Building reliable AI scientists构建可靠的 AI 科学家

Large language model agents can now write code, call simulation tools, and reason over results. We build the scientific AI agents themselves: reusable scientific skills, reliable tool use for atomistic simulation, self-checking and uncertainty awareness, and rigorous benchmarks that measure whether an agent can really be trusted to do research. 大语言模型智能体已经能够编写代码、调用模拟工具并对结果进行推理。我们研究科学智能体本身：可复用的科学技能、可靠的原子模拟工具调用、自我检验与不确定性感知，以及衡量智能体能否被真正信任去做科研的严格评测基准。

Related work相关工作

Atomistic machine learning is machine learning for systems made of atoms — inorganic materials, organic molecules, and drugs. It learns their physics and chemistry from data, spanning graph neural networks (GNNs), machine learning interatomic potentials (MLIPs) that bring quantum-mechanical accuracy to simulations orders of magnitude faster, and generative models for crystal structures. We develop foundation models (e.g. CHGNet) and libraries (e.g. MatGL), the datasets and benchmarks that train and test them, and the understanding needed to trust them — systematic errors, cross-functional transfer, thermodynamics, and scalable distributed inference. Atomistic Machine Learning（原子体系机器学习）是面向由原子构成的体系——无机材料、有机分子、药物分子等——的机器学习，从数据中学习其物理与化学规律，涵盖图神经网络（GNN）、以接近量子化学的精度将原子模拟加速数个量级的机器学习势函数（MLIPs），以及晶体结构生成模型等。我们研发原子体系基础模型（如 CHGNet）与开源库（如 MatGL），构建训练与评测所需的数据集和基准，并深入理解其可靠性——系统性误差、跨泛函迁移、热力学性质以及大规模分布式推理。

Related work相关工作

Quantum chemical simulations — such as density functional theory (DFT) and molecular dynamics — let us study materials and chemistry at the level of electrons and atoms: predicting properties, uncovering the mechanisms behind materials behavior, and guiding the design of new materials. Combined with machine learning and statistical mechanics, they let us explore large chemical spaces. Past work includes battery cathodes, solid-state ionic conductors, and crystal structure prediction. 以密度泛函理论（DFT）与分子动力学为代表的量子化学计算，让我们能够从电子与原子层面研究材料与化学体系：预测材料性质、揭示性能背后的机理，并指导新材料的设计。结合机器学习与统计力学，我们得以探索更广阔的化学空间。已有工作涵盖电池正极材料、固态离子导体与晶体结构预测等。

Related work相关工作

Our ultimate goal is AI scientists that carry out discovery end to end: reading the literature, forming hypotheses, designing and running computational workflows with atomistic ML and quantum simulations, analyzing the results, and deciding what to try next — iterating until they find something new. Automated and robotic labs are one of the tools such an agent can call to validate its predictions, but the core of this direction is the agent's own scientific reasoning and decision-making across the whole research cycle. This is a core direction of the new lab at Fudan AI³. 我们的终极目标是让 AI 科学家端到端地完成科学发现：检索与阅读文献、提出科学假设、借助 Atomistic ML 与量子化学计算设计并执行计算流程、分析结果并自主决定下一步，不断迭代直至获得新的发现。自动化 / 机器人实验平台是智能体可以调用、用于验证预测的工具之一，而这一方向的核心在于智能体贯穿整个科研流程的科学推理与决策能力。这是实验室在复旦 AI³ 研究院的核心发展方向之一。

New projects starting in 2027 — join us to shape them.相关课题将于 2027 年启动——欢迎加入，共同开拓。