Great Experiments in Physics
Proposal: Great Experiments in Physics
An AI-Driven Laboratory-Seminar for CMU Physics Students
This course immerses students in the greatest experiments in the history of physics — not as historical narrative, but as active scientific practice. Each student takes ownership of two landmark experiments over the semester, one from each of two tracks: a quantitative Analysis Track, in which they re-derive a key result from pseudo-data, and a Communication Track, in which they develop a polished public-facing explanation of an experiment whose significance is primarily conceptual. Both culminate in a delivered lecture defended before peers and a faculty coach.
The course has two equally weighted motivations. The first is mastery of experimental physics — its history, methodology, and reasoning. The second is developing genuine fluency with modern AI tools. Students entering research today will use AI throughout their careers, and those who learn to use it well — as a Socratic interlocutor, a research accelerator, and a tireless collaborator — will be meaningfully more capable scientists. This course gives students structured, supervised experience building those skills in a domain that rewards precision and depth.
AI Fluency as a Core Learning Goal
The researchers who will get the most from AI are not those who use it most, but those who use it most skillfully. A well-posed question to an AI can unlock a primary paper, surface a connection across decades of physics, sharpen an analogy until it actually works, or compress hours of background reading into a focused conversation. These are real productivity gains — and they compound. Students who develop these habits early will carry them into everything they do.
This course builds AI fluency through direct, repeated practice in a high-stakes context. The dialogue phase asks students to use AI the way the best researchers already do: not to retrieve answers, but to think more carefully. Students learn to pose questions that expose the structure of a problem, to iterate on explanations until they are genuinely clear, and to use AI as a sounding board that never tires and never lets a vague claim pass unchallenged.
Specific AI skills the course develops:
- Socratic inquiry — using AI dialogue to go deeper into a topic than any textbook or lecture would take you, following threads of curiosity across primary literature, historical context, and physical intuition
- Iterative refinement — learning that the first response is rarely the best one; developing the habit of pushing further, asking follow-up questions, and demanding precision
- Research acceleration — navigating dense primary literature, connecting results across experiments and decades, and building background understanding rapidly without sacrificing depth
- Communication iteration — using AI as an editor and audience surrogate to stress-test explanations, test analogies, and refine until a piece works for a non-specialist
- Critical judgment — developing the intuition to know when AI output should be trusted, when it should be verified, and when a different question would produce a better answer
By the end of the course, students will have internalized habits of AI-augmented inquiry that make them faster, deeper, and more effective — as researchers, as communicators, and as scientists.
Course Structure
The course meets weekly for 12 sessions with a cohort of 6 students. Each student presents twice — once on an Analysis Track experiment, once on a Communication Track experiment. Pairings are assigned by the faculty coach at the start of the semester, deliberately chosen to contrast in intellectual character and stretch students across different modes of engagement.
Each session follows the same format: a 30-minute student lecture followed by 20 minutes of peer and faculty discussion. There are no problem sets or exams.
— Experiment Menu by Track (Example… we should come up with more)
| Experiment | Track | Core Concept |
|---|---|---|
| Millikan Oil Drop | Analysis | Charge quantization; measurement philosophy |
| Rutherford Scattering | Analysis | Nuclear structure; angular distributions |
| Davisson-Germer | Analysis | Wave-particle duality; diffraction |
| Pound-Rebka | Analysis | Gravitational redshift; precision measurement |
| Michelson-Morley | Communication | Null results; what absence of evidence means |
| Wu Parity Violation | Communication | Symmetry breaking; theory-experiment interplay |
| Bell Inequality Tests | Communication | Foundational interpretation; loopholes |
| Homestake Neutrino | Communication | Anomaly persistence; trust in a long-arc result |
| A failed/retracted experiment | Communication | How rigorous scientists reach wrong conclusions |
Example pairings assigned by the coach:
- Rutherford (Analysis) + Bell Inequality (Communication) — contrasting intellectual demands
- Millikan (Analysis) + Homestake (Communication) — thematic echo: both about trusting a result
- Pound-Rebka (Analysis) + Wu (Communication) — precision measurement meets conceptual revolution
Preparation Phases
All students complete the same two phases for each of their experiments:
Phase 1 — AI Dialogue (understanding) Extended Socratic dialogue with an AI assistant covering the physics, historical context, and experimental methodology. What motivated the experiment? What could go wrong? What would a skeptic object to? Students submit a curated log of this dialogue as part of their portfolio — including at least one documented instance where AI produced an incorrect or misleading response and how they identified and corrected it.
Phase 2 — Track Deliverable (doing)
Analysis Track: The student works with AI to generate a realistic pseudo-dataset including physically motivated signal distributions and injected systematic effects. They re-derive the key result, then attempt to break their own analysis — varying cuts, binning, and background models to understand what drives the answer. This phase produces a short written analysis report.
Communication Track: The student uses AI iteratively to develop a polished public-facing explanation of their experiment — a short video, a Quanta-style magazine article, or a piece written for a high school audience. AI serves as both research assistant and editor: students test analogies, stress-test explanations, and refine until the piece works for a non-specialist. This phase produces the finished communication piece.
Phase 3 — Lecture Delivery (defending) All students deliver a 30-minute seminar lecture presenting their findings and defending their methodology and choices under questioning — without AI assistance during the session.
Assessment
Each student is assessed twice — once per experiment. Scores are averaged across both presentations.
| Component | Weight | What it measures |
|---|---|---|
| AI Dialogue Portfolio | 30% | Depth and quality of questioning; critical engagement with AI failures |
| Track Deliverable | 30% | Physical reasoning; execution; critical self-assessment |
| (report or comm. piece) | ||
| Delivered Lecture | 40% | Clarity; physical intuition; ability to defend under questioning |
A significant gap between portfolio/deliverable quality and lecture performance triggers a one-on-one conversation with the faculty coach — the primary safeguard against shallow AI reliance. Students who present twice are also assessed on whether their second presentation shows growth relative to their first.
Faculty Coach Role
The faculty coach attends every session but does not lecture. Their role is to ask the hardest questions: Why that bin width? What if your background model is wrong? Does your analogy actually capture the physics, or does it mislead? Did you verify that with the original paper? The coach assigns experiment pairings at the start of the semester, reviews all portfolios and deliverables, and meets individually with students whose preparation and lecture quality diverge.
One session is devoted to a failed or retracted experiment — cold fusion, Millikan vs. Ehrenhaft, the faster-than-light neutrino claim — led by the coach. The parallel to AI-generated misinformation is explicit: in both cases, the challenge is maintaining critical judgment in the face of authoritative-sounding output.
Why This Works
Physics students will use AI throughout their careers. The question is not whether, but how well — and with what degree of critical awareness. This course addresses that directly, in a domain where errors are detectable and the cost of credulous AI use is visible. A student who generates the wrong angular distribution for Rutherford scattering because they trusted an AI without checking will not make that mistake in their research.
At the same time, the course addresses a real gap in physics training: it explicitly develops science communication alongside quantitative reasoning and AI literacy, preparing students for careers in which all three matter.
Pilot Plan
Run as a 1-unit seminar (12 sessions) in Fall 2026 with 6 advanced undergraduates and first-year graduate students. Collect portfolios, deliverables, and coach evaluations. Assess whether students show measurable growth between their first and second presentations, and whether documented AI failure cases in portfolios correlate with stronger lecture performance — a direct test of whether critical AI engagement produces better understanding.