直播时间：2025年9月23日（周二）20:00-21:30

直播平台：

科学网APP

https://weibo.com/l/wblive/p/show/1022:2321325214035133464623
（科学网微博直播间链接）

科学网微博

科学网视频号

北京时间9月23日晚八点，iCANX Youth Talks第116期邀请到了深圳屹艮科技创始人兼CEO、北京大学深圳研究生院新材料学院长聘副教授郑家新，字节跳动Seed - AI for Science团队研究员余旷担任主讲嘉宾，国防科技大学长聘研究员王珊珊、北京大学信息工程学院助理教授袁粒担任研讨嘉宾，中国科学院北京纳米能源与系统研究所研究员唐伟担任主持人，期待你一起加入这场知识盛宴。

【嘉宾介绍】

郑家新

深圳屹艮科技/北京大学深圳研究生院

AI for Science and Science for AI

【Abstract】

Chips, innovative drugs, and new energy respectively carry the three cornerstone industries of human civilization: information, health, and energy. They embody the wisdom of human civilization and the exploration of the unknown. Their research and development paradigms are gradually transitioning from "experimental trial and error", "theoretical deduction", and "computational simulation" to "data and AI-driven". Currently, AI for Science (AI4S) is highly anticipated. Not only are traditional tech giants making moves in AI4S, but a large number of academic achievements and AI4S startups are also emerging like mushrooms after rain. However, behind the hype, there are few successful cases of "industrial application", the touchstone of research and development, and a lot of doubts have followed: Can AI4S be implemented? If it can, how long will it take? Can large models similar to ChatGPT be developed in the fields of semiconductors, new drug research and development, new energy, and new materials? To truly solve the implementation of AI4S, more and more people now realize that the first step is to do Science for AI (S4AI). Because the richness and constraints of physical models can not only provide AI with rich virtual data, but also greatly improve the prediction accuracy of AI models, thereby truly realizing industrial value.

芯片、创新药、新能源分别承载着人类文明的三大基石产业：信息、健康、能源，凝聚了人类文明的智慧结晶和对未知的探索，其研发范式正在从“实验试错”、“理论推导”、“计算仿真”逐步过渡到“数据与AI驱动”。 当前AI for Science（AI4S）被寄予厚望，不仅传统科技大厂纷纷布局AI4S，大量学术性成果和AI4S创业公司也如雨后春笋般冒出来。然而热潮背后，对于“产业应用”这块研发的试金石，AI4S却鲜有成功落地的案例，大量质疑声也随之而至：AI4S能落地吗？如果能落地，还需要多久？在半导体、新药研发、新能源与新材料领域能发展出类似于ChatGPT的大模型吗？要真正解决AI4S的落地，现在越来越多的人意识到首先要做好Science for AI(S4AI)。因为物理模型的丰富和约束不仅能为AI提供丰富的虚拟数据，同时极大提高AI模型的预测精准度，从而真正发挥产业价值。【BIOGRAPHY】

The founder and CEO of Shenzhen Eacomp Technology, and the director of the Peking University Shenzhen Graduate School - Eacomp Technology Joint Laboratory, is also a tenured associate professor at the School of Advanced Materials of Peking University Shenzhen Graduate School. His main research areas include AI-driven cross-scale simulation and design of materials, and the development of battery design automation software (BDA) for new energy. He received dual bachelor's degrees in physics and mathematics from Peking University in 2008 and a Ph.D. in condensed matter physics from Peking University in 2013. He has served as the vice dean of the School of AI4S at Peking University, the vice dean of the School of Advanced Materials at Peking University Shenzhen Graduate School, and the deputy director of the Science and Technology Innovation Bureau of Nanshan District, Shenzhen. He has published over 160 SCI papers with an H-index of 60. He has led major projects of the National Natural Science Foundation of China, including key projects, major projects, general projects, and youth projects, as well as horizontal projects from CATL. He has participated in three national key research and development plans. He has won the First Prize of Natural Science of Shenzhen and the Shenzhen Youth Science and Technology Award. For many years, he has been continuously included in the list of the world's top 2% scientists released by Stanford University.

深圳屹艮科技创始人兼CEO，北大深研院-屹艮科技联合实验室主任，北京大学深圳研究生院新材料学院长聘副教授、博士生导师。主要从事AI驱动的材料跨尺度模拟与设计、新能源电池设计自动化软件（BDA）开发。于2008年从北京大学获得物理与数学双学士学位，2013年从北京大学获得凝聚态物理博士学位，曾担任北京大学科学智能学院副院长、新材料学院副院长、深圳市南山区科技创新局副局长。至今发表SCI论文160余篇，H-index为60。主持国自然重点/重大项目课题/面上/青年、宁德时代横向课题等，参与国家重点研发计划3项。曾获得深圳市自然科学一等奖、深圳市青年科技奖。多年来连续入选斯坦福发布的全球前2%顶尖科学家单榜（World’s Top 2% Scientists）。

余旷

字节跳动

DMFF：自动微分技术在势能面优化中的应用

【ABSTRACT】

In the development of classical force fields, tuning force field parameters based on experimental properties is a commonly used traditional technique that heavily relies on human physical intuition. With the increasing popularity of machine learning potential (MLPs) and high-throughput experimental methods, the tuning of potential energy surfaces based on macroscopic data is facing new challenges. These challenges include: 1. The enormous number of MLP parameters, with no clear physical meanings for human interpretation; 2. The rapid growth in the volume of experimental data, along with the wide variety of measurable experimental data types which exhibit multimodal characteristics. As both the data volume and number of parameters grow significantly, manual parameter tuning is no longer feasible, and there is an urgent need for automated optimization methods. Recently, based on automatic differentiation technology, we have developed a differentiable molecular simulation platform (DMFF) and realized the gradient propagation from macroscopic properties to force field parameters. DMFF enables us to combine multiple experimental properties to define loss functions and accomplish the automatic optimization of force field parameters. In this report, we will introduce DMFF's automatic differentiation algorithms for thermodynamic ensemble averages, free energy, and dynamic quantities (e.g., infrared spectra), and demonstrate its applications in important systems such as electrolytes and liquid water. These cases show that with the help of automatic differentiation technology, we can effectively learn microscopic potential energy surfaces from batch experimental data, significantly improving the predictive accuracy of MD.

在经典力场开发中，针对实验性质对力场参数进行调优是一项经常使用、但高度依赖人类物理直觉的传统技艺。在机器学习势能面（MLP）和高通量实验手段日益普及的今天，基于宏观数据的势能面调优面临全新的挑战。这些挑战包括：1. 势能面参数数量巨大，且人类对MLP参数缺少物理理解；2. 实验数据数量快速增长，可测量的实验数据种类多、呈现多模态特征。在数据量和参数量都显著增长的情形下，人工调参已不再可行，我们迫切地需要自动化的势能面优化算法。近期，基于自动微分技术，我们开发了可微分的分子模拟平台DMFF，并实现了MD宏观性质到力场参数的梯度传递。利用DMFF可以组合多种实验性质定义损失函数，并完成力场参数的自动优化。在本报告中，我们将分别介绍DMFF对热力学系综平均量、自由能、以及动力学量（例如红外光谱）的自动微分算法，并展示其在电解液、液态水等重要体系中的应用。这些案例表明，借助自动微分技术，我们可以从批量实验数据中有效学习微观势能面，显著提升MD的预测精度。

【BIOGRAPHY】

Dr. Kuang Yu graduated from the College of Chemistry and Molecular Engineering, Peking University in 2008, and obtained his doctoral degree from the University of Wisconsin-Madison in 2013. From 2013 to 2016, he conducted postdoctoral research at the Department of Mechanical and Aerospace Engineering, Princeton University. He then served as a Research Scientist at D. E. Shaw Research (USA) from 2016 to 2018. From 2018 to 2025, Dr. Yu was appointed as an Assistant Professor and later an Associate Professor at the Tsinghua-Berkeley Shenzhen Institute (TBSI) and Tsinghua University Shenzhen International Graduate School (TSIGS). In September 2025, he joined the Seed-AI for Science Team of ByteDance as a Researcher in Computational Materials Science. Dr. Yu's primary research focus lies in the development of molecular force fields based on artificial intelligence technologies. He has published over 60 papers in SCI-indexed journals, including Nature Communications, Physical Review Letters, Journal of the American Chemical Society, and Journal of Chemical Theory and Computation.

余旷博士2008年毕业于北京大学化学与分子工程学院，2013年获威斯康星大学麦迪逊分校博士学位。2013-2016年于普林斯顿大学机械与航天工程系从事博士后研究，2016-2018年为美国D. E. Shaw Research公司研究科学家。2018-2025年受聘于清华-伯克利深圳学院、清华大学深圳国际研究生院，任职助理教授、副教授。2025年9月入职字节跳动Seed - AI for Science团队任职计算材料学研究员。主要工作方向为基于人工智能技术的分子力场构建，并在Nature Communications、Physical Review Letters、Journal of American Chemical Society、Journal of Chemical Theory and Computation等SCI期刊上发表论文60余篇。

【主持人】

唐伟

中国科学院北京纳米能源与系统研究所

【研讨嘉宾】

王珊珊

国防科技大学

袁粒

北京大学