The Interdisciplinary Computing/AI and Physical Sciences Forum is co-hosted by Tsung-Dao Lee Institute (TDLI) and Network & Information Center (NIC) of Shanghai Jiao Tong University. This forum aims to bridge TDLI researchers with concrete examples of "AI for science" that are happening across SJTU campus, and promote research collaborations between TDLI and NIC.
Date: Tuesday, December 9
- 12:00-12:30: Lunch
- 12:30-14:30: Salon
Venue: Meeting Room N102, TDLI
Talk I
Title: An Introduction to the Zhiyuan-1 AI Platform / “致远一号” 集群服务介绍
Speaker: Jiayin Liu (刘佳音), Network and Information Center, Shanghai Jiao Tong University
Abstract:
Zhiyuan-1 is a high-performance computing and large-model service platform newly launched by Shanghai Jiao Tong University in the Fall semester of this academic year. This presentation will provide an overview of the platform's resources and usage framework, including its computing capabilities and service model, procedures for resource application and utilization, and the user support and technical assistance system. The talk will also feature representative application cases to demonstrate how the platform's resources are applied in scientific computing and intelligent model development, aiming to help faculty and researchers gain a comprehensive understanding of the platform and utilize it effectively in their research.
“致远一号” 是上海交通大学于本学期秋季新近投入使用的高性能计算与大模型服务平台。本次报告将围绕平台的资源情况与使用方式展开。内容包括：平台的资源和服务模式，资源申请与使用方式，面向用户的服务体系与技术支持流程。报告还将结合示范性应用案例，展示各类资源在科研计算、智能模型开发中的实践场景，旨在帮助教师和科研人员全面掌握平台能力并高效开展相关科研工作。
Biography:
Jiayin Liu received her master's degree from Shanghai Jiao Tong University in 2024 and subsequently joined the University’s Network and Information Center. She is responsible for user support and services for the Zhiyuan-1 large-model API. She is also involved in the deployment and performance evaluation of local large models, as well as the development of agent systems for various application scenarios. Her work focuses on advancing and optimizing the university's intelligent computing infrastructure.
刘佳音于 2024 年获得上海交通大学硕士学位，随后加入上海交通大学网络信息中心，从事“致远一号”大模型 API 的用户服务与支持工作。同时参与本地大模型的部署与性能测试、以及面向应用场景的智能体系统研发等工作，致力于推进校内智能计算平台的建设与优化。
Talk II
Title: Migration Methods for AI Applications on Ascend Processors: A Case Study of Federated Learning for Wind Turbine Fault Diagnosis / AI 应用在昇腾处理器上的迁移方法 —— 以“基于联邦学习的风机故障诊断”项目为例
Speaker: Kunpeng Xu (徐琨鹏), Network and Information Center, Shanghai Jiao Tong University
Abstract:
In response to current limitations in accessing high-end computing resources, the Network and Information Center at Shanghai Jiao Tong University provides domestic computing services based on Ascend 910B processors through the Zhiyuan-1 cluster. To address the adaptation challenges associated with migrating from NVIDIA GPUs to Huawei Ascend NPUs, the Center offers a mature PyTorch migration solution with comprehensive support throughout the entire process-from migration analysis to performance tuning.
This report presents the "Federated Learning for Wind Turbine Fault Diagnosis" project by Associate Professor Li Yanting's team from the School of Mechanical Engineering as a case study, showcasing the Center's capabilities in application porting and optimization. After tuning, the model's training performance on the Ascend platform increased by 58% compared with the initial migration stage, ultimately reaching 88% of the NVIDIA A100's performance while maintaining equivalent accuracy. This case demonstrates the feasibility and robustness of domestic computing platforms in supporting complex scientific research tasks.
面对高端算力受限的现状，上海交通大学网络信息中心依托“致远一号”集群提供基于昇腾 910B 的国产算力服务。针对从 NVIDIA GPU 向华为昇腾 NPU 迁移所面临的适配挑战，中心提供成熟的 PyTorch 迁移方案，涵盖从“迁移分析”到“性能调优”的全流程技术支持。本次报告将以机械与动力工程学院李艳婷团队的“基于联邦学习的风机故障诊断”项目为例，展示中心在应用移植与性能优化方面的服务成效。经过性能调优后，该模型在昇腾平台上的训练性能较迁移初期提升 58%，达到 A100 的 88%，且模型精度保持一致。该案例充分验证了国产算力支撑复杂科研任务的可行性与可靠性。
Biography:
Xu Kunpeng joined the Network and Information Center at Shanghai Jiao Tong University in 2025. He serves as an AI Application Development Engineer, specializing in the development of AI applications and intelligent agents based on the Zhiyuan-1 cluster.
徐琨鹏于 2025 年加入上海交通大学网络信息中心，担任 AI 应用开发专员，从事基于“致远一号”集群的 AI 应用与智能体开发工作。
Talk III
Title: AI4Tianyu and Tianyu4AI / AI在天语计划中的应用与天语计划对AI的启发
Speaker: Yicheng Rui (芮易成), PhD student, Tsung-Dao Lee Institute, Shanghai Jiao Tong University
Abstract:
The Tianyu Telescope is a 1-meter photometric survey telescope located in Lenghu, Qinghai. Thanks to fast online stacking and reduction, the Tianyu Telescope can observe targets from 10th to 20th magnitude, spanning a factor of 10,000 in brightness. However, this observing mode also brings severe challenges to data processing and observation control. From the initial stage of weather recognition, to target source detection, light-curve extraction in the raw data processing, identification of real versus spurious sources in difference images, and finally scientific discovery at the science end, artificial intelligence is a powerful assistant for photometric survey telescopes represented by Tianyu. Moreover, the 20 TB of light-curve data and 40 TB of image data that the Tianyu Telescope is expected to produce each year will be a valuable resource for research in AI fields such as time-series analysis and image processing.
天语望远镜是位于青海冷湖的一米测光巡天望远镜. 得益于快速的在线叠加结算, 天语望远镜可以对10等到20等的目标进行观测, 其亮度像差10000倍. 但是, 这样的观测模式也给其数据处理与观测控制带来严峻的挑战. 从一开始的天气识别, 到目标源检测 ,原始数据处理中的光变曲线获取, 差分图像的真假源识别, 到科学端的探索发现, 人工智能都是以天语为代表的测光巡天望远镜的强大助手. 而天语望远镜预计每年产生的20TB光变曲线与40TB图像数据也是研究时间序列和图像处理等人工智能领域的宝贵资料.
Hosts:
James Lin (林新华)
Jianglai Liu (刘江来)
Contact: Banghui Ni (nbh@sjtu.edu.cn)
We warmly welcome you to join. Please register via the link below or scan the QR code to help us with the arrangements.
Register deadline：22:00, December 6
https://wj.sjtu.edu.cn/q/wLijWUsp
