Yichen Di

Georgia Institute of Technology
Atlanta, GA, USA

About Me

I am Yichen Di, a Ph.D. student in Computer Science at Georgia Institute of Technology. I received my B.Eng. in Software Engineering from Tsinghua University. My primary research interests lie in AI for Science, with a current focus on AI for Computational Biology. Previously, I worked on neural operators for PDEs over complex geometries and LLM watermarking and robustness.

Previously, I worked as a research intern at the California Institute of Technology with Prof. Anima Anandkumar. I have also worked with Prof. Mingsheng Long, Prof. Lijie Wen, and Prof. Jie Tang at Tsinghua University.

Beyond research, I enjoy music, especially karaoke and piano. I also enjoy staying active through soccer, frisbee, and working out.

Education

B.Eng. in Software Engineering | Tsinghua University
Sept. 2022 – Jul. 2026
School of Software

Research

FG-ConvNO: A Geometry-Aware Neural Operator for Propeller CFD Prediction
Yichen Di, Valentin Duruisseaux, Di Zhou, Xinyi Li, Daniel Leibovici, Jean Kossaifi, Anima Anandkumar
[AAAI2026 AI2ASE Oral]

Factorized Geometry Convolutional Neural Operator (FG-ConvNO) is a geometry-aware neural operator for CFD prediction under complex geometry and limited data. It extends the FIGConv architecture with SDF conditioning, interpolation-based decoding, discrete-continuous convolutional blocks, and auxiliary global supervision. Trained on a newly generated propeller CFD dataset, it outperforms all evaluated baselines, achieving ~4–6% relative improvement.

Transolver++: An Accurate Neural Solver for PDEs on Million-Scale Geometries
Huakun Luo, Haixu Wu, Hang Zhou, Lanxiang Xing, Yichen Di, Jianmin Wang, Mingsheng Long
[ICML 2025]

Transolver++ is a highly parallel neural solver designed for large-scale PDEs, extending the input capacity of Transolver from millions to tens of millions of mesh points. By incorporating an adaptive local mechanism, it captures fine-grained local physics more effectively and achieves state-of-the-art performance on industrial-scale simulations.

[Paper] [Code]

WaterSeeker: Pioneering Efficient Detection of Watermarked Segments in Large Documents
Leyi Pan, Aiwei Liu, Yijian Lu, Zitian Gao, Yichen Di, Shiyu Huang, Lijie Wen, Irwin King, Philip S. Yu
[NAACL 2025, AAAI 2025 PDLM Oral]

WaterSeeker introduces an efficient framework for detecting watermarked text segments within large documents. Combining anomaly extraction with local traversal, it achieves a superior trade-off between detection accuracy and computational efficiency, enhancing the robustness of watermarking detection in real-world scenarios.

[Paper] [Code]

MarkLLM: An Open-Source Toolkit for LLM Watermarking
Contributor, Tsinghua University
[GitHub, 600+ stars]

MarkLLM is an open-source toolkit offering a unified and extensible framework for LLM watermarking algorithms. It integrates diverse methods into automated pipelines for comprehensive watermarking analysis, while also providing visualization tools that make token-level watermarking mechanisms transparent to researchers and practitioners.

[Paper]

Work Experience

Machine Learning Research Intern | DP Technology
Feb. 2026 - June 2026

Developed visual grounding pipelines to link scientific figures with textual entities and concepts, enabling structured extraction of multimodal scientific knowledge for downstream knowledge-base construction.

Summer Researcher | Anima AI + Science Lab, California Institute of Technology
June 2025 – Sept. 2025
Advisor: Prof. Anima Anandkumar

Investigated Neural Operators for PDEs over complex geometries and explored their applications to shape optimization problems in fluid dynamics.

×