NMR 脉冲程序生成模型与评测基准（撰写中）
周子航. (在研). NMR 脉冲程序生成模型与评测基准.
CV
教育经历
- 博士（在读），清华大学，2023 – 至今
- 生命科学学院 · 薛毅实验室 (Rainbow Lab) · 导师：薛毅 教授
- 研究方向：NMR 与人工智能交叉 —— 科学程序（pulse program）生成 + NMR 实验验证
- 目标：用大模型生成可实际执行的 Bruker 脉冲程序，并用模拟与实验闭环验证其正确性
- 本科，生命科学，南方科技大学，2019 – 2023
科研经历
- 2023 – 至今：博士研究生，清华大学
- 微调 NMR pulse program 生成模型（Qwen3.5-27B，SFT + DPO，自建 1,800+ 偏好对训练数据）
- 构建 NMR pulse program 评测基准：observable 粗匹配 + strict score + coverage/confidence 多级架构，50 文件 / 2,192 条有效语句回归验证，并行化结果逐字节一致
- 打通 NMRSIM 验证闭环：AI 生成程序与官方程序在规范化 pulse pattern 层面做等价性评测
- 构建 NMR 知识库与 RAG 检索：70+ 份 Bruker/NMR 手册结构化入库，支撑生成与检索
- 修正与标注 Bruker 脉冲序列解析（phase cycle、多通道脉冲、梯度、shape envelope 等）
开源与工程
- 维护开源项目：ccse（Claude Code Switch Enhanced）、GraphPulse（脉冲序列时序图可视化）、NMRagent（NMR 智能 Agent）、evalpulse（脉冲程序评测基准）、pp2ast（脉冲程序 AST 语法解析器）、nmr-posttrain（脉冲程序 LLM 后训练套件）、topspincli（TopSpin 命令行与模拟工具）、topspinRE（脉冲程序逆向与 AST 解析）、competition（可解释 RNA 能量模型）、xingling（星灵小说站）、AIlove（AI 情侣匹配）、spinachpython、pp2label、prodopy-wt
- 独立开发并部署多个科研与创作服务：NMR Pulse Program AI 生成平台、ccse 在线服务、NMRagent 智能助手、GraphPulse 脉冲序列可视化器、星灵小说站、AI 情侣匹配 及配套自托管服务
荣誉与竞赛
- 优秀奖，第一届全球大学生生命科学挑战赛（华北赛区 · 研究生组 · 赛道二：AI 基因编辑与核酸工具设计），2026年9月
技能
- 编程与工程：Python、Bash、Docker、Linux、Git
- AI 与模型：LLM 微调（SFT / DPO）、数据集构建与审计、Agent 工具开发、模型评测与 benchmark 设计
- NMR 领域：Bruker TopSpin、脉冲序列设计与解析、NMRSIM 模拟验证
Publications
开源与服务
- 长期维护多个 NMR × AI 方向开源项目，并提供使用文档与 issue 支持
Education
- Ph.D. in Life Sciences, Tsinghua University, 2023 – Present
- School of Life Sciences · Yi Xue Lab (Rainbow Lab) · Advisor: Prof. Yi Xue
- Focus: Intersection of NMR and AI —— Scientific Pulse Program Generation + Experimental Validation
- Goal: Generating executable Bruker pulse sequences via domain LLMs with simulation and experimental closed-loop verification
- B.S. in Biological Sciences, Southern University of Science and Technology (SUSTech), 2019 – 2023
Research Experience
- 2023 – Present: Doctoral Researcher, Tsinghua University
- Fine-tuning domain LLMs for NMR pulse programs (Qwen3.5-27B, SFT + DPO, 1,800+ curated preference pairs)
- Developing multi-tier NMR pulse program benchmarks: observable matching + strict statement scoring + simulation coverage, covering 50 files / 2,192 valid statements with bit-level reproducibility
- Establishing NMRSIM validation loop: evaluating physical equivalence between AI-generated and official Bruker programs across normalized pulse patterns
- Constructing NMR knowledge bases and RAG retrieval over 70+ technical manuals
- Structural annotation and parsing of Bruker pulse sequences (phase cycles, multi-channel RF, gradients, shape envelopes)
Open Source & Engineering
- Core repositories: ccse (Claude Code Switch Enhanced), GraphPulse (Pulse Sequence Visualizer), NMRagent (AI Agent for NMR), evalpulse (Pulse Program Benchmark Suite), pp2ast (Pulse Program AST Parser), nmr-posttrain (Pulse Program LLM Post-Training), topspincli (TopSpin CLI Automation), topspinRE (Pulse Program Reverse Engineering & AST), competition (Interpretable RNA Energy Model), xingling (Sci-Fi Portal), AIlove (AI Love Matching), spinachpython, pp2label, prodopy-wt
- Scientific & creative cloud platforms: NMR Pulse Program AI Platform, ccse Service, NMRagent Web Assistant, GraphPulse Timing Visualizer, Xingling Portal, AI Match
Honors & Competitions
- Excellence Award (优秀奖), 1st Global College Students Life Science Challenge, September 2026
- Track: North China Division · Graduate Track · Track 2: AI Gene Editing & Nucleic Acid Tool Design
- Project: AI-Driven Interpretable RNA Energy Model and Its Application in Nucleic Acid Tool Design
- Team: Jianning Kang (Lead), Zihang Zhou, Yaxuan Yuan, Yu Shi
- Core Achievements: Overcame limitations of classic Turner 2004 nearest-neighbor models in free energy precision and temperature dependence; designed
MotifCorrectionNet/Turner-Energy-AI, significantly improving near-degenerate decoy discrimination (60.34% → 80.28%), applied to rational design and experimental testing of temperature-responsive RNA switches/thermometers - Links: GitHub Repository · Official Site
Technical Skills
- Programming & Engineering: Python, Bash, Docker, Linux, Git
- AI & Machine Learning: LLM Fine-Tuning (SFT / DPO), Dataset Curation & Auditing, AI Agent Tools, Benchmark Design
- NMR Domain: Bruker TopSpin, Pulse Program Architecture, NMRSIM Simulation & Validation
Publications
-
NMR 脉冲程序生成模型与评测基准（撰写中）
周子航. (在研). NMR 脉冲程序生成模型与评测基准.
Open Source & Services
- Dedicated to long-term maintenance of NMR × AI scientific repositories, documentation, and user support