Self-Evolving Agentic Engineering
Self-evolving engineering agents that plan experiments, accumulate verified knowledge, and continuously improve models and decisions across tasks.
HELLO, I'M
Building AI that understands the physical world — and improves how we engineer it.
My research explores AI for Science, Industrial AI, AI Agents, and Foundation Models, connecting computational intelligence with real-world engineering challenges.
I am interested in AI systems that represent physical processes, learn efficiently from scientific data, and support autonomous engineering workflows. These themes are closely connected rather than independent research silos.
Self-evolving engineering agents that plan experiments, accumulate verified knowledge, and continuously improve models and decisions across tasks.
Reusable representations and data-efficient learning across scientific tasks and physical systems.
Machine learning architectures that exploit physical structure for accurate, efficient simulation.
TL;DR A Gaussian-particle basis makes neural operators interpretable and resolution-agnostic, with efficient coupling in compact modal space.
The Gaussian Particle Operator (GPO) learns a compact and interpretable basis of Gaussian atoms for physical fields. A Petrov–Galerkin measurement and modal attention exchange information across scales before decoding predictions on spatial samples. The approach supports irregular geometries and achieves competitive accuracy with near-linear scaling for a fixed modal budget.
Repository; full implementation release noted as in preparation.
@inproceedings{li2026gpo,
title={From Basis to Basis: Gaussian Particle Representation for Interpretable PDE Operators},
author={Li, Zhihao and Feng, Yu and Lai, Zhilu and Wang, Wei},
booktitle={Proceedings of the 43rd International Conference on Machine Learning},
volume={306}, pages={66885--66902}, year={2026},
publisher={PMLR}
}TL;DR ReMD integrates multiscale residual correction and physics consistency inside the diffusion process for efficient fluid super-resolution.
ReMD tackles fluid super-resolution by correcting multigrid residuals during diffusion sampling rather than relying on unconstrained image super-resolution. Lightweight physical constraints and a multi-wavelet hierarchy promote coherent large-scale flows and fine vortical structure, while reducing sampling steps against diffusion baselines.
@inproceedings{li2026remd,
title={Physics-Consistent Diffusion for Efficient Fluid Super-Resolution via Multiscale Residual Correction},
author={Li, Zhihao and Dong, Shengwei and Yi, Chuang and Gao, Junxuan and Lai, Zhilu and Liu, Zhiqiang and Wang, Wei and Zhang, Guangtao},
booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
year={2026}
}TL;DR A multi-resolution neural operator combines hierarchical PDE representations and multiscale modeling for efficient dynamic simulations.
M2NO combines multi-resolution learning with multigrid-inspired structure to address multi-scale PDE dynamics. The method organizes computation across spatial scales, enabling efficient learning of physical operators under complex resolution and dynamics requirements.
@inproceedings{li2026m2no,
title={M2NO: An Efficient Multi-Resolution Operator Framework for Dynamic Multi-Scale PDE Solvers},
author={Li, Zhihao and Lai, Zhilu and Zhang, Xiaobo and Wang, Wei},
booktitle={Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining},
pages={771--782},year={2026},doi={10.1145/3770854.3780179}
}TL;DR Multi-graph neural operators fuse geometry, physical relations, and scale for PDE learning on arbitrary domains.
A multi-graph operator framework introduces scale-aware and physics-informed graph structure to learn solutions of PDEs over arbitrary geometries. Graph representations provide interaction paths suited to irregular domains, rather than requiring a single fixed grid.
@inproceedings{li2025amg,
title={Harnessing Scale and Physics: A Multi-Graph Neural Operator Framework for PDEs on Arbitrary Geometries},
author={Li, Zhihao and Song, Haoze and Xiao, Di and Lai, Zhilu and Wang, Wei},
booktitle={Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining},
pages={729--740},year={2025},doi={10.1145/3690624.3709173}
}TL;DR Reformulates neural operators on a d+1-dimensional domain, exposing the evolution of latent embeddings through structured auxiliary dynamics instead of simply increasing embedding width.
We study how neural operators evolve representations in an auxiliary embedding dimension. The method reformulates the learned operator on a d+1-dimensional product domain and uses a structured basis-diversified evolution module to improve expressive power and efficiency across PDE benchmarks.
@inproceedings{song2026reformulating,
title={Reformulating Neural Operators in d+1 Dimensions for Embedding Evolution},
author={Song, Haoze and Li, Zhihao and Zhang, Xiaobo and Gan, Zecheng and Lai, Zhilu and Wang, Wei},
booktitle={Advances in Neural Information Processing Systems},
year={2026}
}TL;DR A lightweight epistemic uncertainty framework that perturbs only the lifting stage of neural operators to obtain better calibrated, spatially faithful uncertainty estimates.
This work addresses epistemic uncertainty in neural operator PDE surrogates. It injects stochasticity only into the lifting module while keeping propagation and recovery deterministic, then calibrates predictive bands. Experiments on PDE benchmarks report stronger coverage, tighter intervals, and improved alignment of uncertainty with local residuals.
@article{song2026structureaware,
title={Structure-Aware Epistemic Uncertainty Quantification for Neural Operator PDE Surrogates},
author={Song, Haoze and Li, Zhihao and Deng, Mengyi and Li, Xin and Pan, Duyi and Lai, Zhilu and Wang, Wei},
journal={arXiv preprint arXiv:2603.11052},
year={2026}
}TL;DR Combines von Neumann–Mullins grain-area kinetics with a graph neural network for reconstructing evolving polycrystalline microstructures.
VN-GraphGG describes two-dimensional grain growth using a graph of grains and junctions. The von Neumann–Mullins law governs grain-area evolution, while a graph neural network reconstructs the microstructure. The separation supports changes to kinetic coefficients without retraining the reconstruction model, including cases with abnormal grain growth.
@article{huang2026vngraphgg,
title={VN-GraphGG: von Neumann--Mullins--based graph grain growth simulation},
author={Huang, Yuntong and Li, Zhihao and Dai, Shuyang and Ma, Chang and Xiang, Yang},
journal={arXiv preprint arXiv:2610.13418},
year={2026}
}TL;DR A learned preconditioner integrates neural operators with Krylov solvers to accelerate numerical PDE solution.
Neural Preconditioning Operator (NPO) employs neural operators trained using condition and residual losses as preconditioners for large sparse linear systems derived from PDEs. Combining algebraic multigrid principles and learned representations reduces solver iterations across Poisson, diffusion, and linear-elasticity problems.
@article{li2025npo,
title={Neural Preconditioning Operator for Efficient PDE Solves},
author={Li, Zhihao and Xiao, Di and Lai, Zhilu and Wang, Wei},
journal={arXiv preprint arXiv:2502.01337},year={2025}
}TL;DR A hyperbolic collaborative-filtering approach improves how recommendation models balance head and tail items.
Hyperbolic Informative Collaborative Filtering (HICF) studies how hyperbolic embeddings model the long-tailed user-item distributions of recommendation systems. An informative hyperbolic margin-ranking objective improves representation of both popular and rare items.
@inproceedings{yang2022hicf,
title={HICF: Hyperbolic Informative Collaborative Filtering},
author={Yang, Menglin and Li, Zhihao and Zhou, Min and Liu, Jiahong and King, Irwin},
booktitle={Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining},
pages={2212--2221},year={2022},doi={10.1145/3534678.3539475}
}TL;DR A broad review of hyperbolic graph representation learning, its geometries, methods, and applications.
This review synthesizes hyperbolic graph neural-network approaches and applications. It surveys the mathematical foundations of non-Euclidean graph representations, model families, and representative learning tasks.
@article{yang2022hyperbolic,
title={Hyperbolic Graph Neural Networks: A Review of Methods and Applications},
author={Yang, Menglin and Zhou, Min and Li, Zhihao and Liu, Jiahong and Pan, Lujia and Xiong, Hui and King, Irwin},
journal={arXiv preprint arXiv:2202.13852},year={2022}
}Reformulating Neural Operators in d+1 Dimensions for Embedding Evolution.
Research Scientist · AI for Science and Industrial AI.
From Basis to Basis: Gaussian Particle Representation for Interpretable PDE Operators.
Towards Scalable and Generalizable Neural PDE Solvers: Learning Across Scales, Geometries, and Physics
Physics-Consistent Diffusion for Efficient Fluid Super-Resolution via Multiscale Residual Correction.
Research Scientist
Research Intern
B.Eng. · Computer Science & Technology
Interested in research collaborations or scientific AI for engineering? I would be glad to connect.