Discover Fundamental Laws
Use observation, experimentation, and theoretical modeling to formulate testable explanations, then establish their reproducibility, scope, and limitations through repeated validation.
Primary Purposes of Scientific Research
Scientific research seeks testable explanations of natural and complex systems while developing knowledge and methods for real-world problems; these aims inform and constrain each other.
Solve Practical Problems
Translate scientific understanding into verifiable models, methods, and engineering solutions for problems in manufacturing, materials, aerospace, and other domains.
Artificial Intelligence Is Driving a Major Leap and Profound Transformation in Scientific Research Paradigms
Scientific research is advancing from observation, theory, computation, and data-intensive discovery toward an intelligent paradigm powered by foundation models and scientific agents.
AI is evolving from an auxiliary analytical tool into a new research infrastructure that connects scientific questions, data, models, experiments, and knowledge—expanding the space scientists can explore, extending cognitive boundaries, and accelerating work on complex scientific problems.
Empirical Paradigm
Archimedes' principle
Theoretical Paradigm
Newton's law of universal gravitation
Computational Paradigm
Global climate models
Data Paradigm
The Human Genome Project
Intelligent Paradigm
Protein structure prediction
Three Meanings of AI for Science
This site organizes AI for Science into three related levels: using AI to solve scientific tasks, using AI to support scientific discovery, and studying AI itself as a scientific system.
Scientific Task Solving
Develop and evaluate AI methods for scientific problems with explicit objectives and evaluation criteria, including equation solving, molecular design, protein structure prediction, and scientific image analysis.
Scientific Laws and Knowledge Discovery
Integrate data, literature, and experimental evidence to identify candidate laws, mechanisms, invariants, and testable hypotheses, together with their validity conditions and uncertainty.
Science of AI
Treat AI systems as empirical and theoretical objects of scientific inquiry, and study scaling, emergence, learning dynamics, and the structure of intelligence.
基于科学数据建模的任务求解
面向表格、时间序列、实验记录与观测数据，研究符合变量语义、实验条件和领域约束的表示学习与预测方法，为科学任务求解提供可检验、可复用的模型基础。
Tabular Data
面向属性表、实验矩阵、材料/分子性质表和科学记录，研究小样本、缺失值、异构字段、分布偏移与领域约束下的学习和不确定性估计。
Time Series
面向实验曲线、传感器序列、仿真轨迹和观测流，刻画多尺度趋势、周期、突变、时序依赖与动态过程，并评估模型在外推和分布变化下的可靠性。
基于科学文献挖掘的知识发现
We organize scientific-literature cognition into five levels: retrieval, structural parsing, information extraction, evidence synthesis, and the formulation of testable hypotheses.
L1
Paper Search
Retrieve and trace literature relevant to an explicit research question, preserving query and citation-expansion paths.
→
L2
Element Interpretation
Parse text, tables, formulas, and figures into structured representations while preserving their document context.
→
L3
Information Extraction
Extract research questions, methods, data, findings, and the evidential relations among them.
→
L4
Knowledge Comprehension
Compare evidence across papers to identify agreement, conflict, applicability limits, and knowledge gaps.
→
L5
Hypothesis Discovery
Formulate falsifiable hypotheses and validation plans from existing evidence; this remains a long-term research direction.
PaperScout
A research agent that plans search, citation expansion, and evidence screening around a scientific question, producing a traceable literature set through iterative interaction.
Project Repo →ChemTable
A benchmark for evaluating multimodal models on structural recognition, symbolic parsing, and semantic understanding of real-world chemical tables, including formulas, table relations, and molecular diagrams.
Project Repo →ScholarSum
A scientific summarization method combining knowledge-graph reasoning and reflective refinement, with emphasis on consistency among methods, evidence, and conclusions.
OpenReview →PaperArena
A benchmark designed to evaluate retrieval, evidence integration, and reasoning by tool-augmented agents across multiple scientific papers.
Project Repo →Mind2Report
A deep-research agent for complex research questions that uses intent-guided retrieval and iterative synthesis to produce evidence-supported analytical reports.
Project Page → GitHub ↗Research Intelligence Vision
Our long-term objective is to develop research agents that identify knowledge gaps under evidential constraints, formulate falsifiable hypotheses, and propose validation plans.
Long-term ResearchScience of AI
将 AI 系统作为复杂科学对象，研究规模扩展、能力涌现、学习动力学与智能结构，并发展可检验的描述变量和理论。
规模与涌现
检验 Scaling Law 在何种条件下成立、表观涌现是连续变化还是类临界跃迁，以及结论如何依赖评价指标、数据与训练过程。
学习与智能机制
分析强化学习与上下文学习如何改变模型行为，并检验记忆、压缩、预测与智能能否获得统一描述。
Agent 与群体智能体
比较人工智能与人类智能的相关机制，并检验 Agent 协作、自主性和群体行为中是否存在可迁移的动力学模式。