刘子提 | Ziti Liu

刘子提 | Ziti Liu

AI for Science, Science for AI

About me

👋 Hi, I’m Ziti Liu — a Ph.D. student (combined Master-Ph.D. program, currently in my 4th year) in Computer Science and Technology at the School of Advanced Interdisciplinary Sciences, University of Chinese Academy of Sciences, co-advised by the Institute of Automation, the Institute of Mechanics and TsingHua University. My research focuses on AI for Science, including both traditional Physics-Informed Neural Networks (PINNs) and the application of Large Language Models (LLMs) in neural network architecture design for scientific problems, especially for PDE solving.

🎓 Education

  • (2022.9 – 2027.6) Ph.D. in Computer Science and Technology, University of Chinese Academy of Sciences, Research direction: AI for Science
  • (2018.9 – 2022.6) B.Eng. in Aircraft Design and Engineering, Huazhong University of Science and Technology , Outstanding Graduate

🛠 Skills & Interests

  • Programming: Python, C++, MATLAB
  • Machine Learning: PyTorch, PINNs, llama, AI for PDEs, scientific computing
  • Languages: English (IELTS 7.0), Mandarin (Native)
  • Others: LaTeX, numerical analysis, data structures, algorithms
  • Hobbies: Fitness, swimming, basketball, piano, photography

🧑‍💻 Projects

  • Generative Physics Engine Scenario Agent (Ongoing)

    • Project: Integrates generative language models to address accuracy issues in physical data inversion with traditional large models.
    • Work: Developed a physics-agent framework based on LLM-SR, significantly reducing the time required by traditional symbolic regression methods. Embedded a reasoning evaluator to enhance LLM’s answer filtering and correct reasoning abilities, improving inference accuracy and stability in physical scenarios.

  • Structured Physics-Informed Neural Networks via Knowledge Distillation

    • Project: Utilized knowledge distillation algorithms to embed physical features into neural network structures, addressing interpretability and applicability challenges.
    • Work & Results: Built a PyTorch-based experimental platform, overcoming manual model construction and strong prior dependency. Developed the gray-box model ψ-NN with higher accuracy, stability, and interpretability (algorithm performance improved by ~90%), reducing computational resource consumption to 7% of the original, suitable for low-power sensor group algorithms.

  • Adaptive Symmetry-Recomposition Physics-Informed Neural Networks

    • Project: Leveraged Lie symmetry properties of PDEs to enhance neural network robustness and accuracy under sparse sampling.
    • Work & Results: Established a PyTorch-based platform, demonstrating high efficiency and accuracy in handling symmetric PDEs and ill-posed sampling. Enhanced PINNs for sparse sensor arrays (sampling density reduced by 50%) and high-precision computation on low-power devices (accuracy improved by 90%).

📚 Publications