The anatomy of Green AI technologies: structure, evolution, and impact - Dataset and Replicability Material

Accompanying material for the paper "The anatomy of Green AI technologies: structure, evolution, and impact" (2025). Dataset Construction The Green AI Patent Dataset comprises 63 326 unique U.S. patents that intersect environmental (“green”) technologies with artificial‐intelligence components, spanning from 1976 to 2023. It was assembled by combining: PatentsView (USPTO) – U.S. patents (snapshot of January 2025) labelled under Cooperative Patent Classification classes Y02 and Y04S for climate‐change mitigation/adaptation and smart‐grid technologies. Artificial Intelligence Patent Dataset (AIPD 2023 - most recent update) – USPTO’s machine‐learning–validated classification of AI‐related patents (predict50_any_ai = 1). Available here: Pairolero, N. et al. The artificial intelligence patent dataset (aipd) 2023 update. USPTO Economic Working Paper 2024-4,USPTO (2024). Available at https://www.uspto.gov/sites/default/files/documents/oce-aipd-2023.pdf. Variables Variable Description Completeness (non-null count) patent_id Unique USPTO patent identifier. 63 326 cpc_subclass Subclasses of "green" CPC taxonomy Y02 / Y04S. Refer to the USTPO's website for more details: https://www.uspto.gov/web/patents/classification/cpc/html/cpc-Y.html 63 326 patent_date Grant date of the patent (YYYY-MM-DD). 63 326 patent_title Title of the patent. 63 326 assignee Disambiguated assignee organization name. 59 479 country Disambiguated assignee country. 59 155 forward_citations Number of times this patent is cited by later patents (forward citations). 63 326 tech_domain BERTOPIC‐derived technology domain (integer 0–15; –1 marks outliers). 62 337 real_value Market‐value proxy associated with the patent, derived from the updated dataset of Kogan, L., Papanikolaou, D., Seru, A. & Stoffman, N. Technological innovation, resource allocation, and growth. The Q. J.Econ. 132, 665–712, DOI: 10.1093/qje/qjw040 (2017). 26 306 BERTOPIC Topic Mapping Each patent was assigned to one of 16 topics (tech_domain), numbered 0–15 (with –1 for outliers). Below is the label, example keywords (with their topic cohesion scores), and the number of patents in each topic: ID Label Top Keywords (score) Count 0 Data Processing & Memory Management processing (0.516), computing (0.461), process (0.449), systems (0.443), memory (0.421) 27 435 1 Microgrid & Distributed Energy Systems microgrid (0.487), electricity (0.421), utility (0.401), power (0.380), energy (0.370) 5 378 2 Vehicle Control & Autonomous Powertrains vehicle (0.477), vehicles (0.468), control (0.416), driving (0.387), engine (0.386) 3 747 3 Irrigation & Agricultural Water Mgmt irrigation (0.511), systems (0.431), flow (0.353), process (0.348), water (0.333) 2 754 4 Photovoltaic & Electrochemical Devices semiconductor (0.518), photoelectric (0.509), electrodes (0.487), electrode (0.473), photovoltaic (0.470) 2 599 5 Clinical Microbiome & Therapeutics microbiome (0.481), clinical (0.371), physiological (0.321), therapeutic (0.320), disease (0.314) 2 286 6 Combustion Engine Control combustion (0.423), engine (0.373), control (0.342), fuel (0.338), ignition (0.318) 2 179 7 Battery Charging & Management charging (0.485), charger (0.449), charge (0.425), battery (0.386), batteries (0.377) 1 541 8 HVAC & Thermal Regulation hvac (0.515), heater (0.474), cooling (0.471), heating (0.464), evaporator (0.455) 1 523 9 Lighting & Illumination Systems lighting (0.621), illumination (0.601), lights (0.545), brightness (0.526), light (0.488) 1 219 10 Exhaust & Emission Treatment exhaust (0.464), catalytic (0.446), purification (0.444), catalyst (0.366), emissions (0.365) 1 064 11 Wind Turbine & Rotor Control turbines (0.498), turbine (0.488), windmill (0.464), wind (0.418), rotor (0.300) 988 12 Aircraft Wing Aerodynamics & Control wing (0.450), aircraft (0.448), wingtip (0.424), apparatus (0.423), aerodynamic (0.418) 697 13 Meteorological Radar & Weather Forecasting radar (0.541), meteorological (0.511), weather (0.412), precipitation (0.391), systems (0.372) 542 14 Fuel Cell Systems & Electrodes fuel (0.375), cell (0.313), systems (0.295), cells (0.291), controls (0.262) 377 15 Turbine Airfoils & Cooling airfoils (0.584), airfoil (0.572), turbine (0.433), engine (0.333), axial (0.321) 352 –1 Outliers – 7 656 Code availability This Zenodo entry contains topic_modeling.ipynb, a fully documented jupyter notebook containing Python code for uncovering latent themes in patent abstracts using BERTopic. It walks through text preprocessing (lowercasing, standard English stopwords plus “herein” and “invention,” tokenization, and boilerplate removal), embedding with the all-MiniLM-L6-v2 SentenceTransformer, dimensionality reduction via UMAP, clustering with HDBSCAN, and topic extraction through class-based TF-IDF. The script also executes a grid search over UMAP and HDBSCAN hyperparameters, computes UMass coherence and topic diversity for each configuration, and saves a CSV of evaluation metrics, enabling straightforward reproduction of our topic-modeling workflow. **Note on Patent Abstracts** The BERTopic analysis in this notebook was performed on the full text of U.S. patent abstracts. To save space and comply with memory constraints, the abstracts themselves are not included in this repository. However, they can be downloaded directly from the PatentsView portal (see “g_patent_abstract” in the data tables at https://patentsview.org/download/data-download-tables). Each record is linked to our processed dataset via the `patent_id` field, so you can seamlessly merge the raw abstracts with your local copy of the Green AI dataset before running or inspecting the topic model. Additional analyses, such as data cleaning, merging, aggregation, and the generation of summary tables and plots, were also performed but are not included here by default, as they consist of straightforward operations using standard open-source libraries (e.g., pandas, NumPy, matplotlib, and seaborn). The full code for these steps can be made available upon request.本配套材料对应论文《绿色人工智能技术剖析：架构、演进与影响》（2025）。 ## 数据集构建 本绿色人工智能专利数据集（Green AI Patent Dataset）包含63 326条唯一美国专利，涵盖1976年至2023年间同时涉及环境（"绿色"）技术与人工智能组件的专利。本数据集通过整合以下两部分资源构建而成： 1. PatentsView（美国专利商标局USPTO）——2025年1月快照的美国专利，归类于合作专利分类（Cooperative Patent Classification, CPC）体系下的Y02与Y04S类，对应气候变化减缓/适应技术及智能电网技术。 2. 人工智能专利数据集（Artificial Intelligence Patent Dataset, AIPD 2023，最新更新版）——美国专利商标局经机器学习验证的AI相关专利分类（predict50_any_ai = 1）。该数据集可参考：Pairolero, N. 等人. 《人工智能专利数据集（AIPD）2023更新版》. 美国专利商标局经济工作论文2024-4, 美国专利商标局 (2024). 可在https://www.uspto.gov/sites/default/files/documents/oce-aipd-2023.pdf 获取。 ## 变量说明 | 变量名 | 描述 | 完整度（非空计数） | |--------|------|-------------------| | patent_id | 美国专利商标局唯一专利标识符 | 63 326 | | cpc_subclass | "绿色"CPC分类体系Y02/Y04S的子类。详细信息可参考美国专利商标局官网：https://www.uspto.gov/web/patents/classification/cpc/html/cpc-Y.html | 63 326 | | patent_date | 专利授权日期（格式为YYYY-MM-DD） | 63 326 | | patent_title | 专利标题 | 63 326 | | assignee | 经消歧的专利权人机构名称 | 59 479 | | country | 经消歧的专利权人所属国家 | 59 155 | | forward_citations | 该专利被后续专利引用的次数（正向引用数） | 63 326 | | tech_domain | 基于BERTopic模型推导得到的技术领域（整数0–15；–1标记为异常值） | 62 337 | | real_value | 该专利对应的市场价值代理指标，源自Kogan, L.、Papanikolaou, D.、Seru, A. 与 Stoffman, N. 的更新数据集：*Technological innovation, resource allocation, and growth*. *The Quarterly Journal of Economics* 132, 665–712, DOI: 10.1093/qje/qjw040 (2017) | 26 306 | ## BERTopic主题映射 每篇专利被分配至16个主题之一（tech_domain），编号0–15（异常值标记为–1）。下表列出主题编号、标签、核心关键词（附带主题凝聚度得分）及各主题的专利数量： | 主题ID | 主题标签 | 核心关键词（得分） | 专利数量 | |--------|----------|-------------------|----------| | 0 | 数据处理与内存管理 | 处理（0.516）、计算（0.461）、流程（0.449）、系统（0.443）、内存（0.421） | 27 435 | | 1 | 微电网与分布式能源系统 | 微电网（0.487）、电力（0.421）、公用事业（0.401）、功率（0.380）、能源（0.370） | 5 378 | | 2 | 车辆控制与自主动力总成 | 车辆（0.477）、多车辆（0.468）、控制（0.416）、驾驶（0.387）、发动机（0.386） | 3 747 | | 3 | 灌溉与农业用水管理 | 灌溉（0.511）、系统（0.431）、流量（0.353）、流程（0.348）、水（0.333） | 2 754 | | 4 | 光伏与电化学器件 | 半导体（0.518）、光电（0.509）、多电极（0.487）、单电极（0.473）、光伏（0.470） | 2 599 | | 5 | 临床微生物组与治疗学 | 微生物组（0.481）、临床（0.371）、生理（0.321）、治疗（0.320）、疾病（0.314） | 2 286 | | 6 | 燃烧发动机控制 | 燃烧（0.423）、发动机（0.373）、控制（0.342）、燃料（0.338）、点火（0.318） | 2 179 | | 7 | 电池充电与管理 | 充电（0.485）、充电器（0.449）、电荷（0.425）、单电池（0.386）、多电池（0.377） | 1 541 | | 8 | 暖通空调（HVAC）与热调节 | HVAC（0.515）、加热器（0.474）、冷却（0.471）、加热（0.464）、蒸发器（0.455） | 1 523 | | 9 | 照明与照明系统 | 照明（0.621）、照明（0.601）、多灯（0.545）、亮度（0.526）、单灯（0.488） | 1 219 | | 10 | 排气与排放处理 | 排气（0.464）、催化（0.446）、净化（0.444）、催化剂（0.366）、多排放（0.365） | 1 064 | | 11 | 风力涡轮机与转子控制 | 多涡轮机（0.498）、单涡轮机（0.488）、风车（0.464）、风（0.418）、转子（0.300） | 988 | | 12 | 飞机机翼空气动力学与控制 | 机翼（0.450）、飞机（0.448）、翼尖（0.424）、装置（0.423）、空气动力学（0.418） | 697 | | 13 | 气象雷达与天气预报 | 雷达（0.541）、气象（0.511）、天气（0.412）、降水（0.391）、系统（0.372） | 542 | | 14 | 燃料电池系统与电极 | 燃料（0.375）、单电池（0.313）、系统（0.295）、多电池（0.291）、多控制（0.262） | 377 | | 15 | 涡轮机叶片与冷却 | 多翼型（0.584）、单翼型（0.572）、涡轮机（0.433）、发动机（0.333）、轴向（0.321） | 352 | | –1 | 异常值 | 无 | 7 656 | ## 代码获取说明 本Zenodo学术存储条目包含`topic_modeling.ipynb`——一份带有完整文档的Jupyter Notebook（Jupyter笔记本），内含用于从专利摘要中挖掘潜在主题的Python代码。该笔记本文档涵盖了文本预处理（小写转换、标准英文停用词加"herein"与"invention"、分词、模板文本移除）、使用all-MiniLM-L6-v2 SentenceTransformer进行嵌入、通过UMAP进行降维、使用HDBSCAN进行聚类，以及基于类别的词频-逆文档频率（TF-IDF）进行主题提取。该脚本还会对UMAP与HDBSCAN的超参数进行网格搜索，为每种配置计算UMass凝聚度与主题多样性，并保存评估指标的CSV文件，从而可便捷复现我们的主题建模流程。 ## 专利摘要说明 本Notebook中的BERTopic分析基于美国专利摘要的全文内容。为节省空间并适配内存限制，本仓库未包含原始摘要文件。但可直接从PatentsView门户下载（参考https://patentsview.org/download/data-download-tables的数据表中`g_patent_abstract`字段）。每条记录均通过`patent_id`字段与本处理后的数据集关联，因此可在运行或检视主题模型前，将原始摘要与本地副本的绿色人工智能数据集无缝合并。 ## 补充说明 额外分析（如数据清洗、合并、聚合，以及汇总表格与图表生成）同样可执行，但默认未包含在本仓库中，因为这些操作均为使用标准开源库（如pandas、NumPy、matplotlib与seaborn）的常规操作。如需获取上述步骤的完整代码，可联系索取。