Carbonate reservoirs in China are deep in buried depth and diverse in types. Among them, karst fracture-cavity reservoirs controlled by paleo-geomorphology, structure and karstification, have large differences in storage space scale, strong heterogeneity, complex spatial distribution pattern, and strong ambiguity of geophysical response characteristics, so they are very difficult to identify, predict, describe and model^[1,2,3]. At present, the reservoir research mainly relies on well data and the seismic data. Since the core and well logging data of single well is limited, and the inter-well analysis is limited by the precision of seismic prediction, it is very hard to accurately characterize carbonate reservoirs of different genetic types^{[4,5,6,7,8,9]}.

With the great improvement of computer computing ability, machine learning algorithms such as multi-point statistical geological modeling method based on training images and deep learning method based on samples have developed rapidly in recent years, and they have expanded in application fields rapidly, and constantly improved in application effect. In terms of clastic reservoir research, domestic and foreign research institutions have significantly reduced the uncertainty of geological model by establishing geological knowledge base of clastic fluvial reservoirs^{[9,10,11,12,13,14]}. As the training images contain various data and a variety of geological prior knowledge, such as sedimentary genesis, sedimentary scale, spatial morphology and distribution law, the knowledge such as statistical analysis of reservoir models in individual wells can be obtained^{[15,16,17,18,19,20]}. The knowledge base can effectively improve the quality of training images, and can be successfully used in multi-point statistical geological modeling.

In carbonate reservoir research, with no effective data and knowledge management means, the knowledge on carbonate reservoirs is fragmented and scattered in databases or professional systems, personal computers and lengthy research reports, which makes it difficult to obtain the knowledge completely. The amount of data and knowledge that can be obtained in reservoir modeling is small and unilateral, and it is difficult to work out statistical laws. At the same time, the definition and framework design of carbonate reservoir knowledge system is different from those of clastic reservoir system. Since it needs to acquire the structure information in caves, the related data acquisition methods and processes are different. The conventional methods for building clastic reservoir knowledge base cannot meet the requirements of building carbonate reservoir knowledge base.

We propose to build a carbonate reservoir knowledge base in line with the characteristics and characterization difficulties of this kind of reservoir, to serve high-quality reservoir research. By constructing targeted knowledge classification system and knowledge description standards, various technical methods can be combed and integrated, and technical service standards and technical service standard package can be established according to the specifications. At the same time, case samples are collected and accumulated continuously, relevant case samples are strictly and effectively managed to work out construction modes, parameter characteristics and quantitative relationship of various samples^[21,22]. When the knowledge base has enough samples and rich knowledge content, analogy and statistical analysis can be carried out based on the knowledge base to get statistical laws. At the same time, more geological information can be integrated into the model, and training images can be established and continuously accumulated, to obtain training images closer with the target areas, so as to improve the understanding accuracy of reservoir and solve the problems in traditional carbonate reservoir modeling^{[23,24,25,26,27]}. With the support of complete, systematic, real-time updated and optimized knowledge system, the reservoir knowledge base can effectively reduce the uncertainty of geological model, improve the refinement degree of the modeling, and more effectively support the efficient development of oil and gas reservoirs.

The construction and application of the carbonate reservoir knowledge base include four specific contents and steps: construction of classification framework system, description criterion and standards, and knowledge acquisition and extraction of reservoir knowledge base, and analogy analysis based on knowledge base and reservoir modeling.

The carbonate reservoir knowledge base construction needs to consider the application needs, goals, main technical methods, knowledge contents, and how to obtain and classify knowledge to facilitate its application.

After analyzing "what technical process and standards (technical standards) should be obeyed? what technical methods (technical research methods) should be applied? what basic data, sedimentary model, quantitative parameters and other related knowledge contents (geological object cases) are needed? ” in carbonate reservoir research, we came up with the knowledge framework of the reservoir thematic knowledge base with three categories of knowledge: technical service standards, technical research methods, and professional knowledge and cases of geological objects. Based on the reservoir types and research technology fields, these three categories are further subdivided into sub-categories, each sub-category has its own knowledge system, and each sub-category is further subdivided into child categories, and so on, from larger to smaller, from coarser to finer, forming a complete carbonate reservoir knowledge classification system as shown in Fig. 1.

Taking geological modeling knowledge structure as an example, the main contents include: summarizing various types of geological modeling methods; establishing modeling process according to the characteristics and difficulties in modeling of different types of oil and gas reservoirs; optimizing typical examples of geological modeling of different types of oil and gas reservoirs, supporting relevant data and geological models; summarizing experience from method optimization, software skills, operation guide, etc., and establishing experience base for users to learn. The most important thing is to form a technical service standard package based on industry standards and enterprise requirements to ensure the quality of modeling.

The knowledge of technical service standards is reflected by the development of enterprise technical service standard packages. Professional technical service standard package is a series of standardized technical specifications and processes, which have been widely discussed, prepared and approved by experts. The purpose is to solidify the professional ability, and solidify the personal ability of experts into organizational ability. The degree of standardization and streamlining of business directly reflects the maturity and growth of business. The technical service standard package is the core knowledge content of technical service standard knowledge base.

Technical service standard packages are classified according to businesses, and basic technical service standard packages can be prepared according to the exploration and development businesses, including the business guide, standard and specification, checklist of inspection elements, operation guide, typical cases, experience and lesson summarization. The key to the successful application of the technical service standard package is to establish a mechanism to ensure the continuous updating of the contents of the standard package, especially the typical cases, and experience and lesson summarization.

Compiling technical standard package is one of the main contents of building, enriching and optimizing knowledge base. For example, in view of the special needs of carbonate reservoir geological modeling, the knowledge base has necessary expansion on the contents of the original technical standard package, and which is included in the basic service standard package after approved by experts, forming an authoritative standard system for the reference of upstream enterprises. The main businesses the technical service standard package of knowledge base covers include eight technical standard packages: reservoir evaluation, reservoir experimental research, geological modeling, logging evaluation, reservoir description, 3D seismic data processing, and 3D seismic data interpretation. For example, the technical service standard package for fracture-cavity carbonate reservoir modeling includes: business guide for fracture-cavity carbonate reservoir geological modeling, geological modeling standards and specifications, checklist of inspection elements for reservoir geological modeling, operation guide for reservoir modeling, typical cases of reservoir modeling, and experience and lesson summarization of geological modeling, etc.

Knowledge is classified according to different professional and technical categories, and its contents are roughly equivalent to "the latest and most complete professional textbooks". The technical categories needed for carbonate reservoir modeling include geological analysis and evaluation technology, logging evaluation technology, geophysical processing and interpretation technology and modeling technology, etc. This is conducive to the acquisition of knowledge by categories, but also allows users to focus on the knowledge in their specialties more quickly and solve the practical problems encountered in reservoir research. Among them, the knowledge classification framework of reservoir geological modeling technologies includes: reservoir geological modeling method, geological modeling process for different types of reservoirs, and so on.

The knowledge of technical research methods in the reservoir knowledge base is an effective expansion and enhancement of the knowledge in the technical service standard packages. It acquires not only a large number of basic and professional knowledge, but also many of the latest technical methods for various complex practical problems. It can help some researchers with basic professional knowledge not systematic enough master the relevant knowledge and tools quickly, complete the tasks they undertake, and also provide effective reference for experienced professionals. The technical research method knowledge has higher degree of freedom and richness than standard knowledge.

When studying a geological object with some technical methods, some regularity knowledge (such as genetic mechanism, distribution law, geological mode, etc.) can be obtained. Recorded and classified, the regularity knowledge form cases, which can provide examples and samples for further researches. For example, all kinds of geological modes based on analysis of outcrops and reservoir cases can provide training images for geological modeling. As the geological object case knowledge is directly oriented to application, its richness and completeness become the key to the success of building knowledge base.

As the carbonate reservoir related knowledge is huge in content and complex in structure, this study selected the fracture-cavity carbonate reservoir as the main database object. Based on the structure and genesis, karst fracture-cavity reservoirs can be divided into three types: weathering crust, underground river and fault-controlled karst. Fig. 2 is the knowledge classification system of fracture-cavity carbonate reservoirs. Through extensive acquisition of field outcrops and reservoir development cases, and on the basis of knowledge classification and description specification, typical cases of reservoir units, modern karst outcrops and paleo-karst outcrops have been constructed. Fig. 3 is the knowledge framework of paleo-underground river outcrop of fracture-cavity carbonate reservoir, which mainly includes three parts: the original data of outcrops and reservoirs, knowledge extraction and knowledge application.

For the description of geological object case (outcrop and reservoir) knowledge, it is necessary to establish the specification of knowledge description, that is, to set up the construction standard of the knowledge base. Since only after standardization can analogy be realized, can statistical data be meaningful, and can the knowledge base provide references for modeling. Some expected statistical laws can be obtained through efficient acquisition, standardization and effective management of outcrops and reservoir research data according to standards.

Taking the outcrop description specification as an example, the knowledge name, knowledge content, display type, generation mode and storage mode, and the format and content of standardized text, diagram and data table should be defined respectively.

Geological modeling needs to apply the type, shape and structure of the karst outcrop directly to the preparation of training images for modeling, so it requires higher precision of description. Through the subdivision of outcrop reservoir types, this subject knowledge base describes its attribute parameters in detail. There are 49 items of contents and attribute parameters of carbonate underground river reservoir in Table 1, and there are more than 450 data fields in the statistical sheet of various characteristics, covering the main characteristics of the underground river. At the same time, the storage standard of all kinds of data and the specification list of 3D outcrop data storage in the knowledge base are specified, including serial number, data item name, pinyin code, type, decimal number, unit of measurement and filling regulations. The statistical contents of cave characteristics are designed, including cave number, section, outcrop area, cave type, cave shape, minimum length, maximum length, average length, minimum width, maximum width, average width, minimum height, maximum height, average height, minimum direction, maximum direction, average direction, width to height ratio, length to width ratio, filling type, filling distribution, filling combination, filling period, cave formation period, and so on.

After the construction of reservoir knowledge base framework, it is necessary to construct knowledge base contents. Through systematic and detailed acquisition, extraction, sorting and storage, the knowledge base contents can be constantly enriched. The systematicity, advancement and the novelty of knowledge contents are the key indicators to judge whether the knowledge base is successful.

In the actual research and production process, due to the lack of effective management of the relevant research data, the result data is scattered, not convenient for sharing and application of data and knowledge. Therefore, how to use the knowledge base platform to classify the scattered data and achievements and realize unified and effective management is the first issue to be solved by the knowledge base.

Since the reservoir knowledge base needs to acquire a large number of characteristic parameters, the data need to be investigated, acquired and sorted out is of huge amount, and the workload is huge, which brings great difficulties to the construction of the knowledge base. Three aspects of information technologies are adopted to improve the efficiency of knowledge acquisition, collation and extraction.

2.2.1. Improving knowledge contents based on multidisciplinary collaborative editing framework

General knowledge bases are usually maintained by a few administrators through the back-end. Ordinary users can only read the contents, but have no authority to add or change knowledge contents. Due to the limitations of personal knowledge, it is difficult to fully understand all the geological attributes of a geological object, resulting in the knowledge base contents not complete enough, and even missing of many key attribute parameters.

Using multidisciplinary collaborative editing framework in the building of reservoir knowledge base can make up for the above defects. The multidisciplinary collaborative editing framework is an open hypertext system that can be created by many users on the Internet. In the front-end interface, anyone can expand and edit the contents, and the contents can be stored after approved by the administrator. By using this way of collaborative knowledge construction, researchers of different fields can work together to complete the construction of thematic knowledge contents. The knowledge contents will be improved over time, the understandings on specific objects will be deepened and become objective gradually, and the contents of the knowledge base will be enriched and improved gradually.

2.2.2. Automatic extraction of geological objects and attribute parameters based on Natural Language Processing Technology

When writing cases, knowledge administrators need to search and read a wide range of materials. Through acquisition, collation, induction and summary of the existing data, the relevant contents are extracted and incorporated into the knowledge base, which is time-consuming and labor-consuming. With the development of deep learning and NLP (Natural Language Processing Technology), it can automatically process large scale documents, and identify and automatically classify the knowledge in the documents through deep learning today. As shown in Fig. 4, the knowledge graph based on natural language processing is automatically generated, which can greatly reduce the time of knowledge acquisition and sorting, and greatly improve the work efficiency of knowledge administrators.

2.2.3. Data acquisition based on modern fine measurement technology

The number and quality of cases in the outcrop knowledge subject database of the carbonate reservoir knowledge base are very important. Therefore, it is necessary to carry out field investigation on each outcrop, acquire geological information, and complete the sorting and storing work. As shown in Fig. 5, the outcrop data acquisition is mainly completed in the field exploration process by using the latest measurement technologies such as laser radar and UAV tilt photogrammetry. In caves, laser radar and close-range photogrammetry technology are mainly used to build the digital model of internal structure, and UAV tilt photogrammetry technology is used to obtain the outcrop image data and build the digital outcrop model. Through the research and development of integrated modeling technology based on images and laser radar, the integration of digital models inside and outside the cave is realized.

The software framework of carbonate reservoir knowledge base platform is designed with layered and loosely coupled framework system, and is composed of one portal and three subsystems that can be integrated and run independently. Each subsystem can run independently and has deep interactive data integration with the other subsystems.

The whole platform system is complete in framework and function, and includes 8 main modules and 967 functional submodules, as follows. (1) Deep carbonate rock portal system: its function is to integrate and display the contents of all linked subsystems, and it includes integrated search, subsystem entrance and user authority integration modules. (2) Deep reservoir knowledge library subsystem: it is similar to Baidu Library, including modules such as uploading, browsing, querying and content recommendation of literatures. (3) Deep reservoir knowledge card subsystem: similar to Baidu Ency-clopedia (Wiki framework), it is an open web page system for collaborative creation by many users that integrates knowledge graph, GIS (Geographic Information System), tag, knowledge recommendation and other functional modules. (4) Fracture-cavity carbonate professional knowledge subsystem: its function is similar to the professional database built for professional application needs. Fine knowledge description, fine processing and organization need to be conducted according to specific professional application needs, to form a professional knowledge base in a special field to support some professional analysis and application. The whole platform has various data management forms, and can manage a variety of unstructured, semi-structured and structured data.

The main application goal of building a knowledge base is to carry out analogy analysis according to the relevant information of the knowledge base and obtain statistical laws. By obtaining characteristic parameter data of the target area from the knowledge base, carrying out detailed statistics and comparative analysis, some preliminary statistical laws can be obtained, to enrich the contents of the knowledge base and provide effective reference for geological modeling.

If the acquired characteristic parameters (such as cave length, width and height) are not statistically analyzed, the acquired data itself is meaningless. If based on the format specification, statistical analysis and mapping are conducted, or summarization and classification are made to list into a table, geological significance is given, these data can become useful knowledge. Fig. 6 is the parameter statistical classification table of the underground river caves in Tabei area and modern karst channels of Miaoqian town in Hunan. Based on the structural characteristics of modern underground rivers, together with the causes of structural elements of Tabei underground river, the landform and structural location of the underground rivers can be analyzed, the internal structural types of the underground river reservoir can be classified, the cave types of the underground river can be comprehensively identified, and the knowledge template of fracture-cave characteristic parameters can be formed to provide the basis for reservoir description and geological modeling.

As shown in Fig. 7, based on the statistical analysis of the knowledge base, the finding that faults control the development and distribution of fractures and caves has been reached: the larger the fault scale is, the larger the fault displacement is, the higher the probability of fracture development around the fault is, and the wider the fracture zone is; the closer to the fault, the more developed the related fractures and caves are; and the fracture development degree near the fault core is the highest. Based on the statistical analysis, the response relationship between fault distance and development frequency of fractures and caves is obtained. The relevant data, after transformed into conditional probability volume of reservoir development, can be used for fine geological modeling of fault-controlled caves and fractures.

The large amount of geological information in the knowledge base is used to construct training images, and then the multi-point geostatistical modeling method is used to further characterize the structural characteristics of the underground river system on the basis of description of the underground river morphology. Making training images based on the knowledge base is the core content. Taking the geological modeling of underground river karst reservoir of the Ordovician Yijianfang Formation in Tahe area as an example, the preparation method of training images is introduced in the following section.

Firstly, paleo-karst and modern karst outcrops similar to the study area are selected from the knowledge base, the shape and structure of the underground river are counted and examined, the characteristic parameters are measured, including the type of underground river, geometric parameters (length, width, height, length-width ratio, width-height ratio) and internal characteristics (filling, physical properties), etc., the distribution of fillings and collapses in the underground river is analyzed, and the corresponding river channel distribution and structural plan are drawn.

Secondly, through comparison, it is found that the widths (50-100 m) and width probability distribution of modern underground rivers are quite different from the widths of paleo-underground river and caves adjacent to Tahe area, so it is necessary to correct the width data of modern caves. As shown in Fig. 8, the cumulative probability curve method is used to find the width (3.7 m) with the same cumulative frequency on the cumulative frequency curve of the ancient cave width for the width (82 m) of any point of modern caves, and the corresponding width list of the paleo-underground river and the modern underground river before and after correction is established.

Finally, as shown in Fig. 9, the 3D training image is made based on the corrected shapes of modern underground river and caves. The image not only simulates the shape and structure of typical modern underground river, but also reflects the geometric parameters of paleo- underground rivers and caves in the study area. It can be better used in the modeling of paleo-underground rivers and caves, and can more accurately simulate the situation of the target area and reduce the uncertainty of the model.

The typical underground river system in the main area of Tahe Oilfield in the Tarim Basin was used for modeling. The area is about 47.8 km², covering 13 development units (Fig. 10). Compared with the modeling method based on seismic volume carving, the modeling method based on knowledge base is further improved in accuracy. By comparing the reservoir thickness and types of 88.5 m in two sparse wells, the coincidence rate of actual drilling increased from 67.0% to 81.8%; the original model depicted the outline of the river channel, mainly the abnormal body of the river channel, while the new model includes two main channels, six branch channels, five dissolution zones and sinkholes. The application of the modeling method based on the thematic knowledge base more effectively reflects the complex structure of the underground river system, and improves the model accuracy and modeling efficiency.

In line with the features and characterization difficulties of carbonate reservoir, an applied reservoir knowledge base has been constructed to integrate management of relevant knowledge and reservoir research and improve the efficiency and precision of carbonate reservoir research. With the goal of serving high quality analysis, evaluation, description and geological modeling of carbonate reservoir, the knowledge base has a framework divided into three categories: technical research method, technical service standard, and professional knowledge and cases related to geological objects.

By building a knowledge classification system and knowledge description standards, the structure of the knowledge base has been set up; theoretical understandings and various technical methods for different geologic objects have been sorted and technical service standard packages have been worked out according to the technical standard; typical outcrop and reservoir cases have been collected and managed according to standard; through systematic extraction, sorting and saving, professional knowledge about geological objects is accumulated constantly. By using encyclopedia based collaborative editing architecture to complete the knowledge contents jointly, using natural language processing (NLP) technology to automatically extract geological objects and related attribute parameters, and using modern fine measurement technology to do data collection, the efficiency of knowledge acquisition, extraction and sorting has been enhanced significantly.

The geological modeling of underground river systems in complex fracture-cavity reservoirs of Tarim Basin was taken as an example to illustrate the construction of carbonate reservoir knowledge base and its application in geological modeling of fracture-cavity reservoir. Drilling data confirms that the geologic model built increased by 14.8% in coincidence rate, and characterized effectively the complex structure of the underground river systems. Clearly, the knowledge base can improve the model precision and model building efficiency.