MSc @ Dalian Maritime University (2027) · CFD + deep learning for flow-field prediction
I build small, inspectable tools at the intersection of physics and machine learning. Every project here runs offline with zero or one core dependency -- because a reviewer should be able to pip install -e . && pytest and see for themselves.
Surrogate modelling of flow in rough rock fractures (first-author paper under review)
- 616 OpenFOAM LES cases -> FNO + U-Net + AWF adaptive-fusion model
- RMSE 0.0045, Ux R² 0.9983, Uy R² 0.9970, 10.4x faster than CFD
- Funded by NSFC Youth Project 42407219
NTDR: point-cloud dimensionality reduction for joint-trace extraction (4th author, Rock and Soil Mechanics 2026, EI/CSCD)
- SVD-based projection + RGB rasterisation + Canny + DBSCAN
- 91% time saving vs manual, 90% overlap with ground truth
| project | what it does | stack |
|---|---|---|
| Fracture-Flow-Bench | JRC-calibrated aperture synthesis + Reynolds lubrication solver + cubic-law deviation analysis (the validation target sitting upstream of my FNO surrogate) | numpy |
| AeroForge-Agent | CFD multi-agent workflow: NL task -> parametric geometry -> OpenFOAM case generated from scratch -> real steady-RANS solve via a native/WSL runtime bridge, validated on the Ahmed-body benchmark (Cd vs experiment) | pydantic, asyncio |
| SmartPort-MultiAgent | Multi-agent container-terminal scheduling: berth / quay-crane / yard agents over an asyncio message bus + blackboard; FCFS vs NSGA-II vs MIP(CBC) vs HiGHS-exact + literature benchmarks (absorbed the archived Berth-Scheduler) | pydantic, numpy, matplotlib |
| SFT-Data-Auditor | Pluggable rule engine that audits LLM SFT training data and attributes each Bad Case to the rule that caught it | stdlib |
| FNO-Flow-Prediction | FNO vs UNet for parametric PDEs (1D Burgers): resolution-invariance, zero-dep core | numpy (+optional torch) |
| LLM-Eval-Harness | Lightweight LLM evaluation: multi-model comparison, LLM-as-Judge, Bad Case attribution | stdlib |
| AIGC-TestData-Synth | Spec-driven test-data synthesis with dedup (1-Gini evenness) and quality filtering | stdlib |
| Driving-Multimodal-Pipeline | Multimodal driving-load evaluation: AHP (geometric mean + CR test) + fuzzy synthesis | numpy |
| LLM-TestCase-Gen | LLM-powered unit-test generator that executes every case against the target function | stdlib |
| AI-Memory-Hub | Multi-platform AI memory sync + job tracking + knowledge-base maintenance | stdlib |
Python PyTorch OpenFOAM FNO U-Net numpy scipy matplotlib Linux Git
"Small, inspectable, runs offline." -- the constraint I design under.