I am a Ph.D. candidate in AI for Science at The University of Manchester, with expected completion in March 2027. My research brings together physics-guided deep learning, generative models, and scientific discovery agents to understand and simulate urban climate.
Alongside my research, I work on reliable LLM agents and model post-training. As a Qingyun Program intern at Tencent CSIG, I developed verification and refinement methods for DataBuddy and worked on tool-using SQL agents. Previously, I conducted research at Shanghai AI Laboratory on physics-guided urban climate modeling, leading to a first-author paper at NeurIPS 2025.
Open to full-time opportunities starting in 2027. I am actively seeking research and engineering roles in AI for Science, LLM agents, and model post-training. Please contact me at jiyang.xia@manchester.ac.uk to discuss opportunities.
🔥 News
- 2026: 🏅 Gold medal in the Kaggle Hull Tactical – Market Prediction competition.
- 2026.09: Completed my Qingyun Program internship with the DataBuddy Agent Algorithms team at Tencent CSIG.
- 2026.08: Our paper on automated LCZ mapping and thermal disparities was published in Urban Climate.
- 2026.05: Received Google Cloud Research Credits to support agentic urban climate research.
- 2025.10: Awarded the NeurIPS 2025 Scholar Award; my first-author paper on physics-guided urban climate modeling was accepted to NeurIPS 2025.
- 2025.07: Presented at the 12th International Conference on Urban Climate, in the session on Digital Twins, AI/ML, and Open Data Science for Urban Climate.
- 2025: Awarded the Doctoral Training and Development Support Fund by The University of Manchester.
- 2025.01: Appointed as a Graduate Teaching Assistant for The Natural Scientists Toolkit course (2024–2025) at The University of Manchester.
- 2024.11: My academic homepage is online!
🗂 Selected Projects
UrbanSurface-Agent
2026 · Scientific discovery agents
An autonomous system that combines LLMs with the SUEWS physical simulator to diagnose urban climate model errors. It proposes physical hypotheses, tests them through bounded parameter interventions, and uses a physics-based judge to distinguish meaningful improvements from compensating errors. Scenario-conditioned memory helps subsequent experiments learn from failed hypotheses. Evaluated across 20 real urban sites; the associated manuscript is under review.
Remote Sensing + AutoML for Urban Climate Modeling
2025–2026 · Urban Climate 2026
A framework that combines multi-source remote sensing with AutoML to map urban surface characteristics at 100 m resolution. The resulting Local Climate Zone maps improve classification accuracy over the WUDAPT baseline and inform CLM5 heatwave simulations. Validation against 37 sensors reduced temperature RMSE from 0.98 K to 0.90 K, revealing heat exposure missed by conventional surface representations.
Urban Climate RAG
2026 · Retrieval and evidence-grounded generation
A domain-specific knowledge system with two-stage retrieval: vector search followed by cross-encoder reranking. Answers are linked to source passages for traceable evidence. The system combines Qwen inference and embedding models with Ollama, FastAPI, Docker, and Open WebUI.
Physics-Guided Urban Climate Emulation
2024–2025 · NeurIPS 2025
A multitask Transformer that models urban surface–atmosphere interactions using physical inductive biases. The model achieves 422× faster simulation than the numerical baseline, improves prediction and cross-city generalization, and supports fine-tuning on real observations. Follow-up research has received NVIDIA academic support.
Physics-Aware Diffusion for Urban Climate Adaptation
2025–2026 · Ongoing research
A generative modeling project that combines physical constraints with remote sensing imagery, meteorological time series, and socioeconomic features. The work explores urban climate adaptation scenarios and a spatially neutral evaluation benchmark, with physical consistency checks to assess generated results.
📝 Publications
Published papers
- Jiyang Xia, et al. Optimized LCZ mapping with Automated Machine Learning Reveals Thermal Disparities during a Heatwave Event. Urban Climate, 2026.
- Jiyang Xia, et al. Learning Urban Climate Dynamics via Physics-Guided Urban Surface–Atmosphere Interactions. NeurIPS, 2025.
- Jiyang Xia, Yue Zhang*. Where Are Potential Areas for Transit-Oriented Development (TOD)—Exploring the Demands for Built Environment for TOD Planning. Sustainability, 2022, 14, 8364.
- Jiyang Xia, Yue Zhang. *Research on Spatial Spillover Effect of Highly Educated Migrants in Yangtze River Delta Urban Agglomerations Based on Urban Innovation. Areal Research and Development, 2022, 1, 162–168.
Manuscripts under review
- Jiyang Xia, et al. When SQL Correctness is Not Enough: Verifying Contract-to-Delivery Conformance in Data Agents. Submitted to ICLR 2027.
- Yu Zhang, Jiyang Xia, et al. UrbanSurface-Agent: Physics-Gated Scientific Discovery for Urban Surface Parameterization. Submitted to ICLR 2027.
- Yu Zhang, Jiyang Xia, et al. LCZFormer: Learning Factorized Urban Surface–Climate Representations for Cross-City Urban Climate Emulation. Submitted to KDD 2027.
🎖 Honors and Awards
- 2026: Gold Medal, Kaggle Hull Tactical – Market Prediction
- 2026: Google Cloud Research Credits
- 2025: NeurIPS Scholar Award
- 2025: Doctoral Training and Development Support Fund, The University of Manchester
- 2024: SoNS Research Training Support Grant, The University of Manchester
- 2023: Excellent Postgraduate Students’ Award of Zhejiang Province (Top 1%)
- 2022: Sophia Yang Scholarship, Zhejiang University (Top 1%)
- 2022: Outstanding Postgraduate Award and “Triple-A” Student Award, Zhejiang University
- 2020: Postgraduate Academic Scholarship, Zhejiang University
- 2018: National “Outstanding Work” Award, National Urban Transportation Innovation Practice Competition
📖 Education
- 2024.01–2027.03 (expected): The University of Manchester, UK. Ph.D. candidate in AI for Science. Research interests: AI for climate science, representation learning, physics-guided deep learning, and scientific agents.
- 2020.09–2023.06: Zhejiang University, China. M.Sc. in Urban Planning, with a focus on spatial big data analysis. GPA: 3.88/4.0; rank: 1/31.
- 2015.09–2020.06: Southwest Jiaotong University, China. B.Eng. in Urban and Rural Planning.
💬 Talks and Presentations
- 2025.07: 12th International Conference on Urban Climate, session on Digital Twins, AI/ML, and Open Data Science for Urban Climate, Rotterdam, Netherlands.
- 2024.08: NERC’s Digital Gathering 2024, Manchester, UK.
- 2022.07: 16th International Association for China Planning (IACP) Annual Conference, session on Transit-Oriented Development, Wuhan, China.
💻 Experience
Tencent CSIG · Qingyun Program Intern
2026.05–2026.09 · DataBuddy Agent Algorithms Team
- Developed lightweight SQL verification and structured feedback to support automatic refinement within the agent loop, integrated into the DataBuddy production workflow.
- Designed asynchronous end-to-end verification to check whether an agent’s final delivery satisfies the user’s requirements, combining execution evidence, reconstructed contracts, deterministic checks, and LLM judgment.
- Built agentic trajectory datasets and feedback loops for SQL subagent post-training, using full supervised fine-tuning and LoRA with DeepSpeed.
Shanghai AI Laboratory · Research Intern
2024.03–2024.09 · AI for Earth
- Developed a physics-guided Transformer for urban climate simulation, spanning problem formulation, data pipelines, model design, and experimental validation. This work led to a first-author NeurIPS 2025 paper.
- Contributed to urban climate agents that combine site characteristics, physical constraints, planning, tool use, and code execution for automated modeling experiments.
China Southwest Architectural Design and Research Institute · Assistant City Planner
2020.06–2020.08