Master of AI, Monash University · graduating Dec 2026莫纳什大学人工智能硕士在读 · 2026 年 12 月毕业

Beichen Li 李北辰

I trained as a chemist and now work in machine learning, on generative models for CJK type and on what interatomic potentials learn from their training labels.我本科学化学，现在做机器学习，研究中日韩字体的生成模型，以及原子间势从训练标签中学到了什么。

Applying for PhD positions · 2027 intake正在申请 2027 年入学的博士项目 Open to ML research & engineering roles同时寻找机器学习研究与工程岗位
About简介Melbourne · AEDT墨尔本 · 澳东夏令时

Chemistry, then machine learning.先学化学，后做机器学习。

I studied applied chemistry at Lanzhou University and am finishing a Master of Artificial Intelligence at Monash. In my research I fix the design and freeze the benchmark before I look at any results, and keep every run reproducible.我本科在兰州大学读应用化学，现在在莫纳什大学读人工智能硕士，2026 年 12 月毕业。做研究时，我在看结果之前定好实验设计、冻结基准测试，并保证每一次实验都能复现。

  • Generative models for typography字体生成模型CJK glyph generation and vector fonts中日韩字形生成、矢量字体生成
  • AI for ScienceAI for ScienceMachine-learned interatomic potentials and their labels机器学习原子间势及其训练标签
  • Rigorous evaluation of ML机器学习的严谨评估Controlled experiments, pre-registration, measurement validity对照实验、预注册研究、测量效度
  • Reliable AI systems可靠的 AI 系统LLMs behind rules, guardrails and fallbacks在规则、防护和兜底机制之下运行的大语言模型
U+9571 · 镱 · one character, two questions一个字，两个问题

Read it as type.把它当作字来读。

Its left half, 钅, is the simplified form of the metal radical. Japanese fonts write that radical as 金, so a Japanese font has never drawn 钅. Extending a font to characters built from parts it has never drawn is the problem behind my font study. In the study I create the gap on purpose: I remove one component family from a font's training glyphs and measure what that costs.

左半边的 钅 是简体的金字旁。日文字体把这个偏旁写作“金”，所以一款日文字体从没画过 钅。把字体扩展到由它没画过的部件组成的字，就是我的字体研究要解决的问题。实验里，我有意制造这种缺口：从字体的训练字形中删去一类部件，再测量这样做的代价。

Research研究Papers in preparation论文均在撰写中

Two studies, planned in advance.两项研究，方案先于结果。

Independent study独立研究 Aug 2026 – now2026.08 – 至今 Paper in preparation · replication under way论文撰写中 · 复现进行中

Component support gaps in CJK font generation中日韩字体生成中的未见部件问题

When a CJK font is extended beyond the few thousand characters it ships with, some new characters contain a the font never drew. I measure what that costs. Each target font is fine-tuned twice: once with every training glyph that contains a given component family removed, and once with the family kept. I then compare the generator's error on the same held-out characters. The design covers (小, 日, 竹, 米), each with 3 random seeds. It is crossed with a generator pretrained on 330 other fonts, with or without that family in the pool. A placebo control instead removes random unrelated characters from the pretraining pool. A replication on a second, architecturally different generator is under way.

把一款中日韩字体扩展到原有的几千字之外时，有些新字会包含字体从未画过的。我测量这件事的代价。每款目标字体微调两次：一次删去训练字形中含某类部件的全部字，一次保留，然后在同一批留出字上比较生成器的误差。实验覆盖（小、日、竹、米），每格 3 个随机种子；另与预训练条件交叉：生成器先在 330 款其他字体上预训练，预训练池含或不含这类部件。另设随机删除对照（即安慰剂对照）：从预训练池里随机删去与该部件无关的字。在另一种架构的生成器上的复现正在进行。

PyTorch · diffusers. From-scratch GAN (pix2pix) and flow-matching generators, trained on NVIDIA L4, L40S and A100 across Colab, AWS and Nectar Research Cloud.

PyTorch · diffusers。从零实现的 GAN（pix2pix）与流匹配生成器，在 Colab、AWS 和 Nectar 研究云的 NVIDIA L4、L40S、A100 上训练。

AI for Science Nov 2026 – now2026.11 – 至今 Analysis plan frozen Nov 2026分析计划已于 2026 年 11 月冻结 arXiv preprint planned Dec 2026计划 2026 年 12 月发布 arXiv 预印本 Code and data released with the submission (MIT / CC BY 4.0)代码与数据随投稿公开（MIT / CC BY 4.0）

Pseudopotential conventions in foundation interatomic potentials通用机器学习势如何继承赝势约定

Foundation interatomic potentials are trained on DFT labels, and the choices behind those labels are not fixed. In 2023 the Materials Project changed the pseudopotential it uses for from Yb_2 to Yb_3. Models of the same architecture (MACE, eqV2, Nequix) have been released twice: once trained on MPtrj, built before the change, and once on OMat24, which uses Yb_3. This gives a natural experiment. I ask whether the models inherit the convention of their labels, where in the network it is carried, and whether retraining only a small, ytterbium-specific set of weights can move a model from one convention to the other without changing its other elements.

通用机器学习原子间势（基础模型）用 DFT 计算的标签训练，而这些标签背后的计算设置并非一成不变。2023 年，Materials Project 把的赝势从 Yb_2 换成了 Yb_3。同一种架构（MACE、eqV2、Nequix）的模型都有两个公开版本：一个用更换之前构建的 MPtrj 训练，一个用采用 Yb_3 的 OMat24 训练。这构成了一个自然实验。我研究这些模型是否继承了训练标签的约定，约定由网络的哪一部分承载，以及只重训一小组镱专属的权重，能否把模型从一种约定移到另一种，而不改变其他元素的结果。

PyTorch · mace-torch · fairchem · Nequix. A fault-tolerant job queue runs training and relaxation jobs across cloud GPUs (A10, A16, T4) and a laptop RTX 4060, with cross-GPU reproducibility checks.

PyTorch · mace-torch · fairchem · Nequix。自写的容错任务队列在云端 GPU（A10、A16、T4）和笔记本 RTX 4060 之间调度训练与弛豫任务，并做跨 GPU 的可复现性检查。

BSc thesis · Lanzhou University · 2023本科论文 · 兰州大学 · 2023Synthesis and fluorescence of a neodymium 2-azido-4,4′-biphenyldicarboxylate complex. I did all the experimental work.2-叠氮基-4,4′-联苯二甲酸钕配合物的合成与荧光性质研究，全部实验工作独立完成。

Projects & open source项目与开源6 selected精选 6 项

Selected projects.精选项目。

LXGW WenKai500+ AI-generated glyphs across the family; the TC edition is on Google Fonts为整个霞鹜文楷系列补全 500 多个 AI 生成的字形，繁体版已收录于 Google FontsCNN · glyphs2023–24 ColorOS Font InjectorSystem fonts on locked, non-rooted phones, swapped inside a 30 ms window在未 Root 的手机上替换系统字体，在约 30 毫秒的窗口内完成Java · Android2026
EdibliFood-label scanner: rules decide every allergen verdict, a LoRA-tuned LLM only explains食品标签扫描应用：过敏原判断全部由规则做出，微调后的大模型只负责解释Node · Android2026
cross-harness-syncLets several AI coding agents hand off one research project across machines, with state kept in Git让多个 AI 编程智能体跨机器接力同一个研究项目，状态保存在 Git 中Python2026 fx-sentinelExchange-rate monitor: an Android app and a Telegram bot with threshold alerts and 7- or 30-day scenarios汇率监控：安卓应用和 Telegram 机器人，支持阈值提醒和 7 天或 30 天情景预测Flutter · Python2026 openfood-auOpen Food Facts pipeline for Australian products澳洲商品的 Open Food Facts 开放数据管线Python · SQLite2026
Experience & education经历与教育

Work and education.工作与教育。

2025 – Dec 20262025 – 2026.12

Master of Artificial Intelligence, Monash University人工智能硕士，莫纳什大学

Deep Learning · Fundamentals of AI · IT Research & Innovation Methods · Industry Experience Studio

深度学习 · 人工智能基础 · IT 研究与创新方法 · 行业实践项目

Apr – Nov 20252025.04 – 2025.11

IT & AI Solutions Assistant, New Asian SupermarketIT 与 AI 解决方案助理，New Asian Supermarket

Deployed an image-recognition model that replaced manual product entry; built dynamic pricing and supplier matching; ran the databases, tools, website and network.

部署了取代人工录入的商品图像识别模型；开发动态定价与供应商匹配系统；维护数据库、内部工具、公司网站和网络。

2019 – 2023

BSc in Applied Chemistry, Lanzhou University应用化学学士，兰州大学

Thesis on a neodymium complex, a lanthanide metal–organic framework.

毕业论文研究一种钕配合物（镧系金属–有机框架）。

Contact联系Melbourne墨尔本

Contact

me@beichen.li
GitHub ↗ LinkedIn ↗ CV (PDF) ↗简历 PDF ↗