MindSpore Technical Forum: New Trend of AI for Science

In recent years, Artificial Intelligence (AI) has been dominating many scientific fields, fueling a new trend: AI for Science. The purpose of AI for Science is to find ways to use AI to help people transform masses of data first into information, and then into usable knowledge. This approach poses challenges to conventional scientific research, and it remains unclear if we are ready for it. In order to learn more about this potentially revolutionary method, the MindSpore community invited key experts to share their insights in the second MindSpore Technical Forum, held on April 16. Key topics included not only AI for Science, but also related models, experimental methods, and new industry forms. The following is a short review of speeches given at the forum.

Mr. Huang Xiaomeng, Associate Professor from the Department of Earth System Science at Tsinghua University, Chief Engineer of the National Supercomputing Center in Wuxi, shared his views on the application of and challenges to deep learning in earth system science. Deep learning has been widely adopted in earth system science for purposes such as remote sensing image creation, extreme weather detection, weather forecasting, climate predictions, and smart reconstruction of physical parameterization in numeric mode. Mr. Huang pointed out that deep learning brings new research methods to the discipline, especially inspiration for how to address the physical inconsistency of black-box models. He also introduced one application of deep learning to earth system science: the earth system numerical simulation program, which aims to create a digital twin earth in a supercomputer to simulate real earth environments.

Mr. Yang Yi, Associate Researcher from the Institute of Systems and Physics Biology, Shenzhen Bay Laboratory, explained how molecular dynamic simulation works in the AI field. Molecular dynamic simulation is an important scientific research method. With the rapid development of AI in recent years, powerful AI-based algorithms such as AlphaFold have emerged in molecular dynamics. Their superb performance has demonstrated that AI will soon bring significant changes to molecular dynamic simulation and even the entire scientific computing field. Mr. Yang Yi and his team developed MindSpore SPONGE, a next-generation intelligent molecular dynamic simulation software, based on MindSpore. MindSpore SPONGE not only overcomes the compatibility problems of conventional molecular dynamic simulation software and AI frameworks, but also implements functions that cannot be implemented by conventional software, such as meta-optimization. Thanks to these features, MindSpore SPONGE has enormous potential to become crucial in various fields, such as drug screening, protein structure prediction and new material design.

Dr. Man-Hong Yung, Huawei's Chief Quantum Computing Software and Algorithm Scientist, introduced quantum computing in the AI field. As a disruptive computing technology, quantum computing has been rapidly improving and creating new industries since it was proposed in 1980. It is now possible to make quantum processors with over 50 qubits, unleashing more computing power than supercomputers. But such processors need practical algorithms and software. To fuel the development of quantum software, the MindSpore team developed MindQuantum, a general-purpose framework for quantum computing which supports quantum-classical hybrid computing and runs smoothly both on classical simulators and quantum processors.

Mr. Dong Bin, Associate Professor of the Beijing International Center for Mathematical Research & Center for Machine Learning, Peking University, explained that deep learning has achieved some success in fields such as computer vision and natural language processing, and its influence has been growing, extending to scientific and engineering research fields. In his speech, Mr. Dong introduced the impact of deep learning on computational mathematics and scientific computing, recent research on the relationship between convolutional neural networks and discrete forms of differential equations, as well as ways to guide the modeling and computing of mechanism and data convergence. He also described designing interpretable data-driven models for system identification and model simplification and combining numerical partial differential equations with machine learning to create data-driven solvers for electromagnetic simulation.

Mr. Yu Pan, Senior Architect of MindSpore, introduced the trend of AI for Science and current MindSpore practices following this trend. MindSpore has been engaged in scientific computing since its inception. The original MindSpore computing engine was upgraded, based on multi-scale hybrid computing and high-order hybrid differentiation, to a unified engine for AI and scientific computing, achieving converged and unified acceleration. MindSpore has made breakthroughs in multiple scientific fields in the past two years. It supports scientific computing, makes breakthroughs in high-level and high-dimensional automatic differentiation, automatic heterogeneous parallelism, and cross-scale computing convergence. It also provides an all-scenario framework with eight suites at its core, to support multiple data-driven or physics-driven AI paradigms.

As a domestic framework for AI and scientific computing, MindSpore continuously improves its running efficiency and user experience while innovating its technologies across eight suites covering manufacturing, meteorology, pharmacy, aerospace, automobile, energy, finance, and material fields. The first four suites have already been put into practice. Among them, MindSpore Elec for AI electromagnetic simulation and MindSpore SPONGE for AI biopharmacy have yielded astonishing outputs.

To learn more about MindSpore, please visit the following links: Official website: https://www.mindspore.cn/en Gitee: https://gitee.com/mindspore/mindspore GitHub: https://github.com/mindspore-ai/mindspore