Li's Flood Lab · University of Colorado Boulder
Flood science, from the raindrop to society.
We advance the scientific understanding and practical mitigation of floods through remote sensing, physics-based modeling, and AI — delivering accurate, timely information for the good of the community.
The mission
We study surface water across scales — local, continental, and global — to predict and monitor floods before they arrive.
Floods are the costliest natural hazard on Earth, and they are growing flashier and less predictable. Our lab couples remote-sensing platforms with AI-integrated hydrologic–hydraulic models to close the gap between raw observations and actionable warning.
The research framework
SURGE — five letters, four thrusts, one pipeline.
Open models
From rain to flood.
Our models transcend catchment-scale hydrology to reach hyper-local hydraulics — turning water into flood. From the CREST lineage to the GPU-native Inunda, the Flood Lab is one of the core contributors to each. “All models are wrong, but some are useful.”
Open data
Four open flood datasets.
From a century of observed U.S. floods to simulation-grade benchmarks for AI — data the community can build on.
U.S. Flood Database (USFD)
Remote sensing, stream gauges, flood reports, and crowdsourcing harmonized into one of the most comprehensive flood records in the United States.
A benchmark-ready flash-flood dataset for the contiguous U.S.
42,466 NOAA flash-flood episodes (1996–2025) traced to gauges and watersheds: observed USGS hydrographs, high-water marks, 2-minute MRMS rainfall and terrain, with a 21-testbed primary benchmark for model intercomparison.
A high-resolution physical flood inundation benchmark for AI
Physics-based, simulation-grade flood inundation truth for training and evaluating the next generation of AI flood models.
An event-by-event atlas of U.S. floods
Dated flood extents from satellites and drones, crowd and storm reports, rainfall and USGS gauge records for each flood, checked layer by layer on an interactive map with a timeline.
Agent benchmarks
Can frontier AI run a flood model?
We build the evaluations that answer it. Two benchmarks test LLM agents on real hydrologic modeling — wrangling data, configuring simulators, calibrating, and verifying results end to end.
Agent's Last Exam
An expert-level exam for scientific agents: end-to-end hydrologic modeling tasks that demand reasoning, tool use, and domain judgment.
terminal-bench-science
Real scientific computing in real terminals — agents drive models from raw data to verified simulation, graded automatically.
Yan, S., Chen, M., Li, Z., Wen, Y., Zhu, S., Zhang, M., Liu, D., Cao, J., Chen, X., Deng, C., Yang, T., & Hong, Y. (2026). AI Agent for Hydrologic Modeling: Definition, Development, and Application. Geophysical Research Letters, 53(13), e2025GL119814. doi:10.1029/2025GL119814
Yan, S., Li, Z., Zhu, S., Wen, Y., Zhang, M., Chen, M., et al. (2025). AQUAH: Automatic Quantification and Unified Agent in Hydrology. Proceedings of the IEEE/CVF International Conference on Computer Vision, 2926–2935.
Open initiative from the lab
Does your AI hydrologic model conserve mass, energy and momentum?
HydroTuring asks one question of any AI hydrologic model — not whether it reproduces a hydrograph, but whether the physics it claims still holds. It is built in the open, and it grows only as fast as the physics people bring to it.
One question, asked of every model
A model passes only when every criterion of every probe passes: its water budget closes, its energy budget closes, its routing conserves momentum. The verdict is one bit, and everything beneath it stays quantitative — so a model that leaks 6% and one that leaks 40% are told apart long before either crosses the line.
A benchmark that only checks closure is trivially gamed, so deliberately broken models live in the repository to prove it. No probe is merged until it passes an exactly conservative model and catches the broken ones it names.
Failing is not one thing either. A model that reported its budget and did not close it has violated conservation; one that never reported enough to be checked has only declined to be falsifiable — and that is where nearly every published rainfall–runoff model sits today.
Bring a probe or a model
Thirty-seven probes are merged today — twenty-four mass, seven energy, six momentum — and the roadmap lists three more directions, one still unclaimed. If you have spent time with a conservation law that AI models get wrong, that law is a probe, and we would rather have it from you than approximate it ourselves.
Every published rainfall–runoff model, every LSTM, every foundation model with a hydrologic claim is in scope, and a model that fails is worth proposing. You do not need to be able to package it yourself — say so on the form and the packaging is assigned to someone who can.
One merged probe earns co-authorship on the benchmark paper. So do five accepted model proposals.
Three steps, for a probe or a model
- 1Open an issue. Propose it first — whether a probe discriminates is cheaper to learn in a paragraph than in three hundred lines.
- 2Fork the repository. Once the proposal is labelled accepted it is assigned to you, and nobody else will build it twice.
- 3Open a pull request. It comes from your fork and closes the issue.
Come talk it over first
Questions, probe ideas, a model you would like evaluated, a tolerance you think is wrong — the HydroTuring Discord is where that conversation happens, before and alongside the issues. You do not need to be invited, affiliated, or known to us.
HydroTuring: A physics probe benchmark testing mass, energy, and momentum conservation in hydrologic models
A conceptual model that loses 15% of its water to an unreported sink still reaches a Nash–Sutcliffe efficiency of 0.93, so hydrograph skill alone cannot show that a model conserves anything. The preprint introduces HydroTuring v0.1.0: 37 black-box probes, validated against 18 exact and 59 deliberately broken reference models, and runs them on 32 models. Among the 28 scored on five or more probes, only the three exact references pass every one. Process-based and AI models pass similar shares of their scored probes (76% and 74%), and 49% of the 1,184 model–probe pairs cannot be scored at all, because the models do not report the states and fluxes the test needs.
Li, Z., Li, C., Shams, S., Wang, H., Yang, Q., Liu, Y., et al. (2026). HydroTuring: A physics probe benchmark testing mass, energy, and momentum conservation in hydrologic models. arXiv:2611.11563.
The team
People
Principal Investigator
Dr. Zhi Li (“Dr. Z”)
Dr. Z is an Assistant Professor at the University of Colorado Boulder. Previously he was the Stanford Doerr School of Sustainability Dean's Postdoctoral Fellow (2023–2025). He earned his PhD (2022) from the University of Oklahoma, with Master's and Bachelor's degrees from the National University of Singapore (2019) and Hohai University (2017).
He studies surface water across scales — spatially (local, continental, global) and temporally (hydrology, hydrometeorology, hydroclimatology) — focusing on flood prediction and monitoring with remote-sensing platforms and AI-integrated hydrologic–hydraulic models. He has published 50+ peer-reviewed articles and one invited book chapter. He is an Editor for Journal of Hydrology X and an Associate Editor of the Journal of Applied Meteorology and Climatology, and serves as a reviewer for 10+ international journals.
Team members
Ke Zhu
Ke Zhu is a doctoral student at the University of Colorado Boulder. He earned his Master's degree (2025) from Southeast University (Civil & Hydraulic Engineering), directed by Prof. Qian Zhu, and his Bachelor's (2022) from Hohai University (Water Science & Engineering). His prior research focused on improving hydrologic models; he now studies the application of foundation models to flood identification.
Katie Miller
Katie Miller is a doctoral student at the University of Colorado Boulder. Her research focuses on using remote sensing to assess socio-environmental vulnerability to flooding. Prior to her PhD, she served as a Lead with NASA's DEVELOP Program, where she led interdisciplinary projects using NASA Earth observations to support environmental decision-making. Katie holds a bachelor's degree in Political Science (UC Berkeley) and master's degrees in Supply Chain Management (Erasmus University) and Sustainability Science (University of Gävle).
Chris Leong
Chris Leong is a graduate student at the University of Colorado Boulder. His research focuses on enhancing remote sensing products for flood inundation and wetland dynamics under drought–flood conditions. He earned his Bachelor's degree (2024) in Earth Systems Science (ESS) from the University of California, Irvine, where he collaborated in research pertaining to atmospheric chemistry and hydrology.
Collaborators
Our work is built with wonderful collaborators, including Prof. Yang Hong (Univ. of Oklahoma), Jonathan J. Gourley (NOAA/NSSL), Steven Gorelick (Stanford), Pierre Kirstetter, Mengye Chen, Shang Gao, Yixin Wen, Siyu Zhu, Songkun Yan, Guoqiang Tang, and Tiantian Yang.
In the classroom
Teaching
Bridging environmental engineering, data science, and AI for the Earth system.
Hydraulic Engineering
An undergraduate course on hydraulic engineering theory and practice: incompressible flow in conduits, pipe system analysis, open channel flow, flow measurement, and hydraulic machinery — built around four labs and a semester-long water distribution design project.
AI for Earth System Science & Engineering
A graduate course (AI4ESSE) on machine learning for hydrology, remote sensing, and the geosciences: from data pipelines and physics-informed models to responsible deployment on real Earth-system problems.
CVEN 3323 · Fall 2026 · CU Boulder
Hydraulic Engineering
Course Overview
Course Information
- Course Number: CVEN-3323
- Course Title: Hydraulic Engineering
- Credits: 3
- Term: Fall 2026
Meeting Times
- Days: T, Th
- Time: 9:30–10:45 AM
- Location: ECCR 245
Course Description
This course studies hydraulic engineering theory and its applications. Topics include incompressible flow in conduits, pipe system analysis and design, open channel flow, flow measurement, and the analysis and design of hydraulic machinery. Lecture material is reinforced by four laboratory experiments and a semester-long project in which students design a water distribution system for a small community using EPANET.
Learning Objectives
Upon successful completion of this course, students will be able to:
- Apply fluid properties and hydrostatic principles to engineering calculations
- Quantify measurement uncertainty using error propagation and the multiple-measurements method
- Apply conservation of mass, energy, and momentum to pipe flow and open channel flow
- Compute friction and minor losses, and solve Type I, II, and III pipe problems
- Estimate water demands and size storage for a distribution system
- Analyze pipe networks by hand and with EPANET
- Select and configure pumps using system and pump characteristic curves
- Analyze uniform and gradually varied open channel flow, and classify flow regimes
- Measure flow using weirs and analyze hydraulic jumps
- Design a water distribution system that meets stated design criteria
- Conduct experiments and communicate results as professional engineering reports
Instructor Information
Instructor
- Name: Zhi Li
- Email: Zhi.Li-2@colorado.edu
- Office: SEEC N151
- Office Hours: Tuesday & Thursday, 11:30 AM – 1:00 PM (Zoom — link on Canvas)
Teaching Assistant
- Name: Ke Zhu
- Email: Ke.Zhu@colorado.edu
- Office Hours: Monday, 10:00 AM – 12:00 PM, ECOT 448 (no office hours 9/7 — Labor Day)
Lab Assistant
- Name: TBD
To contact the professor or TA outside of class and office hours, please use Canvas Mail. Email sent directly to campus addresses may not be answered. Weekday email receives a response within 24 hours; email is not answered in the evenings or on weekends.
Prerequisites
- CVEN 3313 Theoretical Fluid Mechanics, its prerequisites, and their prerequisites
- Fluid mechanics (reviewed in class, but a general understanding is expected): density, specific weight, and specific gravity; hydrostatic head, pressure, and force; pressure, elevation, and velocity head; the Bernoulli principle; conservation of mass
- General Physics 1: magnitude and direction of a vector; resolving a vector into x- and y-components; Newton’s second law; free-body diagrams
- Calculus: partial derivatives; derivatives of polynomials, products, and quotients
Proposed Course Topics
| No. | Main Topic | Items |
|---|---|---|
| 1 | Fluid Properties and Measurement Uncertainty |
|
| 2 | Hydrostatics |
|
| 3 | Conservation Principles in Pipe Flow |
|
| 4 | Friction and Energy Losses in Pipes |
|
| 5 | Water Demands and Storage |
|
| 6 | Pipe Systems and Networks |
|
| 7 | Pumps and Hydraulic Machinery |
|
| 8 | Open Channel Flow |
|
| 9 | Flow Measurement and Varied Flow |
|
| 10 | Water Distribution System Design |
|
Assignments
Concept Checks
Eight Canvas quizzes of roughly eight multiple-choice questions each, testing conceptual understanding of the previous week’s material. The highest six scores count toward the final grade, and you have three attempts at each. Released Thursdays at 5 PM and due Tuesdays at 5 PM. Concept Checks must be done individually. Some include FE (Fundamentals of Engineering) Exam practice questions.
Homework Assignments
Six assignments covering 23 problems, worth 20 points each. Homework is due Thursdays at 5 PM and submitted on Canvas as a PDF. The homework grade is calculated out of 414 of the 460 possible points, so earning 90% of the points earns full credit; any points above 414 are extra credit.
Problems must be written as professional engineering calculations: sketch, equation in symbolic form, parameter values with units, stated assumptions, equation with numbers substituted in, units carried through every step, and a circled final answer with appropriate significant figures. Solutions that do not carry and cancel units in every step receive no credit. You may discuss assignments with classmates, but what you submit must be your own work.
Laboratory Experiments
All students must be registered for a lab section. Four experiments run through the semester; reports may be written individually, with the full lab group, or with any subset of it.
| Lab | Experiment | Report Due |
|---|---|---|
| Lab 1 | Boulder Creek | Tuesday, September 15, 5:00 PM |
| Lab 2 | Measurement Error | Tuesday, September 8, 5:00 PM |
| Lab 3 | Friction Losses in Pipes | Tuesday, October 6, 5:00 PM |
| Lab 4 | Weirs | Tuesday, December 1, 5:00 PM |
Design Project
All students design a water distribution system for a small community, in two parts:
- Default Pipe Network — three assignments (demands and tank volume, tank elevations, pump calculations), done individually or in groups of up to three. This network will not meet every design criterion, but it lets you complete the necessary calculations and get feedback before the final design.
- Final Design — modify the pipe diameters of the default network so the design meets all criteria. Done individually or in groups of up to six. The final report is due Monday, December 7 at 4:00 PM on Canvas.
Exams
Both midterms are held in person, in the regular classroom during class time. You will be provided with the Textbook Resources booklet; no other materials are permitted, though you may use a calculator. Exam dates are fixed even if the content shifts.
- Exam 1: Thursday, October 15 — Chapters 2, 3, and 4
- Exam 2: Thursday, November 19 — Chapters 5, 6, 8, and 9
Grading Policy
| Component | Weight |
|---|---|
| Concept Checks (8 total, lowest dropped) | 7% |
| Homework Assignments | 10% |
| Midterm Exams (2 total, 20% each) | 40% |
| Design Project — Default Design | 15% |
| Design Project — Final Design | 16% |
| Laboratory Reports (4 total, 3% each) | 12% |
Grading Scale
Grades are assigned on the scale below. There is no curve.
- A: ≥ 92% · A−: 90–92%
- B+: 88–90% · B: 82–88% · B−: 80–82%
- C+: 78–80% · C: 72–78% · C−: 70–72%
- D+: 68–70% · D: 60–68%
- F: < 60%
All grades are posted on Canvas. Questions about grading must be submitted to the Canvas assignment “Grading Questions” within one week of the grade being posted, with the problem clearly identified and a written explanation; all grades are final one week after posting.
Course Policies
Late Submissions
- Homework: half credit if received by 8 AM on the Tuesday immediately after the due date; not graded after that, unless special arrangements are made with Prof. Li.
- Lab reports: penalized 25 points per day.
- Default Design assignments: half credit if submitted by the assignment closure date on Canvas; not accepted after that.
- Final Design report: penalized 25% per day, including weekend days.
Use of AI
Everything needed to complete the assignments is provided on Canvas: the Textbook Resources document, solutions to in-class examples, lecture recordings, lab documents, and project documents. Some problems deliberately ask you to apply concepts in ways you have not seen before — an important part of engineering training. You are encouraged to avoid using AI on all assignments in this course. If you do use it, you are required to identify which parts of the assignment were completed with AI and to explain why you needed it instead of the resources provided for the class.
Academic Integrity
Following the expectations of the engineering profession — see the ASCE Code of Ethics and the NSPE Code of Ethics — any assignment you submit must be entirely your own work. On group assignments, everyone named must have contributed, and everyone who contributed must be named or acknowledged. All students must sign the Academic Integrity Policy on Canvas before submitting any assignment.
You are permitted to get help from the professor and TAs, discuss assignments with classmates, and consult the textbook, Canvas resources, and other textbooks. You are not permitted to copy another person’s work, share your work for someone else to submit, use solutions manuals, use or post to sites such as Chegg and CourseHero, or give or receive help on exams. Anything not listed is not permitted unless you are told otherwise. Suspected violations go to the Honor Board and may carry sanctions from a zero on the assignment to an F in the course.
Accommodations
Students who need exam accommodations should provide official documentation to Prof. Li at least two weeks before the exam so there is time to arrange them. If you must miss an exam for a legitimate reason, such as a religious holiday or job interview, make arrangements with Prof. Li beforehand.
Resources
Textbook
- Fundamentals of Hydraulic Engineering Systems, 5th edition, R.J. Houghtalen, A.O. Akan, and N.H.C. Hwang, Pearson, 2017 (ISBN-13: 978-0-13-429238-0)
- Textbook Resources document on Canvas — the key equations, tables, and figures, and the reference you are given during exams
- FE Reference Handbook (relevant sections posted on Canvas)
Software/Tools
- EPANET — free water distribution modeling software from the US EPA, used throughout the design project
- Bechtel lab (ECCE 157/161) — EPANET is installed on the lab machines; Buff card access is requested by email as described in the syllabus on Canvas
- Apporto Civil Engineering Desktop — browser access to the Bechtel lab software, linked from Canvas
CVEN 5833 · Spring 2026 · CU Boulder
AI for Earth System Science & Engineering
Course Overview
Course Information
- Course Number: CVEN-5833
- Course Title: AI in Earth System Science and Engineering
- Credits: 3
- Term: Spring 2026
Meeting Times
- Days: TTH
- Time: 8:30-9:45
- Location: SEEC N124
Course Description
This course explores the application of artificial intelligence and machine learning techniques to problems in Earth System Science. Students will learn how modern AI methods can be applied to analyze, model, and predict complex Earth system processes including climate dynamics, hydrology, atmospheric science, and environmental monitoring.
Learning Objectives
Upon successful completion of this course, students will be able to:
- Understand fundamental AI/ML concepts relevant to Earth system science
- Apply machine learning techniques to analyze Earth observation data
- Develop predictive models for Earth system processes
- Evaluate model performance and interpret results in scientific context
- Critically assess the application of AI methods to environmental challenges
Instructor Information
Instructor
- Name: Zhi Li
- Email: Zhi.Li-2@colorado.edu
- Office: SEEC N104
- Office Hours: TBD
Teaching Assistant
- Name: TBD
- Email: TBD
- Office Hours: TBD
Prerequisites
- Linear Algebra
- Calculus
- Programming (Python)
Proposed Course Topics
| No. | Main Topic | Items |
|---|---|---|
| 1 | Introduction to Programming Language and AI & Earth System Data |
|
| 2 | Supervised Learning and Regression |
|
| 3 | Classification and Non-Linearity |
|
| 4 | Model-Based and Non-Parametric Methods |
|
| 5 | Deep Learning Fundamentals |
|
| 6 | Convolutional Neural Networks (CNNs) for Spatial Data |
|
| 7 | Recurrent and Sequence Models |
|
| 8 | Graph and Generative Models |
|
| 9 | Advanced Sequence Modeling |
|
| 10 | Foundation Models in Practice |
|
| 11 | Societal Impact & Ethics |
|
Assignments
Assignment 1: Data Analysis
Due Date: TBD
Assignment 2: TBD
Due Date: TBD
TBD
Assignment 3: TBD
Due Date: TBD
TBD
Final Project
Due Date:
Comprehensive project applying AI methods to an Earth system science problem...
Grading Policy
| Component | Weight |
|---|---|
| Assignments | 40% |
| Midterm Exam | 25% |
| Final Project | 30% |
| Participation | 5% |
Grading Scale
- A: 90-100%
- B: 80-89%
- C: 70-79%
- D: 60-69%
- F: <60%
Course Policies
Late Submissions
Assignments submitted after the due date will be penalized 10% per day late, up to a maximum of 3 days. After 3 days, assignments will not be accepted without prior approval from the instructor.
Academic Integrity
All work submitted must be your own. Collaboration on assignments is allowed up to the point of sharing code or solutions. Any violation of academic integrity will be reported to the Honor Code Council.
Accommodations
Students with disabilities who need accommodations should contact Disability Services and inform the instructor as early as possible in the semester.
Resources
Online books
Youtube videos
Software/Tools
- 100 Page Python Intro
- Python Data Science Handbook
- Deep Learning with PyTorch
- Jupyter Notebook Tutorial: Introduction, Setup, and Walkthrough
Statistics
Peer-reviewed · 50+ articles
Publications
The full peer-reviewed list — newest first, with the lab author highlighted. The citation trajectory below is tracked via Google Scholar.
Science stories
Gallery
Gallery of the model simulation and scientific communication.
Nov 4, 2026
Third Storm, Highest Crest: How Wet Ground Set Up Pagosa Springs' Biggest Flood Since 1911
On November 1, 2026 the San Juan River crested at 16.81 feet, its highest since 1911. Two October storms had already filled the soil and the river.
Open the story →
Oct 25, 2026
Two Floods in a Year: Why the San Juan Rose Higher in 2026
Two Inunda simulations of the October 2025 and October 2026 floods at Pagosa Springs reveal the differences of two flood events.
Open the story →
Oct 16, 2026
When the Desert Drains: Rachel's Flash Flood in Las Vegas
How the remains of Hurricane Rachel flooded Las Vegas streets before dawn on 11 October 2026, rebuilt with the Inunda flood model at 3 m over the urban core and checked against rescue reports.
Open the story →
Oct 9, 2026
Hurricane Isaias: An AI-Driven Compound Flood Forecast for Pensacola
Google WeatherNext 3 rain and tracks, NOAA storm surge and river forecasts, and the Inunda 30 m flood model. Experimental research forecast issued 9 October 2026, ~23 UTC; not an official warning.
Open the story →
Oct 6, 2026
Validating Flood Simulations Where No Flood Map Exists: SWOT, OPERA and Gauges on the Buffalo National River
SWOT water-surface elevations and OPERA flood extent can test a flood simulation in a park that FEMA never mapped. Each has blind spots, and only together with ground gauges do they constrain the model.
Open the story →
Oct 4, 2026
Kelvin Wave Flood Outlook for Southern Marin
How the 2026 El Niño Kelvin wave is raising San Francisco Bay tides, and where the Inunda flood model puts that water on land this fall.
Open the story →
Aug 1, 2026
Decisions, Not Just Depths: Camp Mystic, 4 July 2025
A 10 m flood simulation of the Guadalupe headwaters, turned hour by hour into evacuation judgments by a judgment model (TypeSafe Jev), then checked against the gauge that recorded the flood.
Open the story →
Sep 1, 2026
The Flood That Broke the Gages
Rebuilding the August 29, 2026 flash flood on Bright Angel Creek, Grand Canyon National Park, with the Inunda flood model.
Open the story →
Oct 11, 2025
Twice in Four Days: the San Juan River Floods Pagosa Springs
A 10 m Inunda simulation of the October 2025 flood, the third largest on record at Pagosa Springs, checked against the river gage and the places people reported under water.
Open the story →
Jul 8, 2024
Three Storms on Halls Bayou
Modelling Hurricane Beryl and the 2022 and 2024 floods in north Houston at 10 m with Inunda, and checking the model against the stream gage.
Open the story →Join us
Opportunities
Please read the lab manual before you consider joining us. There are no grant-supported openings right now, but I strongly encourage you to pursue these fellowships — I'm glad to support strong applications.
Graduate fellowships
- NASA FINESST — Future Investigators in NASA Earth & Space Science and Technology · due February
- NSF GRFP — Graduate Research Fellowship Program · due November
Scholarships
- AWRA Colorado — Rich Herbert Memorial Scholarship · up to $5,000
- Open to CU Boulder students across levels
Postdoctoral fellowships
- CU Boulder Chancellor's Postdoctoral Fellowship · due early November
- Email a CV and short research statement anytime
Business partnerships
- Services: flood simulation & inundation mapping, and water-resources management
- Sponsored research, model licensing, and technical consulting
- For companies, agencies, and consultancies — email to scope a project
News archive