-for-patents/helpful-resources/first-time-here/classification/cpc (CPC) G06N3/045 (“Auto-encoder networks; Encoder-decoder networks”), a sub-group of G06N3/02 Neural Networks.(2)https://worldwide.espacenet.com/patent/cpc-browser#!/CPC=G06N3/045 Most of the sub-groups in this area were newly introduced in 2023 and publications are currently being reclassified. There is not a proper class for GenAI itself yet.
GenAI can be viewed as a generic concept corresponding to the application of specific software methods to a large number of datasets (“modes”), addressing multiple applications. This is reflected in the use of many generic terms in patents today, often without defining the underlying technology used whose implementation is left to general knowledge of the skilled person in the art.
We had to develop a specific approach for capturing patents, involving the generation of digital entities, such as images, text or data through the use of specific machine learning algorithms. To achieve this, we used a two-stage approach: first, we combined classical patent searches together with prompts using our AI tool (see Appendices A.4 and A.5 for the patent searches and prompts) to retrieve a first patent dataset with high recall. Second, we refined the previous set using a trained BERT classifier to increase precision. This approach helps to avoid patents that are not “GenAI” according to the usually accepted definition, but that might generate products or other things (such as 3D printers or cameras) by using AI techniques somewhere in a process.
BERT (Bidirectional Encoder Representations from Transformers) is a large language model (LLM) developed by Google, is built on the Transformer architecture, and was introduced in 2018 (Devlin et al. 2018). In contrast to modern LLMs that only retain the decoder part of the original transformer architecture, BERT only keeps the encoder part to address discriminative tasks. BERT is pretrained on large amounts of text data. The training involves two main processes: 1. Masked language modeling, and 2. Next sentence prediction.
In masked language modeling, words in the input text are randomly selected and replaced with a special [MASK] token. The model is then trained to predict the original words based on the context provided by the surrounding words. This helps learning the bidirectional context of words in a sentence.
In next sentence prediction, the model is trained to predict for pairs of sentences whether the second sentence in a pair follows the first sentence. This task helps capturing relationships between sentences and understanding the overall context of the text input.
In order to use BERT in practical applications, it is fine-tuned to the specific task with labeled data in a supervised manner. BERT has proven to be highly effective and is utilized in various natural language processing tasks, including text classification, entity recognition, sentiment analysis and more.
There are different variants of BERT available, such as BERT Base, BERT Large, cased, uncased and there are models that are based on BERT but differ from it in terms of parameters, training methods, languages, such as RoBERTa, DistilBERT, XLM-RoBERTa etc. In this study, we use the BERT base uncased model with 110M parameters.
Sentence BERT
Sentence-BERT (SBERT) is a modification of the BERT model and is specifically designed for the calculation of text representations (text embeddings or text vectors).
The key idea behind SBERT is to fine-tune BERT on datasets that are crafted for sentence-level tasks. The model learns to produce embeddings that capture semantic information about sentences. Text similarities can be calculated based on calculations of the cosine-similarity of the embeddings. There is a maximum context length, which is typically about 512 tokens, so SBERT is not restricted to sentences in the literal sense. There are different types of SBERT models with varying performance in semantic search, encoding speed and model size.
Patent searches/prompts
To gather the basic dataset we have used three approaches:(3)The patent search definitions and prompts are listed in Appendix A.4.
Method 1a: A patent search using generic terms such as “generative AI” to locate patents using specific keyword concepts only. These concepts are searched more broadly or narrowly in different patent classifications.
Method 1b: Five patent searches for the specific search concepts: Generative Adversarial Networks (GAN), Autoregressive Models, Diffusion Models, LLMs and Variational Autoencoders. These five concepts are considered as almost a synonym to the concept “generative AI.”
Method 2: About one hundred prompts for various concepts of GenAI and its use, by using EconSight’s advanced AI search algorithms.
Patent classification
The fused datasets from the above searches are then transferred to a BERT classifier to increase accuracy and to identify patents that are really “generative AI.” This BERT is fine-tuned on this technology field as a classifier and is then be used to classify the patents according to the technology field into two classes:
Class 1: patent is GenAI
Class 2: patent is not GenAI
To achieve the BERT fine-tuning, we identified a training dataset of patents that can be assigned to GenAI (seed patents, label 1) and patents that cannot be assigned to GenAI (negative-seed patents, label 2). While the seed patents cover the whole range of GenAI topics, the negative seeds are a mixture of patents that are thematically very close to GenAI, but not exactly related to it. This means the selection of AI/machine learning patents that are not GenAI. A sample-based test of the accuracy of the BERT classifier delivers the following values: precision = 0.8, recall = 0.9, F1 score = 0.85.
Data collection, patent counting
Simple patent families are counted as a proxy for individual inventions in the report. A simple patent family is a set of patents in various countries in relation to a single invention. The technical content covered is considered to be identical. All patent documents have the same priority date or combination of priority dates. The first publication by a member of a patent family counts as the publication year.
Most analysis in the report refers to numbers of patent families.
Patent families generally include only patents and not utility models, without assessing their legal status.
Active patent portfolios: for this type of analysis, only active patent families (as in the INPADOC legal status definition) are counted. Active patent families are time-specific (active in a certain year) and highly relevant when analyzing a company’s patent strength.
The origin of the inventor (inventor’s location or residence) is used as a proxy for the source of innovations. For patents with multiple inventors, we count the different locations listed and count the location for multiple inventors of the same origin once.
A.2 Patent indicators
Patent applicant/owner
Patents are filed by an applicant, which can be an organization or a natural person. Applicants are not inventors, even if sometimes they are similar. The applicant is in most jurisdictions and in most cases published with the patent and remains always the applicant. The applicant is not automatically the owner of a patent at a given time, even if that is often the case. Patents can be transferred or sold, or the applicant itself can be sold as a company in a merger or takeover. Therefore the “owner” of a patent might change over time and it is not always published. For proper analysis, to consolidate incorrect spelling and to include merger and acquisition information in the analysis, the report used the ultimate owner concept in the IFI Claims global patent database. The most probable entity was then named as owner.
Patent family
A patent family is a collection of patent applications covering the same or similar technical content and all sharing one or more priority documents. Families are used to count inventions and not several patents corresponding to the same subject matter and filed in different jurisdictions. There are several definitions of patent families, including simple and extended patent families (EPO, n.d.; WIPO, 2015), depending on the number of priority documents shared (ranging from one to all priority documents). Patent family members are the individual patents filed in those jurisdictions where a patent applicant is seeking patent protection (e.g., WIPO, EPO) and all publications in relation to these (patent publications with code types A1, A2, B1 and so on). In the present study, we counted simple patent families (using a representative patent family member for each patent family), unless otherwise specified.
Inventor country
The origin of the inventor (inventor’s location or residence) is used as a proxy for the source of innovation. For patents with multiple inventors, we count the different locations listed and count the location for multiple inventors of the same origin once. If no inventor address is available, the patent priority country is used as a proxy for the source of innovation.
Filing country
Jurisdiction in which a member of a patent family has filed a patent application to seek patent protection.
A.3 Interdependence between models, modes and applications
A.4 Patent searches
GenAI models:
Generative AI total
(TITLEABSTRACTCLAIMS=(GENERATIVE* NEAR3 (AI OR ARTIFICIAL INTEL* OR ADVERSARIAL) OR GENERATIVE SEQ2 PRE_TRAINED NEAR3 (LANGUAGE SEQ2 MODEL OR TRANSFORMER?) OR CHAT_GPT OR VARIATIONAL SEQ2 AUTOENCODER? OR GENERATIVE SEQ2 ADVERSARIAL SEQ2 NETWORK* OR DIFFUSION SEQ2 PROBABILISTIC SEQ2 MODEL? OR AUTO_REGRESSIV* SEQ2 MODEL* ) OR (( IPC=(G06F 18/214, G06F 40/20, G06F 40/284, G06F 40/40, G06N 3/02, G06N 3/08, G06N 20/00, G06Q 10/04, G06V 10/70, G10L 13, G16H 30/40, H04L 51/02) OR TAG=("ECONSIGHT TECHNOLOGY FIELDS\IC5.3.9. NEURAL NETWORKS & DEEP LEARNING") OR CPC=(G06F 18/214, G06F 40/20, G06F 40/284, G06F 40/40, G06N 3/02, G06N 3/08, G06N 20/00, G06Q 10/04, G06T2207/20, G06V 10/70, G10L 13, G16H 30/40, H04L 51/02) ) AND ( TITLEABSTRACTCLAIMS=(GENERATIVE* NEAR3 (AI OR ARTIFICIAL INTEL* OR ADVERSARIAL) OR GENERATIVE SEQ2 PRE_TRAINED NEAR3 (LANGUAGE SEQ2 MODEL OR TRANSFORMER?) OR CHATGPT OR VARIATIONAL SEQ2 AUTOENCODER? OR CONVOLUTIONAL SEQ2 GENERATIVE SEQ2 ADVERSARIAL SEQ2 NETWORK* OR DIFFUSION SEQ2 PROBALISTIC* SEQ2 MODEL? OR DIFFUSIONAL SEQ2 NETWORK? OR DENOISING SEQ2 DIFFUSION SEQ2 PROBABILISTIC SEQ2 MODEL? OR GENERATIVE SEQ2 LATENT SEQ2 OPTIMIZATION OR NEURAL SEQ2 RADIANCE SEQ2 FIELD? OR AUTO_REGRESSIV* NEAR3 MODEL* OR GAN OR GANS OR GENAI OR VAE OR VAES OR (DENOISING OR VIDEO OR MODELS OR STABLE) NEAR3 DIFFUSION* SEQ2 MODEL* OR VARIATIONAL NEAR3 AUTOENCODER* OR GPT_3* OR GPT_4* ) OR ( TITLEABSTRACTCLAIMS=( ((IMAGE* OR TEXT* OR VIDEO* OR SPEECH* OR ("3D" MODEL*) OR GENE SEQ* OR DESIGN OR (PROGRAM* OR COMPUTER* OR SOFTWARE*) SEQ2 CODE* OR MUSIC* OR SPEECH* OR SCENE* OR MOLECULE* OR SYNTHETIC SEQ2 DATA* OR WORD SEQ3 SEQUENC*) NEAR5 (GENERATE* OR GENERATING* OR GENERATION* OR GENERATIV*)) OR GENERATIVE* NEAR3 MODEL*) AND ( TITLEABSTRACTCLAIMS=((TRANSFORMER* OR AUTOENCODER* OR LLM OR LARGE SEQ2 LANGUAGE SEQ2 MODEL* OR GAN OR GENERAT* SEQ2 ADVERSARIAL SEQ2 NETWORK* OR (AUTO_REGRESSIV* OR DIFFUSION SEQ2 PROBALISTIC) SEQ2 MODEL*)) OR TAG=("ECONSIGHT TECHNOLOGY FIELDS\C5.3.23. GAN, GENERATIVE ADVERSARIAL NETWORKS", "ECONSIGHT TECHNOLOGY FIELDS\IC5.3.27. AUTOREGRESSIV MODELS", "ECONSIGHT TECHNOLOGY FIELDS\IC5.3.28. VARIATIONAL AUTOENCODER,VAE", "ECONSIGHT TECHNOLOGY FIELDS\IC5.3.29. DIFFUSION MODELS", "ECONSIGHT TECHNOLOGY FIELDS\IC5.3.36. LARGE LANGUAGE MODELS,LLM") ))))) OR ( IPC=(G06N 3/0475) OR CPC=(G06N 3/0475))
Generative adversarial networks
(TITLEABSTRACTCLAIMS=(GENERATIVE NEAR5 ADVERSARIAL OR GAN OR (GENERATIVE SEQ2 ADVERSARIAL SEQ2 NETWORK*) OR (DUELING OR CONTRARIAN* OR ANTAGONISTIC OR ADVERSARIAL) NEAR3 (NEURAL SEQ2 NETWORK*)) ) AND ( FTERMSMART=(5L096/HA11) OR IPC=(G06N 3/02, G06V 10/70) OR CPC=(G06N 3/02, G06T2207, G06V 10/70))
Variational autoencoder
(CPC=(G06F 40/20, G06F 40/284, G06F 40/40, G06N 3/02, G06N 3/08, G06N 20/00, G06Q 10/04) OR IPC=(G06F 40/20, G06F 40/284, G06F 40/40, G06N 3/02, G06N 3/08, G06N 20/00, G06Q 10/04) OR TAG=("ECONSIGHT TECHNOLOGY FIELDS\IC5.3.9. NEURAL NETWORKS & DEEP LEARNING") ) AND TITLEABSTRACTCLAIMS=(VARIATIONAL NEAR5 AUTO_ENCODER* OR VAE )
Autoregressive models
(CPC=(G06F 40/20, G06F 40/284, G06F 40/40, G06N 3/02, G06N 3/08, G06N 20/00, G06Q 10/04) OR IPC=(G06F 40/20, G06F 40/284, G06F 40/40, G06N 3/02, G06N 3/08, G06N 20/00, G06Q 10/04) OR TAG=("ECONSIGHT TECHNOLOGY FIELDS\IC5.3.9. NEURAL NETWORKS & DEEP LEARNING") ) AND TITLEABSTRACTCLAIMS=(((AUTO_REGRESSIV* OR AUTO_REGRESSION OR SELF_REGRESSIV* OR SELF_REGRESSION* OR RECURSIVE_REGRESSION* OR ITERATIVE FORECASTING) NEAR3 MODEL*))
Large language models
(CPC=(G06F 40/20, G06F 40/284, G06F 40/40, G06N 3/02, G06N 3/08, G06N 20/00, G06Q 10/04, G10L 15/183) OR IPC=(G06F 40/20, G06F 40/284, G06F 40/40, G06N 3/02, G06N 3/08, G06N 20/00, G06Q 10/04, G10L 15/183) OR TAG=("ECONSIGHT TECHNOLOGY FIELDS\EC\5IC\IC5.3.9. NEURAL NETWORKS & DEEP LEARNING") ) AND TITLEABSTRACTCLAIMSDESCRIPTION=(LARGE SEQ2 LANGUAGE SEQ2 MODEL* OR LLM OR LARGE LANGUAGE MODEL* OR (LARGE LANGUAGE MODEL* OR LLM OR LARGE NEAR3 (LANGUAGE SEQ2 MODEL*) OR ((EXTENSIVE* OR MASSIVE OR LARGE OR GIGANTIC OR IMMENSE OR COLLOSSAL) SEQ2 (LANGUAGE OR LINGUISTIC OR SPEECH OR VERBAL) SEQ2 (MODEL* )) OR SUBSTANTIAL LANGUAGE PROCESSOR* ))
Diffusion models
(CPC=(G06F 40/20, G06F 40/284, G06F 40/40, G06N 3/02, G06N 3/08, G06N 20/00, G06Q 10/04) OR IPC=(G06F 40/20, G06F 40/284, G06F 40/40, G06N 3/02, G06N 3/08, G06N 20/00, G06Q 10/04) OR TAG=("ECONSIGHT TECHNOLOGY FIELDS\IC5.3.9. NEURAL NETWORKS & DEEP LEARNING") ) AND TITLEABSTRACTCLAIMS=(DIFFUSION NEAR5 MODEL* OR PROBALISTIC NEAR3 MODEL* OR (STABLE* OR DENOISING) NEAR3 DIFFUSION* OR (DIFFUSION NEAR3 MODEL* OR (PROGAGATION OR DIFFUSION ) SEQ2 (STOCHASTIC* OR PROBALISTIC* OR PROBABILITY) SEQ2 MODEL* OR PROPAGATION SEQ2 STOCHASTIC SEQ2 MODEL* OR DISPERSION SEQ2 STOCHASTIC SEQ2 MODEL* OR SCORE_BASED SEQ2 GENERATIVE SEQ2 MODEL*))
GenAI modes:
Image, video
( CPC=(A61B 5/1128, B23Q 17/249, G01M 11/065, G01N 15/1463, G01N 15/1475, G01N 21/8851, G01N2203/0647, G02B 7/365, G02B 21/244, G05D 1/0251, G06F 16/583, G06F 16/5862, G06F 16/70, G06F 16/78, G06F 16/783, G06F 16/7864, G06F2212/455, G06T, G06T 1, G06T 1/20, G06T 3, G06T 3/4046, G06T 5, G06T 7, G06T 9, G06T 9/002, G06T 11, G06T 13, G06T 15, G06T 17, G06T 19, G06T2207, G06T2207/00, G06T2207/20, G06T2207/20081, G06T2207/20084, G06V 10/70, H04N 5/2226, H04N 5/23229, H04N 5/23254, H04N2013/0074, Y10S 128/922, Y10S 707/914) OR FTERMSMART=(2H029/CD01, 2H029/DB12, 2H070/BB12, 2H095/AC02, 2H106/AA83, 2H109/BA06, 2K103/BB05, 2K203/GB22, 4C161/WW04, 4C601/JC, 5B057/DA, 5B057/DB, 5B057/DC, 5C020/AA13, 5C079/LA40, 5C122/FH, 5C122/FH17, 5C122/FH18) OR IPC=(G06F 16/58, G06F 16/583, G06F 16/70, G06F 16/78, G06F 16/783, G06T, G06T 1, G06T 1/20, G06T 3, G06T 5, G06T 7, G06T 9, G06T 11, G06T 13, G06T 15, G06T 17, G06T 19, G06V 10/70) ) OR( TITLEABSTRACTCLAIMS=((IMAGE OR VIDEO) NEAR3 (SYNTHES* OR CREAT* OR GENERAT*) OR IMAGE_TO_IMAGE OR IMAGE STYLE TRANSFER* OR TEXT_TO_IMAGE* OR VIDEO_TO_VIDEO*) ) AND ( IPC=(G06F, G06T, G06T 1, G06T 3, G06T 7, G06T 9, G06T 13, G06T 15, G06T 17, G06T 19, G06V) OR CPC=(A61B 5/1128, B23Q 17/249, G01M 11/065, G01N 15/1463, G01N 15/1475, G01N 21/8851, G01N2203/0647, G02B 7/365, G02B 21/244, G05D 1/0251, G06F, G06F 16/70, G06F2212/455, G06T, G06T 1, G06T 3, G06T 7, G06T 9, G06T 13, G06T 15, G06T 17, G06T 19, G06T2207, G06T2207/00, G06V, H04N 5/2226, H04N 5/23229, H04N 5/23254, H04N2013/0074, Y02D 10/00, Y10S 128/922, Y10S 707/914) )
Text
( CPC=(G05B2219/13106, G06F 16/243, G06F 16/24522, G06F 16/3329, G06F 16/3334, G06F 16/3335, G06F 16/3337, G06F 16/3338, G06F 16/3344, G06F 16/3347, G06F 16/345, G06F 16/36, G06F 16/367, G06F 16/374, G06F 16/90332, G06F 17/20, G06F 40, G06F 40/16, G06F 40/20, G06F 40/205, G06F 40/279, G06F 40/30, G06F 40/40, G06F 40/56) OR IPC=(G06F 16/36, G06F 17/20, G06F 40, G06F 40/16, G06F 40/20, G06F 40/205, G06F 40/279, G06F 40/30, G06F 40/40, G06F 40/56) ) OR ( TITLEABSTRACTCLAIMS=(PARAPHRASING OR (REWORDING OR REPHRASING) NEAR5 (WORD* OR SENTENCE* OR PARAGRAPH*) OR (SEMANTIC* OR NATURAL NEAR3 LANGUAGE) OR LANGUAGE NEAR3 (GENERAT* OR PRODUCTI*) OR TEXT* NEAR3 SUMMARI?ATI*) ) AND ( IPC=(G06F) OR CPC=(G06F, Y02D 10/00) )
Speech, music, voice
( CPC=(A63B2071/068, A63F2300/1081, B60G2401/19, B60R 25/257, B65H2551/132, B66B2201/4646, G01C 21/3608, G03G2215/00122, G05B2219/40531, G06F 3/167, G06F 16/7834, G10H 1/0025, G10H2210, G10H2240, G10H2250, G10L, G10L 13, G10L 15, G10L 15/08, G10L 15/24, G10L 15/26, G10L 17, G10L 25, G10L 99, H04M 1/642, H04M 1/6505, H04M2201/39, H04M2201/40, H04Q2213/13378, H04Q2213/378, Y10S 379/907) OR IPC=(G10L, G10L 13, G10L 15, G10L 15/08, G10L 15/24, G10L 15/26, G10L 17, G10L 25, G10L 99) OR FTERMSMART=(2C028/BB07, 2H270/QA36, 5B056/HH05, 5B089/KH16, 5C164/PA43, 5C164/PA46, 5D015, 5D045, 5D102/HC33, 5D108/BC17, 5H220/GG06, 5K015/AA06, 5K025/EE26, 5K027/HH20, 5K034/FF07, 5K038/GG04, 5K049/CC10, 5K127/CA27, 5K201/EC09) ) OR( TITLEABSTRACTCLAIMS=((VOICE OR SPEECH) NEAR3 (SYNTHESI* OR GENERAT* OR ANALYSIS*) OR SPEECH_TO_TEXT OR TEXT_TO_SPEECH* OR SPEECH* NEAR4 RECOG* OR VOICE* NEAR4 RECOG* OR VOICE* NEAR4 PRINT* OR VOICEPRINT*) ) AND ( IPC=(G06F) OR CPC=(G06F, G06F 3/01, Y02D 10/00) )
Software, code
( CPC=(G06F 8/00, G06F 8/20, G06F 8/30, G06F 8/40, G06F 8/60, G06F 8/70, G06F 11/36) OR IPC=(G06F 8/00, G06F 8/20, G06F 8/30, G06F 8/40, G06F 8/60, G06F 8/70, G06F 11/36) ) OR ( TITLEABSTRACTCLAIMS=((SOFTWARE OR CODE) NEAR3 (COMPLETION* OR GENERAT* OR DEVELOPMENT* OR PROGRAMMING)) ) AND ( IPC=(G06F) OR CPC=(G06F, Y02D 10/00) )
3D image models
(( CPC=(G06T 13/20, G06T 13/40, G06T 15, G06T 15/04, G06T 15/20, G06T 17, G06T 17/20, G06T 19, G06T 19/003, G06T 19/20) OR IPC=(G06F 3/04815, G06T 13/20, G06T 13/40, G06T 15, G06T 15/04, G06T 15/20, G06T 17, G06T 17/20, G06T 19, G06T 19/20) OR FTERMSMART=(5B050/BA09, 5B050/EA27) ) OR ( TITLEABSTRACTCLAIMS=(TEXT_TO_NERF OR TEXT_TO_3D OR (((THREE_D OR "3D") NEAR3 IMAGE*) NEAR3 MODEL*)) ))
Molecules, genes, proteins
( IPC=(C40B, C40B 10, C40B 20, C40B 30, C40B 30/00, C40B 30/02, C40B 30/04, C40B 30/08, C40B 30/10, C40B 40, C40B 50, C40B 50/02, C40B 60, C40B 70, C40B 80, C40B 99, G06F 19/10, G06F 19/16, G06F 19/18, G06F 19/22, G06F 19/28, G16B, G16B 20/00, G16B 30/00, G16B 40/00, G16C, G16C 10, G16C 20, G16C 60, G16C 99, G16CMISS) OR TITLEABSTRACTCLAIMSDESCRIPTION=((COMPUTATIONAL W2 CHEMISTRY) OR CHEMINFORMATICS OR ((AB D2 INITIO) AND CHEM*) OR (DENSITY D2 FUNCTIONAL D2 THEORY) OR (MOLECULAR D2 MECHANICS) OR (QUANTUM D2 CHEMISTRY) OR CHEMOINFORMATICS OR (DRUG* NEAR6 (DEVELOPMENT* OR TARGETING* OR DESIGN*)) NEAR10 (COMPUT* OR SOFTWARE* OR ALGORITHM*) OR (SYSTEMS* NEAR6 BIOLOGY*) OR PHARMACOPHOR*) OR CPC=(B01J2219/00689, B01J2219/00695, C04B2235/6026, C07K2299, C12N 15/1037, C12N 15/1093, C40B, C40B 10, C40B 20, C40B 30/00, C40B 30/02, C40B 30/04, C40B 30/08, C40B 30/10, C40B 40, C40B 50, C40B 50/02, C40B 60, C40B 70, C40B 80, C40B 99, G06F 19/10, G06F 19/16, G06F 19/18, G06F 19/22, G06F 19/28, G06F 19/70, G06F 19/706, G16B, G16B 20/00, G16B 30/00, G16B 40/00, G16C, G16CMISS, Y10S 423/05, Y10S 977/808) ) OR ( ( IPC=(C01, C07B, C07C, C07D, C07F, C07G, C08, C21) OR CPC=(C01, C07B, C07C, C07D, C07F, C07G, C08, C21) ) AND TAG=("ECONSIGHT TECHNOLOGY FIELDS\IC5.3.9. NEURAL NETWORKS & DEEP LEARNING") )
GenAI applications:
Physical sciences and engineering
TITLEABSTRACTCLAIMS= ((PHYSICAL NEAR2 SCIENCES) OR (ARCHAEOLOGY) OR (ASTRONOMY) OR (CHEMISTRY) OR ((EARTH OR ATMOSPHERIC) NEAR2 SCIENCES) OR (ENVIRONMENTAL NEAR2 SCIENCES) OR (COMPUTER_AIDED DESIGN) OR (PHYSICS) OR (MATHEMATICS) OR (ELECTRONICS) OR (WIRELESS DEVICE?)) OR IPC=(C OR D OR E OR F01 OR F02 OR F03 OR F04 OR F15 OR F16 OR F17) OR CPC=(G16C20/70 OR C OR D OR E OR F01 OR F02 OR F03 OR F04 OR F05 OR F15 OR F16 OR F17)
Industry and manufacturing
TITLEABSTRACTCLAIMS= (INDUSTRY OR INDUSTRIAL OR (SUPPLY NEAR3 CHAIN) OR MANUFACTURING OR (MACHINE NEAR3 TOOL?)) OR CPC=(G06Q10/06 OR G06Q10/08 OR G06Q50/04 OR G06Q50/28) OR IPC=(G06Q10/06 OR G06Q10/08 OR G06Q50/04 OR G06Q50/28)
Life and medical sciences
TITLEABSTRACTCLAIMS= ((LIFE NEAR2 SCIENCE?) OR HEALTH OR BIOLOGY OR HEALTHCARE OR MEDICAL OR (COMPUTATIONAL BIOLOGY) OR (MOLECULAR NEAR2 ("SEQUENCE ANALYSIS" OR EVOLUTION)) OR (RECOGNITION NEAR2 GENES) OR TRANSCRIPTOMICS OR "BIOLOGICAL NETWORKS" OR GENOTYPING OR PROTEOMICS OR GENOMICS OR BIOINFORMATICS OR METABOLOMICS OR METABONOMICS OR GENETICS OR PROTEOMICS OR TRANSCRIPTOMICS OR (DRUG DISCOVERY)) OR CPC=(G16B40 OR A61 OR G16H50/20) OR IPC=(A61 OR G16B40 OR G16H50/20)
Telecommunications
TITLEABSTRACTCLAIMS=(TELECOM? OR TELEPHON? OR PHONE? OR (COMMUNICATION? NEAR2 NETWORK?) OR RADIO OR PHONE? OR WIRELESS OR (COMMUNICATION NEAR2 SATELLITE?) OR TELEVISION) OR CPC=(H04L2012/5686 OR H04L2025/03464 OR H04L25/0254 OR H04L25/03165 OR H04L41/16 OR H04L45/08 OR H04N21/4662 OR H04Q2213/054 OR H04Q2213/13343 OR H04Q2213/343 OR H04R25/507) OR IPC=(H04L12/70 OR H04L25/02 OR H04L25/03 OR H04L12/24 OR H04L12/751 OR H04N21/466 OR H04R25)
Transportation
TITLEABSTRACTCLAIMS= (TRANSPORTATION OR VEHICLE? OR AEROSPACE OR SPACECRAFT OR SPACEFLIGHT OR ROADS OR AUTOMOBILE? OR AUTOMOTIVE? OR TRUCKS OR RAILWAYS OR TRAINS OR FREIGHT OR AIRWAYS OR WATERWAYS OR WATERCRAFT? OR AVIONICS OR AERONAUTICS OR AIRCRAFT? OR DRONE? OR UAV OR HELICOPTER? OR BOAT OR BOATS OR (BUS NEAR2 STATION?) OR AUTOBUS OR MOTORBUS OR STREETCAR OR TROLLEY) OR CPC=(B60W30/06 OR B60W30/10 OR B60W30/12 OR B60W30/14 OR B60G2600/1876 OR B60G2600/1878 OR B60G2600/1879 OR B62D015/0285 OR B64G2001/247 OR G06T2207/30248 OR G06T2207/30236 OR G05D001 OR B64C2201) OR IPC=(B60W30/06 OR B60W30/10 OR B60W30/12 OR B60W30/14 OR B62D15/02 OR B64G1/24 OR G05D1)
Energy management
TITLEABSTRACTCLAIMS=(((ENERGY OR POWER) NEAR2 (MANAGEMENT OR PLANNING OR CHALLENGE)) OR -GRID? OR ( NEAR2 GRID?)) OR CPC=(G01R 31/2846, G01R 31/2848, G01R 31/3651, G21, H01J2237/30427, H01M 8/04992, H02, H02H 1/0092, H02P 21/0014, H02P 23/0018, H03H2017/0208, H03H2222/04, H04W 52) OR IPC=(G21, H01M 8/04992, H02, H03H 17/02, H04W 52)
Agriculture
TITLEABSTRACTCLAIMS=((AGRICULTURE OR AGRICULTURAL OR CULTIVATE* OR BREEDING OR AGRONOMY OR PESTICIDE? OR AGROCHEMICHAL? OR FERTILIZER?)) OR CPC=(A01) OR IPC=(A01)
Security
TITLEABSTRACTCLAIMS= (SECURITY OR SURVEILLANCE OR (INVESTIGATION TECHNIQUES) OR (EVIDENCE COLLECTION) OR (NETWORK FORENSICS) OR (SYSTEM FORENSICS) OR (DATA RECOVERY) OR (COMPUTER FORENSICS) OR (BIOMETRICS) OR (CYBERSECURITY)) OR CPC=(G06F21 OR A61B5/117 OR H04W 12) OR IPC=(G06F21 OR A61B5/117 OR H04W 12)
Entertainment
TITLEABSTRACTCLAIMS= (ENTERTAINMENT OR ((VIDEO OR COMPUTER OR ELETRONIC OR ONLINE) NEAR2 (GAME? OR GAMING))) OR CPC=(A63) OR IPC=(A63)
Business solutions
TITLEABSTRACTCLAIMS=((ELECTRONIC NEAR2 (COMMERCE? OR "DATA INTERCHANGE" OR "FUNDS TRANSFER")) OR (ENTERPRISE NEAR2 (COMPUTING OR "INFORMATION SYSTEMS" OR "RESOURCE PLANNING" OR APPLICATIONS OR (ARCHITECTURE NEAR2 (MANAGEMENT OR FRAMEWORKS OR MODELING)) OR ONTOLOGIES OR TAXONOMIES OR VOCABULARIE OR "DATA MANAGEMENT" OR INTEROPERABILITY)) OR (CUSTOMER NEAR2 SERVICE?) OR (DIGITAL CASH) OR (E-COMMERCE INFRASTRUCTURE) OR (ONLINE NEAR2 (SHOPPING OR BANKING OR AUCTIONS)) OR (SECURE ONLINE TRANSACTIONS) OR (MARKETING) OR (VIDEO CONTENT DISCOVERY) OR (RECRUITMENT) OR (INTRANETS) OR (EXTRANETS) OR (DATA CENTERS) OR ((BUSINESS PROCESS) NEAR2 (MANAGEMENT OR MODELING OR MONITORING OR " CROSS-ORGANIZATIONAL")) OR (BUSINESS NEAR2 (INTELLIGENCE OR RULES)) OR ((SERVICE-ORIENTED OR IT OR EVENT-DRIVEN) SEQ2 ARCHITECTURES) OR (BUSINESS-IT ALIGNMENT) OR (IT GOVERNANCE) OR (INFORMATION NEAR2 (INTEGRATION OR INTEROPERABILITY))) OR CPC=(G06Q 10/10, G06Q 20, G06Q 30) OR IPC=(G06Q 10/10, G06Q 20, G06Q 30)
Military
TITLEABSTRACTCLAIMS=(MILITARY OR WARFARE OR CYBERWARFARE OR TACTICAL OR TACTICS OR ARMY OR WEAPON? OR BATTLE? OR BATTLEFIELD? OR PEACE OR PEACEKEEPING) OR CPC=(B63G, B64D 7, F41, F42, G01S 19/18) OR IPC=(B63G, B64D 7, F41, F42, G01S 19/18)
Education
TITLEABSTRACTCLAIMS= (EDUCATION OR EDUCATIONAL OR (DIGITAL NEAR2 LIBRARY) OR ((CHILD? OR CHILDREN OR PERSON OR PEOPLE OR STUDENT?) NEAR2 INSTRUCTION?) OR ((INTERACTIVE OR COLLABORATIVE OR DISTANCE) NEAR2 LEARNING) OR E-LEARNING OR (LEARNING MANAGEMENT SYSTEM?)) OR CPC=(G09B OR G06Q50/20) OR IPC=(G09B OR G06Q50/20)
Document management and publishing
(TITLEABSTRACTCLAIMS=((DOCUMENT NEAR2 (MANAGEMENT OR EDITING OR PROCESSING OR SEARCHING OR METADATA OR CAPTURE OR ANALYSIS OR SCANNING OR SCRIPTING OR PREPARATION)) OR (TEXT? NEAR2 (MANAGEMENT OR EDITING OR PROCESSING OR SEARCHING)) OR (VERSION CONTROL) OR (GRAPHICS NEAR2 (RECOGNITION OR INTERPRETATION)) OR (((OPTICAL CHARACTER) OR (ONLINE HANDWRITING)) NEAR2 RECOGNITION) OR (MARKUP LANGUAGE?) OR (HYPERTEXT LANGUAGE?) OR (ANNOTATION) OR ((MULTIMEDIA OR MIXED-MEDIA) NEAR2 CREATION) OR (IMAGE COMPOSITION) OR ((HYPERTEXT OR HYPERMEDIA) NEAR2 CREATION)) OR TITLEABSTRACTCLAIMS=((PUBLISHING OR (COPY NEAR2 EDITING) OR PUBLICATION? OR EDITORIAL) OR ((AFFECTIVE NEAR2 COMPUTING) OR (AFFECTIVE NEAR2 (RECOGNITION OR ESTIMATION OR STATE?)) OR ((ARTIFICIAL NEAR2 EMOTION*) NEAR2 INTELLIGENCE) OR ((PHYSIOLOGICAL NEAR3 MARKER) NEAR3 RECOGNITION) OR (EMOTION NEAR2 AI))) OR CPC=(A61B 5/165, G06F 40/10, G10L 25/63) AND IPC=(G06F 40/10, G10L 25/63))
Personal devices, computing and HCI
( TITLEABSTRACTCLAIMS=((PERSONAL NEAR2 COMPUTER?) OR (WORD NEAR2 PROCESSOR?) OR SPREADSHEETS OR MICROCOMPUTER? OR (HUMAN-MACHINE) OR (TOUCH NEAR2 SCREEN?) OR ((DISPLAY OR DISPLAYS) NEAR2 (TECHNOLOGY OR SYSTEM? OR APPARATUS)) OR (USER NEAR2 INTERFACE?)) OR TITLEABSTRACTCLAIMS=((AFFECTIVE NEAR2 COMPUTING) OR (AFFECTIVE NEAR2 (RECOGNITION OR ESTIMATION OR STATE?)) OR ((ARTIFICIAL NEAR3 EMOTION??) NEAR3 INTELLIGENCE) OR ((PHYSIOLOGICAL NEAR3 MARKER) NEAR3 RECOGNITION) OR (EMOTION NEAR2 AI)) ) OR IPC=(G10L 25/63) OR CPC=(A61B 5/165, G10L 25/63)
Banking and finance
TITLEABSTRACTCLAIMS=(FINTECH OR BANKING OR FINANCE OR FINANCING OR INSURANCE? OR REINSURANCE? OR INSURABLE? OR TRADING OR LIABILITY) OR CPC=(G06Q40) OR IPC=(G06Q40)
Arts and humanities
TITLEABSTRACTCLAIMS= ((FINE ARTS) OR (PERFORMING ARTS) OR (ARCHITECTURE NEAR2 BUILDING?) OR (LANGUAGE TRANSLATION) OR (MEDIA ARTS) OR (MUSIC?) OR CINEMA OR CINEMATOGRAPHY OR MOVIE OR WRITTING OR PAINTING? OR SCULPTING OR PHOTOGRAPHY OR THEATRE)
Computing in government
TITLEABSTRACTCLAIMS=(GOVERNMENT OR VOTING OR ELECTION OR E-GOVERNMENT OR (PUBLIC NEAR2 (POLICY OR POLICIES))) OR CPC=(G06Q50/26) OR IPC=(G06Q50/26)
Networks/smart cities
TITLEABSTRACTCLAIMS=(((SOCIAL OR DEVICE?) NEAR2 NETWORK?) OR IOT OR (INTERNET NEAR2 THINGS) OR SMART_CITY OR SMART_CITIES OR (SMART NEAR2 (CITY OR CITIES OR GRID? OR HOME? OR TRANSPORT? OR DEVICE? OR SENSOR?)) OR (VIRTUAL NEAR2 PLANTS))
Cartography
TITLEABSTRACTCLAIMS=(CARTOGRAPHY OR GEOGRAPHIC? OR TOPOGRAPHY OR TOPOGRAPHICS) OR IPC=(G06F 16/29) OR CPC=(G06F 16/29)
Industrial property, law, social and behavioral sciences
TITLEABSTRACTCLAIMS= (((BEHAVIOR OR BEHAVIORAL) NEAR2 SCIENCE?) OR (SOCIAL NEAR2 SCIENCE?) OR (LEGAL NEAR3 (STUDIES OR KNOWLEDGE OR INFORMATION? OR DOCUMENT? OR EVALUATION? OR CITATION? OR OPINION? OR TEXT? OR ARGUMENT? OR CONSULTANCY OR RIGHT? OR ISSUE? OR RISK? OR RESEARCH?? OR MATTER? OR CASE? OR JUDGMENT? OR DISCUSSION? OR CONCEPT? OR ACTION? OR STANDARD?)) OR LAWYER? OR JUDICIAL? OR LEGISLATION? OR ANTHROPOLOGY OR ETHNOGRAPHY OR PSYCHOLOGY OR ECONOMICS OR SOCIOLOGY)
A.5 Prompts
Almost 100 prompts for various concepts of GenAI and GenAI application areas were used in EconSight’s advanced AI search algorithms to help retrieve GenAI patents with high recall. The prompts below were used as the second stage in the GenAI patent retrieval approach discussed in detail in Appendix A.1.
Concept_1 = "Generative AI, generative artificial intelligence for 3D creation or three-dimensional designs."
Concept_2 = "Generative AI, generative artificial intelligence for accounting."
Concept_3 = "Generative AI, generative artificial intelligence in Biotech."
Concept_4 = "Generative AI, generative artificial intelligence for Advertisement."
Concept_5 = "Generative AI, generative artificial intelligence for AI understanding."
Concept_6 = "Generative AI, generative artificial intelligence for algorithm discovery."
Concept_7 = "Generative AI, generative artificial intelligence for analysts."
Concept_8 = "Generative AI, generative artificial intelligence for analytical tasks."
Concept_9 = "Generative AI, generative artificial intelligence for app design or creation."
Concept_10 = "Generative AI, generative artificial intelligence for app creation."
Concept_11 = "Generative AI, generative artificial intelligence for architecture."
Concept_12 = "Generative AI, generative artificial intelligence for bookkeeping."
Concept_13 = "Generative AI, generative artificial intelligence for brain communication."
Concept_14 = "Generative AI, generative artificial intelligence for converting brain signals to images."
Concept_15 = "Generative AI, generative artificial intelligence for converting brain signals to text."
Concept_16 = "Generative AI, generative artificial intelligence for business operations."
Concept_17 = "Generative AI, generative artificial intelligence for chatbots."
Concept_18 = "Generative AI, generative artificial intelligence for computer interaction."
Concept_19 = "Generative AI, generative artificial intelligence for contract analysis."
Concept_20 = "Generative AI, generative artificial intelligence for creating customer experiences."
Concept_21 = "Generative AI, generative artificial intelligence for cybersecurity."
Concept_22 = "Generative AI, generative artificial intelligence for design-to-code conversion."
Concept_23 = "Generative AI, generative artificial intelligence for diffusion modeling."
Concept_24 = "Generative AI, generative artificial intelligence for DNA prediction."
Concept_25 = "Generative AI, generative artificial intelligence for document-to-text conversion."
Concept_26 = "Generative AI, generative artificial intelligence for drug creation."
Concept_27 = "Generative AI, generative artificial intelligence for due diligence litigation."
Concept_28 = "Generative AI, generative artificial intelligence for education."
Concept_29 = "Generative AI, generative artificial intelligence for email generation."
Concept_30 = "Generative AI, generative artificial intelligence for fact-checking."
Concept_31 = "Generative AI, generative artificial intelligence for game design or creation."
Concept_32 = "Generative AI, generative artificial intelligence for gaming design or creation."
Concept_33 = "Generative AI, generative artificial intelligence with GANs (Generative Adversarial Networks)."
Concept_34 = "Generative AI, generative artificial intelligence for image editing."
Concept_35 = "Generative AI, generative artificial intelligence for image-to-text conversion."
Concept_36 = "Generative AI, generative artificial intelligence for image-to-video conversion."
Concept_37 = "Generative AI, generative artificial intelligence for latent diffusion."
Concept_38 = "Generative AI, generative artificial intelligence for law and lawyers."
Concept_39 = "Generative AI, generative artificial intelligence for LLMs (Large Language Models) in finance."
Concept_40 = "Generative AI, generative artificial intelligence for marketing."
Concept_41 = "Generative AI, generative artificial intelligence for medical doctors."
Concept_42 = "Generative AI, generative artificial intelligence for meme creation."
Concept_43 = "Generative AI, generative artificial intelligence for modeling."
Concept_44 = "Generative AI, generative artificial intelligence for molecule modeling."
Concept_45 = "Generative AI, generative artificial intelligence for music."
Concept_46 = "Generative AI, generative artificial intelligence for mute people communication."
Concept_47 = "Generative AI, generative artificial intelligence for generating presentations from language."
Concept_48 = "Generative AI, generative artificial intelligence for generating images from text prompts."
Concept_49 = "Generative AI, generative artificial intelligence for generating videos from text prompts."
Concept_50 = "Generative AI, generative artificial intelligence for protein design."
Concept_51 = "Generative AI, generative artificial intelligence for protein modeling."
Concept_52 = "Generative AI, generative artificial intelligence for real estate."
Concept_53 = "Generative AI, generative artificial intelligence for realistic data generation."
Concept_54 = "Generative AI, generative artificial intelligence for recommendation generation."
Concept_55 = "Generative AI, generative artificial intelligence for regulatory compliance."
Concept_56 = "Generative AI, generative artificial intelligence for Reinforcement Learning."
Concept_57 = "Generative AI, generative artificial intelligence for RPA (Robotic Process Automation)."
Concept_58 = "Generative AI, generative artificial intelligence for social media content."
Concept_59 = "Generative AI, generative artificial intelligence for speech."
Concept_60 = "Generative AI, generative artificial intelligence for speech imitation."
Concept_61 = "Generative AI, generative artificial intelligence for stable diffusion."
Concept_62 = "Generative AI, generative artificial intelligence for storytelling."
Concept_63 = "Generative AI, generative artificial intelligence for summarization."
Concept_64 = "Generative AI, generative artificial intelligence for tattoo generation."
Concept_65 = "Generative AI, generative artificial intelligence for tax documents."
Concept_66 = "Generative AI, generative artificial intelligence for text-to-3D conversion."
Concept_67 = "Generative AI, generative artificial intelligence for text-to-app conversion."
Concept_68 = "Generative AI, generative artificial intelligence for text-to-code conversion."
Concept_69 = "Generative AI, generative artificial intelligence for text-to-design conversion."
Concept_70 = "Generative AI, generative artificial intelligence for text-to-image conversion."
Concept_71 = "Generative AI, generative artificial intelligence for text-to-level conversion."
Concept_72 = "Generative AI, generative artificial intelligence for text-to-medical advice."
Concept_73 = "Generative AI, generative artificial intelligence for text-to-slides conversion."
Concept_74 = "Generative AI, generative artificial intelligence for text-to-software conversion."
Concept_75 = "Generative AI, generative artificial intelligence for text-to-speech conversion."
Concept_76 = "Generative AI, generative artificial intelligence for text-to-video conversion."
Concept_77 = "Generative AI, generative artificial intelligence for text-to-voice conversion."
Concept_78 = "Generative AI, generative artificial intelligence for text-to-website conversion."
Concept_79 = "Generative AI, generative artificial intelligence with Transformers."
Concept_80 = "Generative AI, generative artificial intelligence for travel maps."
Concept_81 = "Generative AI, generative artificial intelligence for urban planning."
Concept_82 = "Generative AI, generative artificial intelligence for user interface creation."
Concept_83 = "Generative AI, generative artificial intelligence for vector generation from video."
Concept_84 = "Generative AI, generative artificial intelligence for video creation."
Concept_85 = "Generative AI, generative artificial intelligence for video-to-image conversion."
Concept_86 = "Generative AI, generative artificial intelligence for video-to-text conversion."
Concept_87 = "Generative AI, generative artificial intelligence for video to 3D."
Concept_88 = "Generative AI, generative artificial intelligence for voice cloning."
Concept_89 = "Generative AI, generative artificial intelligence for website creation."
Concept_90 = "Generative AI, generative artificial intelligence."
Concept_91 = "Generative AI, generative artificial intelligence with neural style transfer (NST)."
Concept_92 = "Generative AI, generative artificial intelligence with diffusion models."
Concept_93 = "Generative AI, generative artificial intelligence with variational autoencoder (VAE)."
Concept_94 = "Generative AI, generative artificial intelligence with autoregressive models."
Concept_95 = "Generative AI, generative artificial intelligence with large language models (LLMs)."
Concept_96 = "Generative AI, generative artificial intelligence with GPT-3, Chat GPT."
Concept_97 = "Generative AI, generative artificial intelligence with GPT-4."
A.6 Scientific publication query with The Lens
We used The Lens (Cambia 2024) as a bibliographical analytics tool for scientific publications. This is a free service, using trusted open bibliographical sources with an extensive coverage of scientific publications, making possible an easy reproducibility of the present analysis.
The base query has been built as follows:
Fields Search terms Comment
publication year range 2010-2023 2023 is entirely included, 2024 is excluded
title, abstract, keyword, field_of_study "GENERATIVE AI" OR "GENERATIVE ARTIFICIAL INTELLIGENCE" OR "LARGE LANGUAGE MODEL" OR "LARGE LANGUAGE MODELS" OR "GENERATIVE ADVERSARIAL NETWORK" OR "GENERATIVE ADVERSARIAL NETWORKS" OR "GENERATIVE PRE-TRAINED TRANSFORMER" OR "DEEP GENERATIVE MODEL" OR "DEEP GENERATIVE MODELS" OR "DEEP GENERATIVE MODELING" OR "LARGE GENERATIVE MODEL" OR "LARGE GENERATIVE MODELS" general terms related to generative AI
"GPT-3" OR "CHATGPT" OR "GPT-4" OR "GPT4" OR ("LLAMA" AND "LANGUAGE MODEL") OR "DALL-E" OR "ALPHAFOLD" OR "GITHUB COPILOT" mainstream foundation model names
note: the search term “GPT3” (without dash separating the version number) is an ambiguous term (also a meteorological/climate model), so it was removed.
"VARIATIONAL AUTOENCODER" OR "VARIATIONAL AUTOENCODERS"
OR "NEURAL RADIANCE FIELD" OR "NEURAL RADIANCE FIELDS"
OR "DENOISING DIFFUSION PROBABILISTIC MODEL" OR "DENOISING DIFFUSION PROBABILISTIC MODELS" related to image/3D generation
"TEXT-TO-IMAGE GENERATION" OR "GENERATIVE TEXT-TO-IMAGE MODELS" OR "GENERATIVE TEXT-TO-IMAGE MODEL" related to text-based image generation
"PROMPT ENGINEERING" OR "LLM PROMPTING" OR "TEXT-TO-IMAGE PROMPT” related to model prompting
"RETRIEVAL AUGMENTED GENERATION" OR "GENERATIVE CONVERSATIONAL AI" OR "GENERATIVE CONVERSATIONAL MODELS" OR "GENERATIVE CONVERSATIONAL MODEL" related to conversational system/AI assistant
Title, keyword, field_of_study "GENERATIVE MODELS" OR "GENERATIVE MODEL" more restrictive to avoid some noise
The number of total results for the base query is 75,870 deduplicated scientific publications, as executed on January 20, 2024 with The Lens.
A.7 Mining software and dataset mentions in the non-patent literature corpus
In scientific publications, software and datasets are usually not formally cited like journal or conference articles. Citations to software and datasets are mostly informal mentions in the body of the text. As a consequence, they are invisible and unavailable in the large bibliographical and citation databases like Web of Science or Scopus. To measure realistic impact of datasets and software, it is necessary to identify their mentions with text mining techniques applied to the full texts of NPL documents.
The corpus of scientific publications used in the study is a set of 75,870 deduplicated publications produced as explained in Appendix 6. We used the open access subset of this corpus to identify automatically the mentions of software and datasets. These mentions can be then exploited to estimate what are the most impactful software and datasets in GenAI.
Our methodology relies on mature text mining tools resulting from several years of developments. These techniques are well-evaluated and able to scale to thousands of documents at a reasonable cost. The process is as follows:
Document full text acquisition: The 75,870 publications of our corpus include 48,784 Open Access publications according to lens.org. We successfully downloaded 34,183 PDFs out of the full open access subset using the Unpaywall database (https://unpaywall.org).
Text mining of mentions: The PDFs were processed by two tools, Softcite for extracting the software mentions (Softcite (2018–2024)) and DataStet for extracting dataset mentions DataStet (2022–2024).
Aggregation: The extracted mentions (789,218 software mentions and 978,297 dataset mentions) were then aggregated and we kept the top 500 datasets and top 500 software.
Data cleaning: We manually corrected these two sets to remove errors and merge the same software mentioned with name variants not handled at aggregation time.
The two mentioned recognizers rely on deep learning techniques, a BERT model called SciBERT, trained on scientific content. Softcite was trained on 5,000 manually annotated scientific articles and DataStet on two sets of respectively 22,000 and 6,000 manually annotated sentences. They perform with high accuracy, around 81% F1-score for software names and 89% F1-score for dataset names. More details on the implementation, evaluation and application of these tools are available in Lopez et al. (2021) and in Bassinet et al. (2023).
While the mentions were extracted from documents corresponding to less than half of the whole GenAI corpus, this is a very large and representative subset. Open Access publications are dominantly from scientific publishers (gold open access) and are not associated to loss of quality (STM 2023). Therefore, we can consider that the derived statistics provide a good approximation of the dataset and software relative impact.
A.8 Descriptions/example patents for GenAI applications
Software/other applications
Many GenAI patent families cannot be assigned to a specific application based on patent title, abstract and claims. These patents often describe topics such as search engines or chatbots that can in theory be used for many different use cases. These patents are assigned to the category “Software/other applications.” In addition, there are also more specific use cases of GenAI within the software space. For example, it can help to automate many of the tasks involved in software development, such as coding, testing and debugging. This could free up developers to focus on more creative and strategic tasks. It can also allow users to code without programming skills because text-based instructions can be used to create simple applications. Companies such as Microsoft are already selling GenAI for software engineering, including GitHub Copilot, which is now integrated with OpenAI’s GPT-4. These patents are instead collected in the category software.
Example patent: WO2023172817 – SYSTEMS AND METHODS FOR A CONVERSATIONAL FRAMEWORK OF PROGRAM SYNTHESIS
Applicant: Salesforce
Abstract: Embodiments described herein provide a program synthesis framework that generates code programs through a multi-turn conversation between a user and a system. Specifically, the description to solve a target problem is factorized into multiple steps, each of which includes a description in natural language (prompt) to be input into the generation model as a user utterance. The model in turn synthesizes functionally correct subprograms following the current user utterance and considering descriptions and synthesized subprograms at previous steps. The subprograms generated at the multiple steps are then combined to form an output of program in response to the target problem.
Image:
Life sciences
GenAI has the potential to have a profound impact on life sciences in a number of ways: for example, biotech and pharma companies have begun to use GenAI foundation models in their research and development for what is known as generative design. Foundation models can speed up the process of developing new drugs by screening and designing molecules best suited for new drug formulation. Another promising area is personalized medicine. The ability to generate insights and patterns from vast quantities of patient data will spark more personalized treatments. GenAI can also be used to create models of individual patients’ genomes to predict their response to drugs. As a result, The McKinsey Global Institute (MGI) has estimated that the technology could generate US$60 billion to US$110 billion a year in economic value for the pharma and medical-product industries (McKinsey 2024b).
Example patent: US20230115171 – SYSTEM AND METHOD FOR GENERATING CUSTOMIZABLE MOLECULAR STRUCTURES FOR DRUG DISCOVERY
Applicant: Innoplexus
Abstract: A system and method for generating customizable molecular structures for drug discovery. The system includes a processor communicably coupled to a memory and executes a deep neural network based molecular encoding model. The processor receives input datasets of drug-like molecules from private and public databases and are employed as training dataset. The processor further executes a plurality of deep generative models configured to receive input data relating to small molecules which includes desirable molecules and undesirable molecules. The plurality of deep generative models generates molecular structures like the input desirable molecules. The deep neural network based molecular encoding model is configured to map similarities between the molecular structures generated. The deep neural network based molecular encoding model computes intra-model and inter-model distances. Further, the deep neural network based molecular encoding model samples the molecular structures generated from the plurality of deep generative models to obtain desired molecular structure.
Image:
Document management and publishing
GenAI can be used to improve and automate document management and publishing. For example, it can help with the generation of documents such as contracts, invoices and reports. This can save businesses time and money. In addition, GenAI tools can create attractive marketing materials tailored to the interests of the target audience. GenAI is also capable of identifying and extracting key information from documents.
Example patent: US20230229866 – SYSTEM AND METHOD FOR MANAGEMENT OF LIFE CYCLE OF CONTRACTS
Applicant: Tata Consultancy Services
Abstract: Contracts are a fundamental tool for coordinating economic activity and need to be managed throughout the lifecycle of contracts. The existing methods are incomplete, expensive, time-consuming, and error-prone. A method and system for management of lifecycle of contracts have been provided. The system leverages a combination of artificial intelligence (AI) techniques appropriate for different micro services in contract lifecycle management. The deep learning and natural language processing (NLP) techniques help in understanding of clauses of the contract, risk levels involved in the contract. The system is configured to automatically generate contracts for a customer based on the other criteria of the customer. The system also identifies alternate options to risky clauses and mandatory clauses to be included. The system is also configured to manage the workflows based on context of contract to seek exception approvals from appropriate stakeholders during contract creation and alert appropriate stakeholders on delivery governance issues.
Image:
Business solutions
GenAI is expected to play an important part in many business areas. For example, it can transform the way organizations manage internal knowledge. I can empower employees to seamlessly access and retrieve stored knowledge by posing queries in a natural, conversational manner. This capability enables teams to swiftly gather relevant information, enabling them to make sound decisions and devise effective strategies with greater agility. GenAI will also transform the customer operations landscape, enhancing both efficiency and customer satisfaction. The technology has already garnered significant adoption in the customer service domain, for example as GenAI–fueled chatbots that give immediate and personalized responses to complex customer inquiries.
Example patent: WO2022201195 – RETAIL ASSISTANCE SYSTEM FOR ASSISTING CUSTOMERS
Applicant: RN Chidakashi Technologies Private Limited
Abstract: A system and method for retail assistance system (102) for assisting customers while shopping in a retail store. The retail assistance system (102) is configured to detect one or more customers entering the retail store using an input unit, determine a personality profile of the one or more customers by analyzing a facial expression and one or more personal attributes of the one or more customers, determine one or more personalized recommendations for the one or more customers by analyzing the personality profile, past purchase history of the one or more customers, and visit history of the one or more customers in the retail store using a machine learning model, and enable the at least one of customer to choose the one or more personalized recommendations.
Image:
Industry and manufacturing
There are numerous ways in which GenAI can play a role in various industries and in the manufacturing process. Traditional AI has already long been used for tasks such as anomaly detection, production analytics or setpoint optimization. GenAI now enables new features. For example, GenAI can be used to optimize designs of new products leading to cost reductions in production. Manufacturers are also adopting GenAI application programming interfaces to connect design and engineering tools to build digital twins of their facilities (Shapiro 2023). The Boston Consulting Group has identified three ways that GenAI can support the path to the factory of the future (BCG 2023):
Assistance system: GenAI improves the efficiency of hands-on tasks such as programming or machine maintenance.
Recommendation system: GenAI tools can also provide recommendations that help workers to identify the best methods for specific tasks, for example, by automatically creating text or images that provide maintenance instructions.
Autonomous system: The third role of GenAI in factories is the role of autonomous systems. For example, GenAI will enable robots to translate operator prompts into a sequence of actions that the system then executes to perform material handling tasks. This would reduce the need for task- and environment-specific training, data labeling and frequent retraining. It therefore has the potential to reduce engineering costs, replace manual activities and increase productivity.
Example patent: US20220035346 – PREDICTIONS FOR A PROCESS IN AN INDUSTRIAL PLANT
Applicant: ABB
Abstract: To generate real-time or at least near real-time predictions for a process in an industrial plant, a set of neural networks are trained to create a set of trained models. The set of trained models is then used to output the predictions, by inputting online measurement results in an original space to two trained models whose outputs are fed, as reduced space inputs and reduced space initial states, to a third trained model. The third trained model processes the reduced space inputs to reduced space predictions. They are fed to a fourth trained model, which outputs the predictions in the original space.
Image:
Transportation
GenAI has many use cases in transportation. For example, it will play a key role in autonomous driving. For this purpose, synthetic data are being employed to train autonomous vehicles, allowing them to be thoroughly tested in a realistic 3D virtual world (Nvidia 2023). GenAI models also allow the simultaneous generation of multiple scenarios, the prediction of future vehicle trajectories, and advancement of decision reasoning chains. These approaches enhance safety, efficiency and flexibility while significantly reducing risk and associated costs. Another use case will be in-car personal assistants with GenAI skills promising to boost navigation and infotainment to new heights. GenAI can also optimize public transportation systems. By analyzing vast amounts of data on factors like population density, traffic patterns and passenger preferences, AI algorithms can devise more efficient routes and schedules for public transit networks.
Example patent: US20210294341 – METHOD AND APPARATUS FOR GENERATING U-TURN PATH IN DEEP LEARNING-BASED AUTONOMOUS VEHICLE
Applicants: Hyundai, Kia
Abstract: A method for generating a U-turn path in an autonomous vehicle includes calculating a drivable area, generating multiple paths drivable in the drivable area, filtering a driving strategy path among the multiple paths based on deep learning, and determining a final path from the filtered candidate paths.
Image:
Security
The rise of GenAI in cybersecurity presents both challenges and opportunities. On the one hand, GenAI is used for valuable applications in cybersecurity, ranging from assisting threat hunters in retrieving relevant data for ongoing investigations to providing real-time insights that enhance vulnerability management processes. Similar to how GenAI can identify and replicate patterns in language, it can recognize and analyze patterns in cybersecurity threats and vulnerabilities. GenAI also plays a crucial role in enhancing system security by generating intricate, unique passwords or encryption keys that are nearly impenetrable to unauthorized access or decryption (Stanham 2023). On the other hand, there is also a clear downside – just as enterprises are harnessing GenAI to strengthen their cybersecurity defenses, cybercriminals are also employing this technology to devise sophisticated attack strategies that can bypass existing security measures. Mentions of GenAI on the dark web have risen significantly in 2023 (Ali and Ford 2023). Therefore, GenAI will accelerate the arms race between hackers and companies.
Example patent: US20200068398 – USE OF GENERATIVE ADVERSARIAL NETWORKS (GANs) FOR ROBUST TRANSMITTER AUTHENTICATION
Applicant: IBM
Abstract: A method is provided for transmitter authentication including generating a noise vector using a generative adversarial network generator model, wherein a signature of a first transmitter is embedded into a signal output by the first transmitter based at least on the noise vector; and using the signature to identify the first transmitter.
Image:
Telecommunications
In the telecom industry, GenAI can be used to transform various aspects of network operations, customer experience and service delivery. For example, telecommunications providers can leverage GenAI to streamline network operations and enhance customer satisfaction. By employing AI algorithms, network bottlenecks can be identified proactively, resource allocation can be optimized, and potential maintenance requirements can be anticipated. Additionally, GenAI can generate tailored service recommendations for customers, such as data plan upgrades or value-added services.
Example patent: US20230101761 – METHOD AND APPARATUS FOR DYNAMIC TONE BANK AND PERSONALIZED RESPONSE IN 5G TELECOM NETWORK
Applicant: IBM
Abstract: Generating a personalized automated voice response in a telecommunications network is provided. An incoming call from a caller for user equipment of an operator in the telecommunications network is identified. In response to identifying the incoming call, it is determined whether to provide an automated response to the incoming call. In response to determining to provide the automated response to the incoming call, a personalized response message from the operator of the user equipment to the caller is generated based on characteristics of communications between the caller and the operator of the user equipment. The personalized automated voice response comprising the personalized response message in a synthesized voice of the operator of the user equipment is generated. The personalized automated voice response is sent to the caller.
Image:
Personal devices
GenAI will also increasingly play an important role for personal devices such as mobile phones. GenAI models can run on a smartphone and analyze the data of one’s device to anticipate one’s next move – essentially transforming one’s phone into a personal AI assistant. For example, Samsung Electronics revealed a new GenAI model called Samsung Gauss in November 2023 that is designed for AI applications on devices and will soon be available on new products (Samsung 2023). It can, among other things, help compose emails, translate content and generate or edit images.
Example patent: US20230223008 – METHOD AND ELECTRONIC DEVICE FOR INTELLIGENTLY READING DISPLAYED CONTENTS
Applicant: Samsung Electronics
Abstract: A method for intelligently reading displayed contents by an electronic device is provided. The method includes obtaining a screen representation based on a plurality of contents displayed on a screen of the electronic device. The method includes extracting a plurality of insights comprising at least one of intent, importance, emotion, sound representation and information sequence of the plurality of contents from the plurality of contents based on the screen representation. The method includes generating audio emulating the extracted plurality of insights.
Image:
Banking and finance
Banks have been using traditional AI tools to automate tasks, generate predictions, detect fraud or for marketing purposes for many years. GenAI now brings new opportunities for established banks as well as for emerging fintech companies. For example, GenAI can serve as virtual experts to boost employee performance. For instance, Morgan Stanley is developing an AI assistant powered by GPT-4, designed to assist tens of thousands of wealth managers in swiftly locating and synthesizing answers from a comprehensive internal knowledge base (McKinsey 2023). GenAI can also improve the way banks manage their back-office operations. For example, GenAI tools can assist customer service representatives by transcribing conversations, extracting relevant information and generating notes during phone calls, allowing the human agent to concentrate on providing better customer service.
Example patent: CN116521840 – FINANCIAL QUESTION AUTOMATIC EXTRACTION AND REPLY METHOD AND SYSTEM BASED ON CONVOLUTIONAL NEURAL NETWORK
Applicant: Ping An Bank (Ping An Insurance)
Abstract: The invention provides a convolutional neural network-based financial question automatic extraction and reply method, which is characterized in that the convolutional neural network-based financial question automatic extraction and reply method comprises the steps of retrieving a financial question containing a preset keyword from a preset channel; transcoding the financial question according to a first preset algorithm to form a question code; inputting the question code into a trained machine learning model to obtain answer information; and uploading and filling the answer information through the preset channel. In addition, the invention also provides a system and computer equipment thereof. According to the technical scheme, the problem that existing network financial question answering adopts manpower or cannot be intelligently answered is effectively solved.
Physical sciences and engineering
GenAI is expected to have an impact on various engineering and physical sciences disciplines, with the potential to enhance research, design and innovation. For example, by analyzing vast amounts of data on material properties and structures, AI algorithms can identify patterns and relationships that lead to the design of new materials with desired properties such as strength, conductivity and durability. However, there are also challenges regarding the use of GenAI in product design as a recent study by the MIT showed (MIT 2023). GenAI models are trained to replicate the patterns and characteristics of a given dataset. Therefore, this approach is effective in mimicking existing designs, but it may not align with the creative goals of engineers and designers who seek to introduce novel innovations. But in design, divergence from existing norms and the pursuit of unique concepts are sometimes crucial for innovation.
Example patent: US20200159886 – ARTIFICIAL INTELLIGENCE-BASED MANUFACTURING PART DESIGN
Applicant: Boeing
Abstract: Systems, methods, and apparatus for artificial intelligence-based manufacturing part design are disclosed. A system for designing a part comprises at least one processor configured: to encode the desired part design to generate an encoded desired part design; to identify a group of part designs within a space that is similar to the desired part design by comparing the encoded desired part design to encoded realized part designs, encoded imagined part designs, real metadata, and imagined metadata within the space; to generate an encoded optimal part design by analyzing the group of part designs according to objectives and weightings provided by a user; and to decode the encoded optimal part design to generate an optimal part design. Further, the system comprises a display configured to display, to the user, the optimal part design, which the user may use as a guide to modify the desired part design accordingly.
Image:
Education
GenAI offers many opportunities in the area of education. It can personalize educational content to match each student’s learning preferences, pace, abilities and learning styles, offering real-time feedback and guidance. This approach enhances the effectiveness of learning and addresses the diverse needs of students. In addition, it can provide high-quality learning resources to remote or underserved regions (UNESCO 2023). On the downside, GenAI also poses challenges especially regarding the use of GenAI tools to cheat during exams and writing assignments. Moreover, AI models may amplify existing societal biases, resulting in the generation of culturally insensitive or biased content. Additionally, there is a risk of students becoming overly reliant on AI-generated assessments, potentially hindering the development of critical thinking and problem-solving abilities.
Example patent: CN110853457 – INTERACTIVE MUSIC TEACHING GUIDING METHOD
Applicant: Chinese Academy of Sciences
Abstract: The invention discloses an interactive music teaching guiding method, which provides a music library for a user to select. After the user selects a song version as a reference audio track (if the song version selected by the user belongs to an unpublished edited version, a music score can be uploaded), the tone scale of each note of the reference audio track is calibrated, and a reference audio track waveform diagram is drawn. The human voice of the user is collected, the tone scale of each note is calibrated, the time axis is kept consistent with the reference audio track waveform diagram, and the human voice audio track waveform diagram of the user is drawn in real time.
Image:
Entertainment
Prominent areas of discussion surrounding GenAI are the opportunities and challenges of GenAI use in the entertainment industry. Since GenAI generates the types of content that the entertainment industry relies on (scripts, stories, marketing campaigns etc.) it could be used to increase creativity and productivity as well as lower production costs. For example, the GenAI company Runway AI and its video editing tools were involved in making the Oscar-winning movie “Everything Everywhere All at Once” (Kingson 2023). However, there are also concerns about the rapid development of GenAI. While it can create content quickly and efficiently, it may also lead to the displacement of human creative workers and the production of unoriginal content. GenAI also poses risks to the industry on complicated legal, ethical and technical fronts. These include unresolved copyright guidance on GenAI training and the protectability of AI-generated output or unauthorized use of studio IP.
Example patent: US20200406144 – SYSTEMS AND METHODS FOR DYNAMICALLY GENERATING AND MODULATING MUSIC BASED ON GAMING EVENTS, PLAYER PROFILES AND/OR PLAYER REACTIONS
Applicant: Activision (Microsoft)
Abstract: The application describes methods and systems for dynamically generating a music clip for rendering at client devices in a multi-player gaming network. Player data and event data are acquired and classified into two or more profiles. The music clip is then generated by identifying a mood based on one of the two or more event profiles and one of the two or more player profiles and modulating one or more music elements of a segment of audio data based on the identified mood.
Image:
Arts and humanities
Similar to its impact on the entertainment industry, GenAI also offers opportunities and challenges in the area of arts and humanities. Text-to-image GenAI tools such as DALL-E or Midjourney can generate paintings in seconds. For example, artist Jason M. Allen won the 2022 Colorado State Fair’s annual fine art competition with a painting that was generated with Midjourney (New York Times 2022). However, this led to a backlash from artists who accused the artist of cheating. In general, GenAI has the potential to make art more accessible to a wider audience, foster new forms of artistic expression and augment human creativity. On the other hand, GenAI-art has been met with criticism for its potential to replicate existing artworks without significant creative input, raising concerns about the originality and authenticity of AI-created pieces. This has sparked anxieties about the impact of AI on the art world, with some fearing that AI could diminish the value of human creativity.
Example patent: US20230316590 – GENERATING DIGITAL PAINTINGS UTILIZING AN INTELLIGENT PAINTING PIPELINE FOR IMPROVED BRUSHSTROKE SEQUENCES
Applicant: Adobe
Abstract: The present disclosure relates to systems, methods, and non-transitory computer readable media for generating painted digital images utilizing an intelligent painting process that includes progressive layering, sequential brushstroke guidance, and/or brushstroke regularization. For example, the disclosed systems utilize an image painting model to perform progressive layering to generate and apply digital brushstrokes in a progressive fashion for different layers associated with a background canvas and foreground objects. In addition, the disclosed systems utilize sequential brushstroke guidance to generate painted foreground objects by sequentially shifting through attention windows for regions of interest in a target digital image. Furthermore, the disclosed systems utilize brushstroke regularization to generate and apply an efficient brushstroke sequence to generate a painted digital image.
Image:
Computing in government
Gen AI’s ability to access, organize and leverage data will create new possibilities for improving government offerings. For example, customer services could get a boost from GenAI-powered chatbots that answer questions from or customize services for residents. Alternatively, when working on citizens’ service requests, GenAI can also assist government employees by unlocking data across agencies to provide information and services more intuitively. Another area that can benefit is government procurement that traditionally is a complex and time-consuming process, involving multiple steps, stakeholders and complex legal and regulatory constraints. GenAI could help to simplify and automate procurement processes by providing intelligent recommendations and facilitating negotiations.
Example patent: CN115526440 – RISK MANAGEMENT ASSESSMENT METHOD BASED ON CROWD SIMULATION
Applicants: Hitachi, Tsinghua University
Abstract: The invention provides a risk management and control assessment method and system based on crowd simulation. The method comprises the following steps: acquiring global population distribution information of a target area; according to the global population distribution information of the target area, generating a motion trail of pedestrians in the target area; performing crowd risk assessment on the target area according to the motion trails of the pedestrians. According to the method, risk assessment is carried out based on the individual simulation trajectory and the individual information, risk simulation based on individual granularity is realized, heterogeneity of each individual is highlighted, and a real risk condition can be reflected more accurately.
Networks/smart cities
GenAI also has various use cases in helping smart cities to address challenges with traffic, transportation and infrastructure. GenAI can analyze data from sensors, cameras, etc. to optimize the management of infrastructure, such as traffic lights, energy grids and water systems. For instance, it can help to optimize traffic flow by dynamically adjusting signal timing or suggesting alternative transportation options.
Example patent: CN115393494 – URBAN MODEL RENDERING METHOD AND DEVICE BASED ON ARTIFICIAL INTELLIGENCE, EQUIPMENT AND MEDIUM
Applicant: Baidu
Abstract: The invention provides an urban model rendering method and device based on artificial intelligence, equipment and a medium, relates to the technical field of artificial intelligence, in particular to image processing, digital twinning and virtual reality technology, and can be applied to smart cities, urban governance and public security emergency scenes. According to the specific implementation scheme, the method comprises the steps of obtaining a second precision model of a first precision model, and rendering a second precision model on the upper layer; sending a rendering request to a rendering server, obtaining a rendering image of a first precision model sent by the rendering server in response to the rendering request, and displaying the rendered image on the lower layer to render the first precision model, wherein the rendered first precision model is overlapped with the rendered second precision model; The second precision model is a result of removing the model map and the material information by the first precision model.
Industrial property, law, social and behavioral sciences
GenAI also has use cases in the fields of industrial property, law, social and behavioral sciences. For example, GenAI can be used to assist in the creation and analysis of patent designs. It can identify potential design flaws and assess the originality and inventiveness of patent applications. This can reduce the time and cost involved in securing intellectual property protection. GenAI can also be used to assist lawyers in conducting legal research and analyzing case law. In addition, it can help to review and draft contracts. Within social and behavioral sciences, it can be used to collect and analyze large amounts of data from social media, surveys, etc. and to identify trends and relationships in the data that would be difficult to detect using traditional methods.
Example patent: CN110895568 – METHOD AND SYSTEM FOR PROCESSING COURT TRIAL RECORDS
Applicant: Alibaba
Abstract: The invention discloses a method and a system for processing court trial records. The method comprises the steps of acquiring court trial records recorded in the court trial process, the court trial records comprise at least one theme module, and the theme modules at least record legal information generated in different court trial stages in the court trial process; determining an information extraction model corresponding to each theme module based on the identification information of different theme modules and corresponding preset parameters; and based on the information extraction model, extracting information elements of each theme module from the court trial record, and the information elements are used for initializing the legal knowledge graph. The technical problem that the record abstract of the online internet court is difficult to obtain in the prior art is solved.
Cartography
GenAI will also lead to changes in the field of cartography, the science and art of mapmaking. GenAI models possess the ability to produce entirely novel data, including maps, images and text, directly from existing datasets. This significantly expedites the analysis of geospatial data, unveiling concealed insights that were previously inaccessible. For example, GenAI can help to analyze satellite imagery and lidar data to reconstruct 3D models of cities, providing a more realistic and immersive view of urban environments.
Example patent title: US20230046926 – 3D BUILDING GENERATION USING TOPOLOGY
Applicant: Here Global BV
Abstract: Embodiments provide systems and methods for three-dimensional building generation from machine learning and topological models. The method uses topology models that are converted into vertices and edges. A BGAN (Building generative adversarial network) is used to create fake vertices/edges. The BGAN is then used to generate random samples from seen sample of different structures of building based on relationship of vertices and edges. The embeddings are then fed into a machine trained network to create a digital structure from the image.
Image:
Military
GenAI is also being explored as a tool in the military for decision support, intelligence analysis and offensive as well as defensive autonomous systems. Autonomous systems have been postulated to have high value for military operations. The benefits range from performing tasks faster than humans/human-operated system for time-critical missions (e.g. air defense or cyber operations) to performing well in difficult and dangerous missions where human performance tends to deteriorate over time. Moreover, synthetic data play an important part in military applications, as they allow for the generation of diverse datasets that have a beneficial effect on training AI systems. Synthetic data could also eliminate legal challenges related to collecting, storing and disposing of sensitive data, thus potentially allowing for more sharing of data among allies (Deng 2023).
Example patent: KR1020230141170 – MINE DETECTION METHOD USING GENERATIVE ADVERSARIAL NEURAL NETWORK
Applicant: Republic of Korea Army
Abstract: According to an embodiment of the present invention, the mine detection method using a generative adversarial network comprises the steps of: receiving signal image information of the underground facility; performing machine learning by using generative adversarial networks and generating information of a mine by using the machine-learned neural network on the basis of the input data of the mine.
Energy management
The energy sector can leverage GenAI to achieve a diverse range of objectives, including optimizing energy consumption patterns and anticipating demand and supply fluctuations. By analyzing vast amounts of data, including performance indicators, load distribution and other metrics, GenAI algorithms can reveal patterns and insights that empower companies to enhance grid efficiency and make informed decisions. It can also help technicians act on HVAC maintenance needs with rapid access to troubleshooting guides and standard operating procedures.
Example patent: US11152785 – POWER GRID ASSETS PREDICTION USING GENERATIVE ADVERSARIAL NETWORKS
Applicant: X Dev (Alphabet/Google)
Abstract: Methods, systems, and apparatus, including computer programs encoded on computer storage media, for using a neural network to predict locations of feeders in an electrical power grid. One of the methods includes training a generative adversarial network comprising a generator and a discriminator; and generating, by the generator, from input images, output images with feeder metadata that represents predicted locations of feeder assets, including receiving by the generator a first input image and generating by the generator a corresponding first output image with first feeder data that identifies one or more feeder assets and their respective locations, wherein the one or more feeder assets had not been identified in any input to the generator.
Image:
Agriculture
GenAI can be used in agriculture for different purposes. For example, GenAI can be used to process data from various sources, such as satellite imagery and IoT sensors, to create detailed maps of fields. These maps can guide farmers in implementing precision agriculture techniques, optimizing resource use and improving overall efficiency. AI-powered image recognition systems can also be trained to identify signs of pests and diseases in crops. In addition, GenAI models can simulate the effects of different genetic combinations, aiding in the development of new crop varieties with desirable traits.
Example patent: US20230108422 – METHODS AND SYSTEMS FOR USE IN PROCESSING IMAGES RELATED TO CROPS
Applicant: Monsanto (Bayer)
Abstract: Systems and methods are provided for use in processing image data associated with crop-bearing fields. One example computer-implemented method includes accessing a first data set including images associated with a field, where the images have a spatial resolution of about one pixel per at least about one meter, and generating, based on a generative model, defined resolution images of the field from the first data set. In doing so, the defined resolution images each have a spatial resolution of about X centimeters per pixel, where X is less than about 5 centimeters. The method also includes deriving index values for the field, based on the defined resolution images of the field, and predicting a characteristic (e.g., a yield, etc.) for the field based on the index values and, in some implementations, at least one environmental metric for the field.
Image:
