Optimizing throughput of Seq2Seq model training on the IPU platform for AI-accelerated CFD simulations

To address the challenges of limited recovery due to the dominant flow field in polymer flooding development of heterogeneous sandstone reservoirs, as well as the low prediction accuracy and delayed response of traditional numerical simulations and the poor adaptability of conventional machine learning models to complex flow fields, this study proposes a flow field simulation technology system that integrates a dynamic feature attention LSTM (DFA-LSTM) with traditional numerical simulation. First, a dynamic feature attention module is constructed to dynamically assign feature weights based on the temporal rate of change (<tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathrm{S} \_\mathrm{t}$</tex>) of flow field parameters and the spatial correlation factor (<tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{R} \_\mathbf{t}$</tex>). This module is then integrated with an improved LSTM gating mechanism. Furthermore, a machine learning-assisted numerical simulation framework is constructed. The model is trained using 1000 data sets from a sandstone reservoir from 2018 to 2023. After Z-Score normalization and SMOTE sampling preprocessing, the dataset is split into a <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$7: 2: 1$</tex> ratio. The experiment included a control group (traditional numerical simulation), a comparison group 1 (traditional LSTM-assisted simulation), and an experimental group (DFA-LSTM-assisted simulation). The results showed that the experimental group achieved a pressure MAE of 0.021 MPa, a decrease of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{6 2. 5 \%}$</tex> and <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{4 0. 0 \%}$</tex> compared to the control group and comparison group 1, respectively. The five-year recovery factor prediction error was <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{1. 2 \%}$</tex>, a decrease of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{6 8. 4 \%}$</tex> and <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{5 2. 0} {\%}$</tex> compared to the control group and comparison group 1, respectively. A single simulation took 28.6 hours, a 35.7% reduction compared to the control group. This technology can accurately characterize the flow field and guide parameter adjustments, providing support for improving recovery factors.

Rapid prediction of unsteady flow fields around airfoils is crucial for aerodynamic design and optimization of aircraft. However, traditional wind tunnel experiments and numerical simulation methods incur high computational costs for flow field calculations. This paper proposes an improved method based on deep neural operator networks for rapidly predicting unsteady flow fields of different symmetric airfoils under high angle-of-attack conditions. The method uses incoming flow conditions and airfoil parameters as inputs, reconstructing the flow field through spatiotemporal coordinates to achieve joint prediction of pressure and velocity. Experimental results demonstrate that this method accurately captures the unsteady flow field characteristics of different symmetrical airfoils at various high angles of attack. The inference time for a single time-step case takes only about 0.6 s, which is significantly faster than traditional numerical simulation methods. The proposed method can substantially enhance computational efficiency while maintaining flow prediction accuracy, offering a viable new approach for rapid flow field prediction and aerodynamic design optimization of aircraft.

Deep-learning based surrogate modeling for fast and accurate simulation in realistic 3D reservoir with varying well controls

The symplectic integration method is popular in high-accuracy numerical simulations when discretizing temporal derivatives; however, it still suffers from time-dispersion error when the temporal interval is coarse, especially for long-term simulations and large-scale models. We employ the inverse time dispersion transform (ITDT) to the third-order symplectic integration method to reduce the time-dispersion error. First, we adopt the pseudospectral algorithm for the spatial discretization and the third-order symplectic integration method for the temporal discretization. Then, we apply the ITDT to eliminate time-dispersion error from the synthetic data. As a post-processing method, the ITDT can be easily cascaded in traditional numerical simulations. We implement the ITDT in one typical exiting third-order symplectic scheme and compare its performances with the performances of the conventional second-order scheme and the rapid expansion method. Theoretical analyses and numerical experiments show that the ITDT can significantly reduce the time-dispersion error, especially for long travel times. The implementation of the ITDT requires some additional computations on correcting the time-dispersion error, but it allows us to use the maximum temporal interval under stability conditions; thus, its final computational efficiency would be higher than that of the traditional symplectic integration method for long-term simulations. With the aid of the ITDT, we can obtain much more accurate simulation results but with a lower computational cost. We use the inverse time dispersion transform to the 3rd-order symplectic integration.It can greatly reduce the time-dispersion error, especially for long travel times.It is superior to the conventional 2nd- and high-order finite-difference methods.It allows the maximum temporal interval under stability conditions thus is efficient.We can obtain more accurate simulation results but with a lower computational cost.

Surrogate model for reservoir performance prediction with time-varying well control based on depth generative network

&lt;p class="Author"&gt;Artificial Intelligence (AI) based framework for the Internet of Things (IoTs) have gained worldwide attention in recent years, mainly with the explosion of Micro-Electro-Mechanical Systems (MEMS) technology. Basically, MEMS has facilitated the development of tiny and smart sensors for the IoT-based framework. An AI-based IoT model is an emerging technology that helps in both fault-tolerant as well as energy-efficient data transmission purposes. For efficient data transmission in an IoT-based model, the concept of Wireless Sensor Network (WSN) plays a vital role that comprises various sensor nodes that communicate together to monitor and gather information from the various Region of Interest (RoI). Generally, sensor nodes are tiny in size and having a small battery life, limited sensing, processing, and communication capabilities. So, the fault-tolerant mechanism for energy efficient data transmission in IoT is a good initiative with the combination of Deep Learning as an AI approach. In this research article, the concept of deep learning-based fault-tolerant mechanism in IoT frameworks for energy efficient data transmission is proposed in an optimized manner. Here, the concept of the Grouped-Bee Colony (GBC) algorithm is designed for the fault detection mechanism as an optimization approach along with the Deep Learning as an AI.&lt;em&gt;&lt;/em&gt;&lt;/p&gt;

Integrating physical knowledge and artificial intelligence approaches for simulation and optimization of direct urea fuel cell

Network simulation plays a crucial role in the field of network research, education, and industry. However, before conducting a simulation on traditional network simulators, operators need to develop a simulative behavioral model, which requires intimate knowledge of the simulator implementation. Besides, the behavioral model cannot be migrated directly into real-world devices due to its tight coupling with the simulator platform, resulting in redundant and error-prone codes rewriting. Recently, P4, a high-level domain specific language (DSL), has attracted great attention from both academia and industry for its advantages of enabling operators to define behaviors of the programmable data plane. Inspired by the idea of DSL, we present NS4, a P4-driven network simulator supporting simulation of P4-enabled networks to address the problems existing in traditional simulators. Taking advantage of P4, NS4 simplifies the development of a behavioral model and bridges the gap between simulation and deployment. Furthermore, to the best of our knowledge, NS4 is the first research effort to enable simulation of a P4-enabled network, providing a useful tool for P4 research and development. In this paper, we designed and implemented NS4, consisting of data plane models integrated with ns-3, the state-of-the-art network simulator, and control plane models to interact with the P4 pipeline. Then we evaluated its effectiveness and efficiency by simulating several representative P4 programs. Results show that NS4 can simulate large-scale P4-enabled networks at a low cost.

In the ever‐evolving AI‐driven education, integrating AI technologies into teaching practices has become increasingly imperative for aspiring STEM educators. Yet, there remains a dearth of studies exploring pre‐service STEM teachers' readiness to incorporate AI into their teaching practices. This study examined the factors influencing teachers' willingness to integrate AI (WIAI), especially from the perspective of pre‐service STEM teachers' attitudes towards the application of AI in teaching. In the study, a comprehensive survey was conducted among 239 pre‐service STEM teachers, examining the influences and interconnectedness of Technological Pedagogical Content Knowledge (TPACK), Perceived Usefulness (PU), Perceived Ease of Use (PE), and Self‐Efficacy (SE) on WIAI. Structural Equation Modeling (SEM) was employed for data analysis. The findings illuminated direct influences of TPACK, PU, PE, and SE on WIAI. TPACK was found to directly affect PE, PU, and SE, while PE and PU also directly influenced SE. Further analysis revealed significant mediating roles of PE, PU, and SE in the relationship between TPACK and WIAI, highlighting the presence of a chain mediation effect. In light of these insights, the study offers several recommendations on promoting pre‐service STEM teachers' willingness to integrate AI into their teaching practices. Practitioner notes What is already known about this topic? The potential of AI technologies to enrich learning experiences and improve outcomes in STEM education has been recognized. Pre‐service teachers' willingness to integrate AI into teaching practice is crucial for shaping the future learning environment. The TAM and TPACK frameworks are used to analyse teacher factors in technology‐supported learning environments. Few studies have been conducted for examining factors of pre‐service teachers' willingness to integrate AI into teaching practices in the context of STEM education. What this paper adds? A survey was designed and developed for exploring pre‐service STEM teachers' WIAI and its relationships with factors including TPACK, PE, PU, and SE. TPACK, SE, PU, and PE have direct impact on pre‐service STEM teachers' WIAI. SE, PU, and PE have been identified as mediating variables in the relationship between TPACK and WIAI. Two sequential mediation effects, TPACK → PE → SE → WIAI and TPACK → PU → SE → WIAI, among pre‐service STEM teachers were further identified. Implications of this study for practice and/or policy Pre‐service STEM teachers are encouraged to explore and utilize AI technology to enhance their confidence and self‐efficacy in integrating AI into teaching practices. Showcasing successful cases and practical experiences is essential for fostering awareness of AI integration in STEM education. It is recommended to introduce AI education courses in teacher training programs. Offering internship and practicum opportunities related to AI technologies can enhance their practical skills in integrating AI into education.

Swarm robotics and cooperative AI represent a transformative approach in robotics and artificial intelligence, drawing inspiration from collective behaviors observed in natural systems such as ant colonies, bird flocks, and fish schools. By leveraging distributed control, decentralized decision-making, and local communication among multiple agents, swarm robotics enables scalable, robust, and flexible multi-agent systems capable of tackling complex tasks. Cooperative AI integrates algorithms for coordination, task allocation, and collaborative learning to optimize group performance while maintaining individual autonomy. Applications span autonomous exploration, search and rescue, environmental monitoring, logistics, agriculture, and defense. This article explores the principles, architectures, coordination strategies, learning mechanisms, and real-world applications of swarm robotics and cooperative AI. It also addresses challenges including scalability, communication constraints, fault tolerance, ethical considerations, and human-swarm interaction. Future directions involve integrating advanced AI techniques such as reinforcement learning, neuro-symbolic reasoning, and edge intelligence to enhance adaptability, resilience, and efficiency in collective robotic systems. Swarm robotics and cooperative AI promise to redefine automation, multi-agent intelligence, and collective problem-solving in dynamic and uncertain environments.

This study aims to produce empirical insights into how students in Chinese higher vocational colleges view and use generative AI in EFL learning. This quantitative study uses the technology acceptance model (TAM) to analyze EFL learners' attitudes, intentions, and behaviors about utilizing generative AI in EFL learning with Structural Equation Modelling. 911 English as a foreign language(EFL) learners completed a questionnaire measuring personal innovativeness, perceived ease of use, perceived usefulness, attitude, behavioral intention, and actual use. Structural equation modeling investigations show that Perceived Ease of Use does not directly predict learners' attitudes but can impact them through the full mediator Perceived Usefulness. Positive attitudes towards generative AI and personal innovativeness are associated with increased behavioral intention. Behavior intention predicts the actual application of generative AI in English learning. This study demonstrates the effectiveness of generative AI as a language-learning tool for EFL learners, promoting personalized and sustainable learning.

The current study investigates the factors affecting students' behavioral intentions to utilize "artificial intelligence" (AI) applications in learning by extending the "unified theory of acceptance and use of technology" (UTAUT) model with the inclusion of cognitive flexibility. Data were collected from 315 undergraduate students and analyzed employing a two-step approach involving "structural equation modeling" (SEM) with AMOS. The measurement model was first run to assess and confirm the reliability and validity of the variables, followed by the structural model to test the hypothesized relationships. The results revealed that "performance expectancy (PE), effort expectancy (EE), and social influence" (SI) exerted significant positive effects on students' behavioral intentions to use AI applications, while "facilitating conditions" (FC) exhibited no significant impact. Additionally, "cognitive flexibility" (CF) did not directly influence behavioral intention but exerted a significant indirect effect through its influence on PE, EE, and SI. These findings underscore the significance of students' cognitive adaptability in shaping their perceptions of technological usefulness and ease of use and highlight the value of integrating psychological variables into technology acceptance models in educational contexts. Implications for educational technology implementation and future research directions are discussed.

Project-based learning is a teaching method used in classrooms that integrate authentic and real-world problems in the learning environment. In light of the Industrial Revolution 5.0 where humans are working alongside advanced technology and A.I. generated assistance, the research introduces a project-based learning approach in ESL classrooms with the assistance of AI technologies. The project aims to analyse the usefulness and ease of use of AI-based language learning applications in project-based learning among undergraduates in a public university according to the Technology Acceptance Model (TAM). The research focuses on two main theories as a theoretical framework: project-based learning (PBL) and Technology Acceptance Model (TAM). It is a 6-week proposal writing project conducted with undergraduate students from a public university in Selangor. Project-based learning was introduced, and the participants were informed that they can make full use of any AI-based technologies to support their project. Submission of the proposal was scheduled for the sixth week in which it will be presented in class. A set of questionnaires was used as the instrument to obtain their perception of using AI-based technologies in project-based learning upon task completion. 50 respondents were involved in the study. The findings indicate that students had a positive view of using AI applications in their projects, preferred group work for its enjoyment, challenge, and manageability, and found AI language applications easy to use and beneficial for their learning. Finally, the project’s adaptability makes it suitable for various educational contexts, and integrating AI technology in ESL classrooms helps students complete language tasks effectively.

Grid energy storage system (GESS) has been widely used in smart homes and grids, but its safety problem has impacted its application. Battery is one of the key components that affect the performance of GESS. Its performance and working conditions directly affect the safety and reliability of the power grid. With the development of data analytics and machine learning, the accuracy and adaptability of the battery state estimation model can be greatly improved. This paper proposes a new method to model battery, with low-quality data. First, it designs a data cleaning method for GESS battery operating data, including missing data filling and outlier data repair. Then, the repaired data is used to model battery. A battery mathematical model is proposed based on a deep learning algorithm to realize accurate GESS state estimation. The performance of the developed deep learning method is compared with conventional BP neural network and generalized regression neural network to highlight the technical merits. Results derived in this paper provide a solid basis for high-efficiency GESS operation and energy management.

Air-quality monitoring and forecasting are crucial for atmosphere pollution control and management. We propose an innovative data-driven framework for air quality index (AQI) prediction by integrating GNSS radio occultation (GNSS-RO) observation and weather modeling. Empowered by the state-of-the-art machine learning approach, our method can effectively predict regional AQI with a comparable accuracy much more quickly than the traditional numerical modeling and simulation approach. In a real case study using a representative region of China, our data-driven approach achieves a 2000 times speedup; meanwhile, the prediction error measured by rRMSE is only 2.4%. We investigate further the effects of different models, hyperparameters, and meteorological factors on the performance of our AQI prediction framework, and reveal that wind field and atmospheric boundary-layer height are important influencing factors of AQI. This paper showcases a direct application of GNSS-RO observation in assisting in forecasting regional AQI. From a machine learning point of view, it provides a new way to leverage the unique merits of GNSS atmospheric remote sensing technology with the help of the more traditional weather forecasting modeling approach.