Final Project

For your final project, you will take on an open-ended hydroclimate hazard analysis related to your work or research interests. Through the project, you will:

Practice working on open-ended hazard analysis problems.
Compare at least two methods that make contrasting assumptions.
Apply a software tool beyond what the labs cover to a problem of your choice, characterize how it differs from other approaches, and describe its key assumptions.
Communicate your recommendation to technical and non-technical audiences.

Project Description

The final project is your chance to learn a method or a software tool beyond what the labs cover. You should be ready to answer technical questions about the method (what it assumes, what it estimates, and how it fails) and questions about how the tool puts it into practice. Compare its answer with at least one other approach to the same question, which can be a more “standard” or “traditional” approach from the course. Then use methods from the course to assess how far each answer can be trusted. If you do not have data from your own research, I am happy to suggest a site and record. The project is your own work, as described under collaboration in the syllabus.

Example Projects

Below are some potential examples, listed alphabetically. Each names a method and a tool or dataset. The pairings are suggestions, not assignments.

Compound events. What is the joint distribution of two or more hazards, such as storm surge and river flow?
- Method: a tail-dependent copula, such as a vine copula, or a stochastic event generator that simulates both drivers.
- Tool: pyvinecopulib in Python, or copula in R.
Downscaling. How much fine-scale structure does bias correction miss, and does the missing structure change the hazard?
- Method: generative downscaling, which simulates fine-scale fields conditioned on coarse model output.
- Tool: NVIDIA CorrDiff, or RainFARM for a lighter spectral option without a GPU.
Event attribution. How much more or less likely is a specific observed extreme at your site than it once was?
- Method: a storyline built from downscaled climate model runs, compared with a nonstationary extreme value fit from the course.
- Data: downscaled projections such as NASA NEX-GDDP-CMIP6.
Historical and paleoflood information. How much does information from before the gauge record change the flood frequency curve?
- Method: a Bayesian fit that folds in historical or paleoflood information through perception thresholds.
- Tool: USACE RMC-BestFit, or the same model written in Stan or PyMC.
Multisite simulation. How likely is it that several places in your region get extreme rainfall at once?
- Method: a multisite stochastic rainfall generator, or a copula across sites.
- Tool: CoSMoS in R, or AWE-GEN-2d.
Regional frequency analysis. What is the design value at a short-record or ungauged site, and how much does it change when you borrow information from elsewhere?
- Method: storm transposition, which resamples the regional storm catalog, or downscaling a climate model to the site, compared with regionalized extreme value models.
- Tool: RainyDay for transposition.

Deliverables

Submit each written piece on Canvas. You may use any language or tool, as long as someone else can reproduce your analysis from what you submit.

This is a good use of an LLM. Have an AI agent check that your analysis is reproducible and identify limitations.

Selection proposal

Due Fri Oct 30. Submit one page that states:

The question you will answer, and the decision it informs.
The decision-maker your memo is for, named by role, such as “the city engineer deciding whether to upsize a culvert.”
Your site and data, with their sources.
The computing resources your tool needs, such as a GPU, a MATLAB license, or a particular operating system, and whether you have access to them.
The two or more approaches you will compare, and the assumption on which they differ.
A prediction: which approach you expect to give the larger answer, and why, based on that assumption.
Your fallback: an approach, such as a second course method, that takes the tool’s place in the comparison if the tool will not run.

After you submit, sign up for a 15-minute meeting with me during the week of Nov 2 to settle the scope: [TODO: sign-up link]. Many of the example methods are taught after the proposal is due, so you may change the approaches, the assumption, and the prediction at that meeting; your question, site, and decision-maker should stay.

Checkpoint

Due Wed Nov 18. Submit a short write-up of two pages at most, with your code so far. By this date, you must have fit at least one course method to your data, and report whether your tool runs on your data.

Nov 18 is the last day to get your tool running. If it has not run on your data by then, I tell you to switch in my checkpoint feedback. Your fallback from the proposal becomes the second approach in your comparison, and you write an account of the tool in place of its results. The account says what it assumes, where it broke, and what running it would take. It also shows the attempt: your environment, the commands you ran, and the errors they produced. The account is graded on the same rubric as a working analysis.

Include a draft of your AI use statement, described below.

Peer review

In class Mon Nov 23. Bring a draft of the methods section of your paper. You will read a classmate’s draft and answer two questions about it. What assumption separates the approaches being compared? Which result, if it came out differently, would change the recommendation? This session counts toward your participation grade, as described in the syllabus, and carries no project points.

Paper, memo, code, and AI use statement

Due Fri Dec 4.

Paper, ten pages at most, excluding references. It states the question, explains each approach and what it assumes, compares their answers, and assesses their credibility with methods from the course. Quote the prediction from your proposal and say whether it held; a wrong prediction costs no points.
Memo, one page, addressed to the decision-maker you named in your proposal. It gives your recommendation and how confident you are in it, in terms that person can act on, with no equations.
Code, as a repository link or a zip archive, with a README. Following the README should reproduce every figure and number in the paper.
AI use statement, a separate document of one page at most. Say how you used AI tools in the project and why you think that use was appropriate, including what you take “appropriate” to mean. I am looking for a considered answer. The statement is required and unscored, and I do not grade your paper until I have it.

Defense

During the final exam period, in the three-hour slot the registrar assigns this course: [TODO: date and time once the registrar posts it]. The defense has two scored parts. In the talk, you present for 12 minutes, with a hard stop at 14 minutes, followed by 8 minutes of questions from me and your classmates. After all the talks, you meet with me one-on-one for about 10 minutes of questions. I ask detailed questions about your method, your tool, and your own analysis, and I prepare the questions from your paper. I record the one-on-one, as with the oral examinations, so that grading can be audited for consistency. Arrange any accommodation that affects the defense format with me in advance, as described in the syllabus.

Grading

The project is out of 100 points and is worth 30% of your course grade.

Component	Points
Selection proposal	5
Checkpoint	5
Paper	25
Memo	15
Code	5
Talk	15
One-on-one questions	30

The proposal and the checkpoint earn full credit for a complete submission that reports where your project actually stands. Each piece loses 20% of its points per day late, which is 1 point per day on the proposal or the checkpoint. Extensions must be arranged in advance, as described in the syllabus.

The table scores the proposal and the checkpoint on completeness, the paper and the one-on-one questions on three criteria each, and each other component on one. Every criterion below is scored at one of four levels: excellent earns full points, proficient 80%, developing 50%, and missing zero. The table describes the first three levels; missing means the piece was not submitted or does not address the criterion.

Criterion	Points	Excellent	Proficient	Developing
Selection proposal	5	Every item listed under “Selection proposal” is addressed.	One item is missing.	Two or more items are missing.
Checkpoint	5	At least one course method is fit to your data, the tool’s status is reported, and the AI use statement is drafted.	One of those three is missing.	Two or more are missing.
Question and method (paper)	8	The question is specific and tied to the decision. You explain what the method assumes, what it estimates, how it fails, and how the tool implements it.	The question is clear and the method is explained, but one of its assumptions or failure modes is missing.	The question is vague, or two or more of the method’s assumptions and failure modes are missing.
Comparison (paper)	9	You compare at least two approaches, explain how far apart their answers are through the assumption that separates them, and say whether your prediction held.	The approaches are compared, but the difference is reported without being explained.	Only one approach is carried through, or the comparison is a side-by-side list of results.
Credibility and uncertainty (paper)	8	You use course methods, such as diagnostics, checks against data you did not fit to, or uncertainty intervals, to judge how far each answer can be trusted.	Uncertainty is quantified for some answers but not used to judge them.	Answers are reported without uncertainty or any check.
Memo	15	A non-technical reader could act on it: the recommendation, its uncertainty, and what would change it.	The recommendation is clear, but the uncertainty or what would change it is missing or stated in technical terms.	The memo summarizes the analysis without making a recommendation.
Code	5	Following the README reproduces every figure and number.	It reproduces the main results with small fixes.	It runs only with major fixes, or the README does not say how to run it.
Talk	15	Within 12 minutes, and a classmate could restate your question, the assumption that separates the approaches, and your recommendation.	Clear, but over 12 minutes or missing one of those three.	Hard to follow, or a tour of methods without a result.
Method (one-on-one)	10	You explain what the method assumes, what it estimates, and the conditions under which its assumptions fail.	You state the method’s principal assumptions, but not when they fail.	You can name the method, but not its assumptions.
Tool (one-on-one)	10	You explain how the tool puts the method into practice: its inputs, the choices it makes for you, and the settings you changed.	You explain the tool’s inputs and outputs, but not the choices it makes for you.	You can run the tool, but not explain what it does between input and output.
Your choices (one-on-one)	10	You justify your modeling choices with evidence, state what would change your answer, and revise a choice when a flaw is identified.	You give a substantive justification for your choices.	You justify your choices only by convention, or cannot explain parts of your own analysis.

If your tool did not run, the account of it replaces its results under “Question and method” and “Credibility and uncertainty”, and your fallback counts as the second approach under “Comparison”.