Academic literature on the topic 'AI-generated learning analytics'

The rapid advancements in generative AI have led to text-to-video models creating highly realistic content, raising serious concerns about misinformation spread through synthetic videos. As these AI videos become more convincing, they threaten information integrity across social media, news, and digital communications. Using AI-generated videos, bad actors can now create false narratives, manipulate public opinion, and influence critical processes like elections. This technology's democratization means that sophisticated disinformation campaigns are no longer limited to well-resourced actors, creating an urgent need for reliable detection methods and human-machine cooperation to maintain public trust in visual information across our digital transformation landscape. The accessibility of these tools to a broader audience amplifies the potential for widespread misinformation, making robust detection systems crucial for maintaining social media integrity.Through our research into video generation models, we identified that state-of-the-art systems like diffusion transformers operate on patches of noisy latent spaces. We deliberately mirrored this architecture in our classifier design, enabling it to analyze videos using the same fundamental structural unit generation models used to create them. This architectural alignment allows our system to adapt to emerging generation techniques while maintaining detection efficacy.We designed an explainable video classifier using deep learning and neural networks that detect AI-generated content and show evidence for its decisions. The classifier uses three main parts: a convolutional encoder that turns video frames into latent representations, a patch vectorizer that breaks these representations into analyzable chunks, and a transformer that processes these chunks to make the final decision. This human-centered computing design lets us efficiently process videos while maintaining explainability through Integrated Gradients, which reveal which input parts influenced the model's decisions.We use integrated gradients to show which parts of a video led to the model's decision. This method looks at how the model's decision changes as we move from a blank video to the actual video, showing us which pixels matter most for classification. These pixel-level maps provide clear evidence of why the model thinks a video is AI-generated or real, providing transparency critical for building trust in automated content verification systems.We will test our model on the GenVideo dataset, a comprehensive collection of videos labeled as real or AI-generated from diverse sources, including Stable Diffusion, Sora, Kinetic 400, and MSRVTT. This large-scale data analytics evaluation will check how well it classifies videos and explains its decisions, helping determine if the model can work as a practical tool for machine learning-based content verification, considering that wrong AI classifications could harm content creators' reputations.Our work adds to the growing field of explainable AI in content authentication and shows why we need clear evidence when making high-stakes decisions about video content. Future work will look at detecting hybrid videos (real videos with AI elements added) and making our visual explanations more useful for human decision-makers in content verification. The insights gained will inform the development of more sophisticated detection systems capable of addressing evolving challenges in digital content authenticity.

Abstract Getting intelligent insight from large amount of dataset is critical for Energy companies to optimize their operations across various business segments such as drilling, production and completion etc. The paper proposes end-to-end workflow to 1) extract data form rig and production reports and store dataset into databases 2) build a conversational generative AI enabled chatbot which is trained to answer questions related to drilling and production monitoring, queries dataset, frequently performed diagnostic analysis and can generate recommendations to improve operations. The chatbot is integrated with large language models (LLM) and machine learning models (ML) on the cloud and based on questions asked by user it provides answers in conversational settings. Chatbot is hosted in cloud and is integrated with various databases, document repositories and several machine learning model. The machine learning models are built to enable chatbot's capability to answer questions related to drilling and production analytics. Chatbot is integrated with user interface where user can type or ask questions. Using natural language process (NLP) and artificial intelligence (Al), chatbot understands intent of question and if needed asks relevant follow-up questions to provide the answer. Chatbot can also perform statistical analysis, generate SQL queries on datasets and can use those statistics to answer questions. Further if enabled, chatbot can also search information from drilling and production reports and scientific articles. Three case studies are presented. In case study#1, chatbot was integrated with operator's historical PDF drilling reports (Volve dataset), which traditionally are not easy to extract and analyze at scale. Several thousand drilling reports were extracted and stored in database. Various capabilities were added to chatbot such has Cross-documents insights and trend, for example, well progression, operation history, can be generated and displayed on user interface and further analysis can be performed in conversational manner. The dataset created was used to perform comparative analysis identifying wells having significant higher non production time (NPT) due to repair or fishing events. In this manner, chatbot can compare one well's operational statistics with other well and generate various visuals which helps identifying possible ways to improve drilling operations. Similarly, chatbot was also trained to provide answers for production diagnostics such as comparing well's relative performances and root cause identification for poor performing wells. When analyzed on test dataset chatbot was able to identify 20% uplift in production for wells supported on plunger lift. Finally, chatbot was enabled to support NLP based searches. Engineers can ask specific questions such as "provide operational log for well F4 when fishing happened and sort the result by reporting date in ascending order. Show me both SQL query and the resulted table" and chatbot will generate SQL query and resulted table. The work demonstrates that generative AI has great potential to transform the Energy industry.

Abstract Having a robust field development plan (FDP) for mid-size mature oil fields generally poses considerable challenges in the context of the integrational elements of production forecast, operational environment, projects and surface facilities. An integrated FDP combined with data analytics and artificial intelligence (AI) has been introduced and deployed in a heavily compartmentalized offshore field of Turkmenistan. An integrated approach through data-centric analytics and AI has been proposed for an optimal FDP. It consists of four aspects: model integration, time-series forecast (TSF) of production, AI-assisted operation-schedule generation, and evaluation and selection of scenarios. Firstly, model integration is performed as bringing together both multi-discipline raw data from field measurement and their interpretations that change non-linearly. Secondly, model integration aids in the application of AI for production forecast. A unique AI technique was built to allow raw data and interpretation. Illustratively, the model is capable of forecasting decline curves matching the history production. Meanwhile, engineers’ production forecast inheriting from simulation, machine learning or type curves is also constructed by understanding how/why human-driven forecasts differ from the measured decline and incorporating those insights. In addition, AI-assisted scheduler efficiently allocates resources for operational activities, considering the well planning nature, intrinsic operation properties, project planning process, surface facilities and expenditures. Resources are thus utilized for optimal schedules. Finally, evaluation and selection of FDP scenarios take place by considering the multidimensional matrix of factors. Multiple scenarios are generated and scored, reacting to the change of factors. AI-powered optimization is availed to recommend the most efficient tradeoffs between production and carbon generation. The implementation of the integrated FDP approach has been successfully applied for the generation of production profiles and operation schedules, which reduces the time by 80% and increasing accuracy by 55%. Production forecast for existing wells and future wells proved to be reliable. It achieved the production targets with proper allocation of schedules, by considering multi-discipline constraints. Through AI-assisted scheduler, different types of rigs were properly assigned to the planned wells, which requires additional rigs based on the outcome. The model was agile to the change and sensitivities of wells requirement, projects uncertainties and cost changes. The optimum FDP scenario was recommended for the business decision, operation guide and execution. This approach represents a novel and innovative means of integrating and optimizing FDP considering complex factors using AI methods. It is efficient in merging raw data and interpretations for model integration. It accommodates changes and uncertainties from multiple aspects and efficiently generates optimum FDP in a few days rather than months for giant fields. It is the first robust tool that unites subsurface properties, reservoir engineering, production, drilling, projects, engineering and finance for the corporate FDP.

Aim/Purpose: As industry and higher education are moving to incorporate generative AI into courses and training there is a need to understand how varied the responses are to a series of prompts. This study's purpose was to evaluate the consistency of AI responses to prompts for teaching analytics using a single visualization across several sections of the same course over a week. The students will eventually use the data to create a linear regression model and while the instructor will provide all the necessary knowledge the students will have freedom to engineer new features which may be identified using AI. Background: The use of generative AI is essential for most students. But it can be difficult to know if they are seeing comparable results from a series of complicated prompts, especially as we move up to more advanced topics within a graduate program. Generative AI will usually provide very generic answers as you upload visualizations. But as you begin to ask for information more aligned with your goals your answers could vary depending on the level of personalization, and your results will vary by platform and level of service. Understanding how different responses are is essential as you may want to incorporate a series of prompts that act as a basis for how to explore a figure for specific purposes such as evaluating model assumptions or fit. These prompts can also serve as a basis of motivation to learn additional theory or to critically think about the response that AI has provided. It may also be essential to help students to train their AI to provide prompts expected for the problem. Methodology: We collected information from a follow-along assignment that entailed creating a visualization from a publicly available Kaggle Competition that pertained to BMI, smoking, and insurance charges. The students were instructed to use ChatGpt and to upload the visualization which in general should lead to a generic description of the visualization. The students were then asked to prompt regarding a trend that was notable at a BMI of 30 and then ask ChatGPT for features that they should use for a regression model for calculating insurance charges. The responses were submitted and scored by 2 scorers working independently for among other aspects that the original response noted a trend at 30 BMI, if the difference at 30 BMI was noted correctly in the second prompt, and extent of the features provided in the third prompt. Specifically, we evaluated if the AI noted if there should be an interaction between smoking and BMI, whether BMI should be categorized, and if the model should explore an interaction between BMI as a category and the smoking status. After the assignments were independently scored the reviewers then deconflicted any differences. Contribution: This study shows that there were a wide variety of differences in the responses indicating that instructors need to take the time to evaluate how well students can utilize the technology and how well their AI has been personalized to the task if students are to be evaluated with the same available resources. Findings: The study indicated that most of the students only received a basic response to the original prompt, but a small percentage saw the trend identified before the second prompt. AI correctly identified the trend if specifically prompted with 94% of the responses being suggested to explore a basic interaction between BMI and smoking, 43% to explore categorizing BMI, and only 8% advised to explore an interaction between the categories and the smoking status. Recommendations for Practitioners: Ethical use of AI and the value of prompt engineering is necessary in today’s environment. We evaluate how well an assignment given over two weeks would provide consistent results. With the results suggesting a wide variety of responses, educators and trainers need to ensure that students can come to comparable results. This may require more prompting for some students and may also require taking additional time to help some students to train their AI to provide results in the manner needed for the problem. Recommendations for Researchers: More work needs to be done to continue to evaluate different AI platforms and personalization to ensure that educational research done on AI is being performed in a comparable manner. There is a need to make sure that interventions or suggested use of AI is conducted with students having equitable resources. Impact on Society: The use of AI is quickly becoming a requirement for many individuals to perform at work. Evaluating and determining how students with diverse backgrounds can obtain equitable results from the system is important. Future Research: More research needs to be done across platforms. More research is needed to evaluate a balance in generative AI use and the ability to maintain material into memory and critical thinking. Studies should be undertaken to evaluate if student’s perception of theory or learning specific technologies or concepts are changed as they are exposed to diverse ways to prompt AI and the responses generated.

Abstract As the oil and gas pipeline industry shifts toward digitalization, machine learning and artificial intelligence (AI) play an increasingly important role in asset integrity management, including operation monitoring, leak and intrusion detection, corrosion protection, and flow assurance, among others. This paper introduces an integrated approach using fiber optics, inspection reports, and fluid flow simulations and demonstrates how machine learning and AI can help operators by producing unified insights. Fiber-optic distributed acoustic sensing (DAS) technologies are routinely used to monitor pipeline activities; critical events such as product leaks, digging near the pipeline, and pigging are captured by quantitatively analyzing unique signatures on the fiber-optic generated space-time image. This can be treated as a pattern recognition or machine learning problem. YOLO, a state-of-the-art fast object detection algorithm, was used to demonstrate accurate tracking of pipeline inspection gauges (PIGs), among other activities, using a small quantity of training data. In addition, using AI, routine inspection reports and flow simulation results were automatically calibrated, cross-validated, and then contextualized with the fiber-optic DAS generated events. The event detection and classification algorithm used in this work achieves high location accuracy, superior to current industry-standard methods. As a result, this method significantly improves the tracking of PIGs. More importantly, these detections are automatically calibrated with inspection reports for cross-validation. Traditionally, fiber-optic systems are an independent and isolated sensor technology, which require field teams to perform manual activities approximately every 2 km along the entire pipeline for georeferencing. This is inefficient and does not provide the location accuracy needed to link the fiber-optic system to other sources of data, such as inspection reports or flow simulation results. This lack of integration has been a longstanding challenge that prevented operators from easily isolating important signals or repeated trends for each weld, valve, meter, or road crossing, for example. With our machine learning - assisted integrated management system, various sources of data can be consolidated and analyzed to provide valuable information that was previously unavailable. This paper presents the novel use of fast machine learning models to accurately detect and track pipeline activities. In addition, data analytics aids in the calibration and cross-validation of different monitoring technologies under a single integrated pipeline integrity management platform, providing operators with unique insights.

Abstract A practical framework that outlines the critical steps of a successful process that uses data, machine learning (Ml), and artificial intelligence (AI) is presented in this study. A practical case study is included to demonstrate the process. The use of artificial intelligent and machine learning has not only enhanced but also sped up problem-solving approaches in many domains, including the oil and gas industry. Moreover, these technologies are revolutionizing all key aspects of engineering including; framing approaches, techniques, and outcomes. The proposed framework includes key components to ensure integrity, quality, and accuracy of data and governance centered on principles such as responsibility, equitability, and reliability. As a result, the industry documentation shows that technology coupled with process advances can improve productivity by 20%. A clear work-break-down structure (WBS) to create value using an engineering framework has measurable outcomes. The AI and ML technologies enable the use of large amounts of information, combining static & dynamic data, observations, historical events, and behaviors. The Job Task Analysis (JTA) model is a proven framework to manage processes, people, and platforms. JTA is a modern data-focused approach that prioritizes in order: problem framing, analytics framing, data, methodology, model building, deployment, and lifecycle management. The case study exemplifies how the JTA model optimizes an oilfield production plant, similar to a manufacturing facility. A data-driven approach was employed to analyze and evaluate the production fluid impact during facility-planned or un-planned system disruptions. The workflows include data analytics tools such as ML&AI for pattern recognition and clustering for prompt event mitigation and optimization. The paper demonstrates how an integrated framework leads to significant business value. The study integrates surface and subsurface information to characterize and understand the production impact due to planned and unplanned plant events. The findings led to designing a relief system to divert the back pressure during plant shutdown. The study led to cost avoidance of a new plant, saving millions of dollars, environment impact, and safety considerations, in addition to unnecessary operating costs and maintenance. Moreover, tens of millions of dollars value per year by avoiding production loss of plant upsets or shutdown was created. The study cost nothing to perform, about two months of not focused time by a team of five engineers and data scientists. The work provided critical steps in "creating a trusting" model and "explainability’. The methodology was implemented using existing available data and tools; it was the process and engineering knowledge that led to the successful outcome. Having a systematic WBS has become vital in data analytics projects that use AI and ML technologies. An effective governance system creates 25% productivity improvement and 70% capital improvement. Poor requirements can consume 40%+ of development budget. The process, models, and tools should be used on engineering projects where data and physics are present. The proposed framework demonstrates the business impact and value creation generated by integrating models, data, AI, and ML technologies for modeling and optimization. It reflects the collective knowledge and perspectives of diverse professionals from IBM, Lockheed Martin, and Chevron, who joined forces to document a standard framework for achieving success in data analytics/AI projects.

Abstract For decades scientists have leveraged computers to solve the complex chip-formation problem — the hope is that one day the computer will be able to solve every technological problem such as modeling of large plastic deformation occurring under friction conditions. Despite the significant number of modeling and simulation studies, there is still limited knowledge about the real nature of the chip formation process. Until today, better agreement with practical results is still being sought-after. Analyzing this situation, some scientists founded the industry 4.0 project to create new opportunities and benefits for industry. In manufacturing, this initiative requires data integration from CAD to CAM with associated Quality Management (QM), and leveraging advanced analytics such as artificial intelligence (AI) with machine learning (ML) to solve complicated problems. In metal cutting, one such complicated problem is modeling of chip formation. Prior to leveraging AI, or rather to create the premise for a physics-based approach, the phenomena occurring during chip formation need better and more realistic modeling. This paper is presenting advances in physics-based modeling of chip formation with particularization for AlMg5. Adequate process mechanics will be developed resulting in modeling of chip-formation and simulation of chip flow. Moreover, it could be shown that the theoretically developed cutting result is in good agreement with the existing experimental result. There is only one disadvantage during production in the machining cell — the generated chip is very long and it is blocking the production process. Further investigation was necessary. A chip breaker was developed as a function of the existing production spectrum with individual geometry and kinematics. For different cutting conditions and tool geometry considerations for a physics-based machine learning process are proposed.

Abstract The purpose of the installed Rod Pump Controller system was to demonstrate production optimization and OPEX reduction, with the aid of Machine Learning and AI based application that helped Operators and Petroleum Engineers better manage the wells. Another one was to help customer evaluate the "as-a-Service" payment model to operate the wells. Such payment models are becoming more relevant across the industry, where International or National Oil Companies allows technology vendors to manage the entire Electrical and Automation scope for wellhead monitoring and optimization. Several issues can develop with this application that can create costly repair situations. Traditionally these have been monitored by personnel visiting each unit on a prescribed basis to ensure they are still operating and then making adjustments during the visit. This makes it ideal for monitored control to ensure maximum liquid production and reduced operational costs. Production optimization for rod pumps is best accomplished by regulating the speed of the pumpjack as reservoir levels rise and fall, in order to gain maximum production without over pumping the well and damaging the equipment. Better management of motor speed can also provide power savings for the end user. Edge computing is the concept of pushing applications, data, and computing power away from centralized points to the logical extremes of a network. In the oil and gas field, this would be the rod pump well sites. Edge analytics is enabled through a combination of Machine Learning and Control Room/Cloud learning for model development. This approach takes advantage of the unlimited processing power and abundance of historical data available in the control/server room. The model is executed in real-time in the Edge minimizing lag and providing real-time feedback and insights to the operations and productions teams. As part of the project, it was able to demonstrate an end-to-end cyber secure architecture for the rod pump analytics application. The software platform, inclusive of the Edge gateway and the cloud solution, collected all generated dynacards from the RTU and predicted dynacard shape for every stroke. By doing so, the software provided real-time analytics of rod pump performance along with production and energy consumption parameters. Finally, the benefit of the solution are make better production optimization, cost reduction and improve production revenue/uptime, reduce well service down time. Hence all these things will reduce CO2 emission and drive sustainability strategy for customer.