DSC 180ab B03 - Week 9: Explore and propose

This is the last working week of Phase I, and the most open. Rather than more set tasks, you will look back over what you have learned, then test one or two ideas for Phase II cheaply, so that you choose your project based on evidence rather than guesswork. By the end of the week your team should have a short written proposal and a 3-minute pitch.

Good Phase II projects usually start from something that surprised you in Phase I: a gradient that went wrong, an emulator that failed on one kind of scenario, or a result you didn’t believe.

Tasks

1. Look back (a day)

Before choosing anything, spend a day consolidating what you have learned:

• Re-read your NOTES.md and your weekly answers. List the three results from Phase I you understand least well, and the one that surprised you most.
• Pick one of those and try to explain it to a teammate without notes. Where your explanation gets vague is often where a good project question is hiding.

2. A feasibility probe (most of the week)

Pick a Phase II direction from the ideas on the course page, or your own, and run the smallest experiment that tells you whether it is worth a quarter of work. Team members can probe different ideas. Aim for one clear figure. Some suggestions:

Idea	A small probe	It looks promising if…
New emulator architectures	Train a small U-Net or Fourier neural operator on ClimateBench tas only, with a 1-hour training budget	It beats your week 3 baseline somewhere specific, such as regional patterns or aerosol-driven responses
Physically constrained emulators	Add a global-mean energy-balance layer to your network and test it on one of your task 1 scenarios	It behaves more sensibly out of distribution, at little cost in skill
Gradient-based scenario analysis	Extend the week 8 inverse problem with a regional target or an emulator ensemble	The answers are stable and physically interpretable
Longer-window gradients in JCM	Average week 6 gradients over many starting states (the “ensemble adjoint” of Lea et al., 2000) and compare with finite differences	Averaging extends the window over which gradients agree with finite differences
Calibrating JCM	Use gradient descent on one or two SPEEDY parameters, such as cloud albedo (shortwave_radiation.albcl), against CERES global means, following JCM’s optimization notebook	The loss is well behaved and the calibrated values are physically plausible
A JCM–JEM benchmark	Time a one-year coupled JCM–JEM run with rising CO₂, and estimate the cost of 20 scenarios of 100 years	It fits within your compute, and the result looks like a sensible climate
Decision support	Identify one user and one question, then sketch a Streamlit prototype on your weeks 7–8 emulator	You can state what the user would decide differently with your tool

Alongside the probe, read one or two papers on your idea (the reading list has starting points), and talk it over with Amirhossein or me before you commit.

3. Write a proposal

Write a 1–2 page proposal as a team, covering:

1. The question, in one or two sentences, and why it matters (and to whom).
2. Data and models. What you will use, and whether you have access to it already.
3. Approach. What you will build or run.
4. What success looks like. A metric, a baseline to beat, or a figure you hope to make.
5. Risks and a fallback. What is most likely to go wrong, and a smaller version of the project that would still be worthwhile.
6. What your probe showed, with its figure.
7. A plan for the first three weeks of winter quarter.

Present a 3-minute pitch in section. We will use the feedback to finalize the proposals in week 10.

Questions

Submit on the form linked in Slack before section.

1. What did your probe show, and how did it change your plans?
2. What is the biggest risk in your proposed project, and what is your fallback?
3. What do you need from us to get started: data, compute, or people to talk to?