On July 13, Flying Paddle joined hands with leading experts and scholars in the field of artificial intelligence scientific computing (AI for Science), teachers from universities and scientific research institutions, and related industry practitioners to hold the Flying Paddle Scientific Computing Offline in Shanghai Baidu Flying Paddle Empowerment Center Exchange. At the meeting, experts in various fields discussed and exchanged topics such as deeply integrated platforms, industry trends, scientific research results, and open source construction. This Flying Paddle Science and Technology Computing Exchange Conference brought together innovative achievements in various fields and promoted the construction and development of AI for Science in China.
At the meeting, Zhang Yanbo, the senior technical product manager of Flying Paddle, and He Sensen, an engineer, respectively introduced the current status of exploration and product construction in the field of Flying Paddle scientific computing, as well as the functions and usage methods of PaddleScience v1.0, a component of Flying Paddle’s scientific computing tool. The meeting also invited many well-known scholars at home and abroad to publish cutting-edge academic reports on AI for Science. Associate Professor Meng Xuhui from Huazhong University of Science and Technology published a report entitled "Scientific Computing: Composite Neural Networks Fused with Multi-Fidelity Data"; Assistant Professor Wang Yunbo from Shanghai Jiaotong University published a report entitled "World Model: Intuitive Physical Reasoning and Decision-making"; Sweden Academician Erik Dahlquist and Madeleine Martinsen, academician of Melladalen University, published a report entitled "Industrial Digitalization: Smart Industries demands Smart Services"; Wang Shuo, a young researcher at Fudan University, published a report entitled "Medical-Industrial Integration: Combining Data-Driven and Biomechanical Cardiovascular Disease Modeling” report. In addition, Mr. Hua Haobo from the National Supercomputing Center in Zhengzhou, Dr. Zhang Zhiqiang from Communication University of China, and Dr. Zhu Weiguo from Beijing Jiaotong University, as representatives of outstanding developers of Flying Paddle AI for Science, discussed "Scientific Computing and Domestic Chips" and "Scientific Computing" respectively. Related reports were published on the three topics of "Computing and Humanities and Social Sciences" and "Combination of Flying Paddle and Other Computing Components", fully reflecting the extensive adaptation and support of Flying Paddle in the direction of AI for Science. 
This event attracted more than 40 teachers, students and practitioners from universities and enterprises all over the country to participate in the meeting, and the online live broadcast has accumulated more than 8,000 views. Through this exchange, it is hoped that more developers will have the opportunity to gain an in-depth understanding of the practical tools of Paddle AI for Science, as well as cutting-edge technologies and field applications.
Paddle Scientific Computing Components:
PaddleScience v1.0
Based on the high-level API and high-order automatic differentiation mechanism of Paddle's deep learning framework, Paddle has been upgraded simultaneously and launched a scientific computing tool component: PaddleScience v1.0. Aiming at the high-dimensional, time-consuming, and cross-scale challenges faced by traditional numerical calculation methods, the integrated mathematical calculation and physical data processing method provides physical mechanism, data-driven and other paradigms to solve problems. At the same time, it builds classic AI for Science cases around computational fluid dynamics (CFD), structural finite element simulation, weather forecasting, etc., and provides reusable open source codes of cases for researchers to promote the integration of AI and basic science.
The official version of PaddleScience v1.0 includes the following four features:
- API schema update
From the perspective of user usage habits, taking into account deep learning and CFD & CAE user experience, API elements are updated from the perspectives of data preprocessing, model selection, network optimization solution, and result post-processing to improve user experience.
- Rich scene cases
Provide basic cases such as 2D & 3D cylindrical flow, vortex-induced vibration, convective heat dissipation, and equation inversion. At the same time, new cases of 2D & 3D structural force analysis in the structural field, weather forecast, pollutant diffusion and other related cases in the meteorological field, It supports direct reuse and secondary development, and users can directly experience the practice of related projects from the Github code warehouse or AI Studio.
- Underlying model update
Added classic neural network models such as CNN, U-Net, Transformer, and GAN, and FNO operator learning models, and provided corresponding verification cases.
- API upgrade update
Provide a newly designed API for users to customize partial differential equations and define various boundary conditions, support 2D&3D basic geometry definition, STL complex shape analysis and Boolean operations, etc., and provide quasi-random sampling, local encryption sampling and other functions.
For specific functions, please refer to: https://paddlescience-docs-hss.readthedocs.io/zh/latest/
Academic Frontiers, Exploring the Mysteries of Science
Associate Professor Meng Xuhui from Huazhong University of Science and Technology published a report entitled "Scientific Computing: Composite Neural Networks Fused with Multi-Fidelity Data". In the report, Mr. Meng introduced the underlying mathematical principles of the PINN network and the composite neural network, and proposed that the computational efficiency and accuracy of PINN can be improved by fusing high-fidelity and low-fidelity data.
Assistant Professor Wang Yunbo from Shanghai Jiaotong University published a report entitled "World Model: Intuitive Physical Reasoning and Decision-Making". In the report, Mr. Wang introduced the concept of the world model, modeling the physical simulation as an inverse problem of image generation, replacing the numerical solution of the dynamic equation with the inverse graphics optimization of the neural network fluid model, so as to realize the inverse calculation of the fluid based on real physical observations Attributes.
Academician Erik Dahlquist and teacher Madeleine Martinsen from Melladalen University in Sweden published a report entitled "Industrial Digitalization: Smart Industries demands Smart Services". The two teachers showed everyone the demand for digitalization and artificial intelligence in the field of energy mining, and proposed the concept of a general intelligent body for industrial operation and maintenance based on the AI-AR technical framework.
Wang Shuo, a young researcher at Fudan University, published a report titled "Medical-Industrial Integration: Combining Data-Driven and Biomechanical Cardiovascular Disease Modeling". Teacher Wang introduced the modeling process of cardiovascular diseases, and proposed to use neural network to realize the prediction of vascular plaque stress, which can greatly reduce the time-consuming solution and the demand for grid quality compared with the finite element method.
Through the wonderful reports of various experts, we can see that artificial intelligence is affecting and changing the realization of scientific research from multiple fields and perspectives, and AI for Science is becoming a new paradigm of global scientific research.
Exchange and discussion, co-creation and commendation
In recent years, Flying Paddle has continued to devote itself to the product innovation of AI for Science and the construction of a cross-type scientific research ecosystem, and released the AI for Science co-creation plan in early 2023, expecting to join hands with partners such as enterprises, universities, scientific research institutes and supercomputing, Jointly build a top open source project based on Paddle AI for Science, create an active and forward-looking AI for Science open source community, and promote scientific research innovation and industrial empowerment through the closed loop of industry, education and research. The plan has attracted more than 40 domestic and foreign college teachers, students and practitioners to sign up for participation. Under the one-on-one guidance of flying paddle experts, 11 teams have completed the project delivery, and the results cover aerodynamics, structural mechanics, and meteorology. , Computational Communication and many other fields. Among them, the project "Rapid Prediction of Atmospheric Pollutant Diffusion Based on Data-Driven U-Net Model" was included in the First National Data-Driven Computational Mechanics Symposium (Dalian), and several teams are writing academic journal manuscripts based on co-creation projects. In the future, I believe we will see more co-creation achievements of flying paddles published in AI for Science related conferences or journals.
The vigorous development of Paddle AI for Science is inseparable from the in-depth cooperation and strong support of scholars and practitioners in various fields. In order to commend the experts who have made positive contributions in the construction of the AI for Science community, the conference site also awarded them the "Baidu Flying Paddle AI4S Academic Tutor Certificate".
In the future, Paddle will continue to increase its technical support for AI for Science, empower the development of artificial intelligence and science, gather scientific research achievements, human resources and product innovation, and provide solid strength for the development of AI for Science.