Pu Ren

Research

My research aims to build principled and practical AI systems for science and engineering. I study AI models from three complementary perspectives: developing methods with appropriate inductive biases, empirically understanding model behavior, and applying these models to real-world scientific and engineering problems. My current research focuses on three directions.

Efficient and scalable neural physics simulators: developing neural PDE solvers, generative models, and scientific foundation models that incorporate physical structure for efficient simulation, prediction, and scientific discovery.

Weight-space analysis of AI models: empirically studying how physical priors, model geometry, and learned representations shape the generalization, transferability, and reliability of SciML models and LLMs.

Agentic AI for science and physical AI: building AI systems that can reason about physical systems, interact with real-world environments, design experiments, diagnose failures, and support auto-research and auto-engineering.

More info about me: I was a postdoc working with Prof. Michael Mahoney at Lawrence Berkeley National Lab. I received my Ph.D. from Northeastern University, where I was advised by Prof. Hao Sun and Prof. Ryan Wang. I was also fortunate to collaborate with Prof. Jerome F. Hajjar, Prof. Yang Liu, and Prof. Jian-Xun Wang.

News

[June 2026] TensorMesh is now online. TensorMesh is a fast, differentiable, ans user-friendly finite element library for PyTorch. It works for general scientific computing tasks, including physics-constrained machine learning, differentiable simulation, inverse problems, etc.
[May 2026] One paper has been accepted by KDD 2026 as an Oral presentation.
[Apr 2026] Three papers have been accepted by ICML 2026. IRNO has been selected as a Spotlight.
[Mar 2026] Our CGM-GM is published online by Nature Communications. We are also building a broader CGM family for geophysical applications: CGM-FAS and CGM-Wave.
[Sep 2025] Our CGM work is featured at SCEC2025 Plenary Talk.
[Apr 2025] I will be visiting UC Irvine to give a talk about our recent work on AI-enabled scientific modeling.
[Mar 2025] I will be visiting Rice University to give a talk about our recent work. Thanks for Hengrui's invitation!
[July 2024] I will co-organize a workshop at NeurIPS 2024 on Foundation Models for Science.
[May 2024] Our work on data-efficient pretraining and in-context learning for SciML has been accepted by NeurIPS 2024.
[Jan 2024] Our work on generative models for time series data has been accepted by ICLR 2024.
[June 2023] Our work on physics-encoded learning has been accepted by Nature Machine Intelligence.

Selected Publications

For more of my research, please see the full publications page.

Spectral Signatures of Large Language Models,
Z. Zhang*, I. V. Prasad*, Y. Hu, Z. Liu, H. Luo, P. Ren#, and Y. Yang,
The 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining (ACM SIGKDD), Oral, 2026.

Iterative Refinement Neural Operators are Learned Fixed-Point Solvers: A Principled Approach to Spectral Bias Mitigation,
X. Liu, S. Shang, X. Wang, P. Ren, and Y. Yang,
International Conference on Machine Learning (ICML), Spotlight, 2026. [arXiv] [Blog] [Code] [Slides]

Unveiling Multi-regime Patterns in SciML: Distinct Failure Modes and Regime-specific Optimization,
Y. Wang*, Y. Hu*, X. Zhong*, X. Wang*, H. Lu, T. Pang, M. W. Mahoney, Y. Yan, P. Ren#, and Y. Yang#,
International Conference on Machine Learning (ICML), 2026. [arXiv] [Code]

Learning earthquake ground motions via conditional generative modeling,
P. Ren, R. Nakata, M. Lacour, I. Naiman, N. Nakata, J. Song, Z. Bi, O. A. Malik, D. Morozov, O. Azencot, N. B. Erichson, and M. W. Mahoney,
Nature Communications, 2026. [DOI] [arXiv] [Code]

Model Balancing Helps Low-data Training and Fine-tuning,
Z. Liu*, Y. Hu*, T. Pang, Y. Zhou, P. Ren, and Y. Yang,
Empirical Methods in Natural Language Processing (EMNLP), Oral, 2024. [DOI] [arXiv] [Code]

Encoding Physics to Learn Reaction–Diffusion Processes,
C. Rao*, P. Ren*, Q. Wang, O. Buyukozturk, H. Sun, and Y. Liu,
Nature Machine Intelligence, 2023. [DOI] [arXiv] [Code]

PhyCRNet: Physics-informed Convolutional-Recurrent Network for Solving Spatiotemporal PDEs,
P. Ren, C. Rao, Y. Liu, J. X. Wang, and H. Sun,
Computer Methods in Applied Mechanics and Engineering, 2022. [DOI] [arXiv] [Code]

Selected Awards

[May 2021] Second Place of 2021 EMI Dynamics Committee Student Paper Competition
[May 2020] Third Place of 2020 EMI SHM&C Committee Student Paper Competition
[Sep 2019] Vilas Mujumdar Fellowship
[Oct 2018] China National Scholarship
[Apr 2015] Finalist of Mathematical Contest in Modeling (top 0.5%)

Invited Talks

[Apr 2025] Baking Physics into AI for Computational Modeling, UC Irvine.
[Mar 2025] Learning Dynamics from Sparse Data, Rice University.
[Feb 2025] Embedding Physics into Deep Learning for Spatiotemporal Systems, Zhongguancun Academy.
[Oct 2024] Generative AI for Earthquake Simulations, Berkeley Lab AI for Science Summit.
[Oct 2023] Learning Dynamics from Sparse Data, Lawrence Berkeley National Lab.
[Aug 2023] Encoding Physics to Learn Reaction-Diffusion Processes, Luoyi Science.
[June 2022] Embedding Physics into Deep Learning for Spatiotemporal Systems, Lawrence Berkeley National Lab.
[Mar 2022] Embedding Physics into Deep Learning for Spatiotemporal Systems, Argonne National Lab.
[Oct 2020] Physics-Reinforced Deep Learning for Structural Metamodeling, Northeastern University.
[Mar 2020] Bayesian Tensor Learning for Structural Monitoring Data Imputation and Response Forecasting, Northeastern University.

Services

Area Chair/Reviewer for AI Conferences: ICML, NeurIPS, ICLR, IJCAI, AAAI, AISTATS, NLDL, etc.

Reviewer for Journals: Nature Communications, Journal of Machine Learning Research, SIAM Review, SMAI Journal of Computational Mathematics, Computer Methods in Applied Mechanics and Engineering, IEEE Transactions on Neural Networks and Learning Systems, IEEE Transactions on Geoscience and Remote Sensing, Mechanical Systems and Signal Processing, and many others.

Teaching

[Sep 2025 - May 2026], Capstone Project, UC Berkeley.
[Sep 2022 - May 2023], Capstone Project, UC Berkeley.
[Sep 2021 - Dec 2021], Structural Dynamics, Northeastern University.