Student_Guide_Building_to_Understand_Creating_Simulations_V1.docx

Scientific Inquiry with AI • Creating AI-Generated Simulations • Student Guide Page 1
AI Simulation Creation
Creating AI-Generated Simulations as a Mode of Scientific Inquiry
Scientific Inquiry with AI
Student Guide
Detailed Instructions & Course Material
Why This Lesson
You have used simulations before: a slider you drag, a curve that updates. This week you build
one. With an AI, you will specify a phenomenon you choose, generate a working simulation, and
then do the part that matters most: prove whether it is right.
The power of AI is that the slow part, writing the code, becomes fast. The danger is that AI
produces fluent, confident output that can be quietly wrong: a missing force, a hidden shortcut, a
number that looks plausible and is not. The skill you are building is judgment: making the model,
then checking it until you trust it.
Learning Goals
By the end of this lesson, students will be able to:
Phenomena
● Use a validated simulation to explore parameter regimes, generate quantitative predictions,
and observe emergent behavior not obvious from the equations or rules alone.
Experimental Practices
● Translate a conceptual or mathematical model into a precise natural-language specification
that an AI system can build from.
● Validate AI-generated simulations on two independent axes: technical validation, which asks
whether the artifact runs as specified, and scientific validation, which asks whether it
respects governing relationships, limiting cases, conserved quantities, dimensional
consistency, and order-of-magnitude expectations.
● Iteratively refine a simulation through targeted prompting, using discrepancies between
output and independent predictions as the driver of each revision.
● Design and run controlled experiments inside a simulation by varying one parameter at a
time, sweeping ranges, and seeking edge cases that probe the model's limits.
Habits of Mind
● Predict before prompting and before running so the simulation retains the power to surprise.
● Verify, do not trust: treat AI output with calibrated skepticism, because its capacity to be
wrong makes verification an act of learning.
● Reflect on what building the model taught that using a finished tool would not, and where
both model and measurement begin to fail.
V1

Scientific Inquiry with AI • Creating AI-Generated Simulations • Student Guide Page 2
What You'll Need
● One paid AI chatbot, such as ChatGPT, Gemini, or Claude (the paid tier your institution
provides). Any one is fine, as long as it has a live code or preview pane.
● A web browser. Your simulation will be a single self-contained .html file you open like any
web page. No install, no API key.
● NotebookLM (free, Google account) for the pre-class reading.
● The reading: Leveraging generative artificial intelligence for simulation-based physics
experiments (provided).
● A PhET (or equivalent) simulation in your area of interest, for the historic-tools exploration.
● The Prelab Tutor document and the Simulation Improvement Playbook (both provided).
Keep the Playbook open while you build.
The Core Idea
● Students do not merely use a simulation; with an AI they build, validate, and refine one, and
the act of getting the model right is the learning.
● A simulation is a model: it has predictive power and a bounded domain where it holds. It is
not a window onto reality.
● A model earns trust by passing independent checks: reproducing known cases, conserving
what must be conserved, and agreeing with a calculation the student can make themselves.
● Two rules carry the whole lesson: predict before you prompt, and verify before you trust.
● The driving question, in the worked example: how does the swing period of a pendulum
depend on how far back you pull it before release?
The Three Parts at a Glance
Part You will... Time
Part 1 - Pre-Class Explore a historic simulation tool, complete a Socratic tutor
session, research the science, and write a model
specification by hand.
~1.5 h
Part 2 - In Class Watch a live fail-and-validate demo, then co-investigate with
AI: build a validated v1, peer-test a classmate's simulation,
and revise to a v2.
~2 h
Part 3 - Post-Class Fact-check an AI-generated report, extend your simulation,
and complete retrieval practice.
~2 h
V1

Scientific Inquiry with AI • Creating AI-Generated Simulations • Student Guide Page 3
Executive Summary: The Core Philosophy
● A simulation is a model. It has predictive power and a limited domain where it holds. It is not
a window onto reality.
● AI can build a working simulation from a good specification in minutes. It can also encode
the wrong model and present it confidently.
● A model earns trust by passing independent checks: it reproduces known cases, conserves
what must be conserved, and agrees with a calculation or estimate you can make yourself.
● Two rules carry the whole lesson: predict before you prompt, and verify before you trust.
● Validating and refining is where the learning lives, the part a student handed a finished
simulation never gets to do.
Worked Example: The Damped Pendulum
Here is the whole arc applied to one phenomenon, so you can see the shape of it. It is a physics
example because it has clean checks, but your topic can be anything; the moves are the same.
Specify (before class)
● Driving question: how does the swing period depend on how far back I pull the pendulum?
● Variables: angle theta and angular velocity omega. Parameters: length L, gravity g = 9.81
m/s squared, a damping strength, release angle theta-zero.
● Governing relationship: theta-double-dot = -(g/L) sin(theta) - (damping) omega. The
small-angle version swaps sin(theta) for theta, a shortcut I will test, not assume.
● Assumptions: point mass, rigid massless rod, damping proportional to speed, planar motion.
● Two checks I can use: with no damping and a small angle, the period should be T = 2 pi
sqrt(L/g), about 2.0 s for a 1-m pendulum; with damping off, total energy should stay
constant.
Build and validate (in class)
● Predict first: about a two-second period; energy flat when undamped; the swing dies out
faster if I increase damping.
● Generate, then check the mechanics: do the length, damping, and release-angle controls all
work? Does it plot theta over time?
● Check the science: measure the period at a small angle. Is it near 2.0 s? Turn damping off.
Does energy hold steady? If energy drifts, the model or the math is wrong; refine and
re-check.
Turn it into an instrument (the payoff)
● Sweep the release angle. The small-angle formula says the period should not change with
amplitude. The real pendulum disagrees: pull it back further and the period grows, a fraction
of a percent by 20 degrees, several percent by roughly 57 degrees (1 radian), and much
more as you approach 180 degrees. If your sim's period does not budge, the AI quietly used
the small-angle shortcut and you have caught a wrong model.
● Find where it breaks. Crank the damping up: past a critical point the pendulum stops
swinging entirely and just creeps back (overdamped). Push the release angle toward 180
V1

Scientific Inquiry with AI • Creating AI-Generated Simulations • Student Guide Page 4
degrees and the simple formula falls apart. Note that real air drag is not quite proportional to
speed, so even your assumption has a limit.
Part 1: Pre-Class Preparation ~ 1.5 hours
A. Exploration of Historic Simulation Tools (~15 minutes)
Open one professionally built simulation in your area of interest, such as a PhET simulation
(https://phet.colorado.edu/) or equivalent. Touch every control. Before you move a slider, predict
what it will do, then evaluate how the simulation benefits learning of the specific topic. You are
seeing what a good, finished simulation is worth, right before you build your own.
B. Socratic AI Tutoring Session (~20 minutes)
Load the provided Prelab Tutor document into your AI chatbot. The tutor draws out, rather than
tells, the core frame: what it means for a simulation to be a model with bounded validity, what
makes a model trustworthy, and how creating a simulation differs from simply using one. The
tutor will not build or specify any simulation for you.
● The tutor builds curiosity about how simulations can support inquiry.
● The tutor asks diagnostic questions about models, variables, assumptions, and validation.
● The tutor uses Socratic questioning and structured feedback to surface misconceptions
about AI output, model validity, and evidence.
Share or submit the chat log with your instructor as evidence of completion.
C. Assigned Readings and Resources (~20 minutes)
Load “Leveraging generative artificial intelligence for simulation-based physics experiments: A
new approach to virtual learning about the real world” into NotebookLM. Use its features to
summarize the paper's findings, focusing on how student-built, iteratively validated AI
simulations compare to traditional labs and pre-built simulations for conceptual learning, and
ask questions about the approach.
D. Model-Specification Exercise: Write, Self-Test, AI Critique (~30 minutes)
This is the most important pre-class task, the analog of a traditional pre-lab. Without AI, write a
one-page specification with these fields: the phenomenon and one driving question; the state
variables and parameters, with units (or the equivalent quantities for a non-science model); the
governing relationships and your key assumptions, meaning what you are neglecting and why;
and at least two validation checks, such as a known case, a quantity that must stay fixed, a
consistency or dimensional check, or an analytic result. Then, and only then, submit it to your AI
for critique only (Appendix A, prompt 0): ask it to flag missing assumptions, ambiguity, and gaps.
Revise. Do not ask it to build anything yet.
Worked fields (damped pendulum): phenomenon, a mass swinging on a rod, losing
energy to damping; driving question, how does the period depend on the release angle?
V1

Scientific Inquiry with AI • Creating AI-Generated Simulations • Student Guide Page 5
variables, theta (rad), omega (rad/s); parameters, L (m), g = 9.81 m/s squared, damping
strength, theta-zero (rad); governing relationship, the full nonlinear model, not the
small-angle shortcut; assumptions, point mass, rigid massless rod, damping proportional to
speed, planar motion; check 1, undamped small angle gives T = 2 pi sqrt(L/g) is about 2.0 s;
check 2, energy stays constant when damping is off.
E. Student Guide and Build-Guide Review (~10 minutes)
Review Part 2 of this guide so you arrive in class ready to work, and confirm you can reach your
AI chatbot and open a .html file in your browser.
HTML file vs. Artifact, GPT, or Gem. When an AI builds a simulation, the result can appear
as an in-app live view, such as a Claude Artifact, a ChatGPT GPT, or a Gemini GEM. These
are excellent for fast iteration, but they require using a specific AI model and are not portable
files. Your deliverable is a single self-contained .html file that you download and open in a
browser. It runs offline, looks the same regardless of which AI built it, and is yours to validate
and submit. Iterate in the live view if you like, but always ask the AI to put everything in one
self-contained HTML file and then download that file.
Part 2: In-Class Investigation ~ 2 hours
The Investigation, Stage by Stage
Your instructor leads the class through a shared rhythm, specify, generate, validate, refine,
probe, while you apply it to your own phenomenon.
Stage 1 – Opening Retrieval Exercise (~5 minutes)
Answer a quick retrieval question connecting back to the prelab: state the two axes of validation
and name one concrete check for your own phenomenon.
Stage 2 – Co-Investigation with AI (~100 minutes)
First, watch your instructor run a live fail-and-validate demonstration: a deliberately flawed
simulation, caught and corrected in real time. Then, with AI, getting a first working version is
fast, so class runs in three rounds: build a validated v1, test a classmate's simulation and trade
feedback, then revise to a v2.
Build round (~45 minutes). Steps 1 through 6 build and prove the model; steps 7 and 8 use it
as an instrument. Get to a working, validated first version.
V1

Scientific Inquiry with AI • Creating AI-Generated Simulations • Student Guide Page 6
● State the phenomenon and the driving question in one sentence.
● Predict first. Write the expected behavior, a sketch, a limiting value, a trend, before you
prompt. This is what lets the simulation surprise you.
● Write the specification prompt. Turn your Phase-1 spec into a build prompt: variables and
units, governing relationships, the controls and outputs you need, and the format, one
self-contained .html file you can download. Precision here is the learning.
● Generate, then validate technically. Does it launch, respond to every control, and produce
the outputs you asked for? Note what's missing, but don't judge the science yet.
● Validate scientifically. Run your checks: a known case, a quantity that must stay fixed, a
consistency check, a calculation at one point. Record each check and its result.
● Refine through targeted prompting. For each discrepancy, name the likely cause, then write
a specific fix. Re-validate after every change, and verify the fix yourself rather than trusting
the AI's word that it worked.
● Switch to instrument mode: sweep a parameter. Vary one thing across a range and watch
the response. Look for thresholds, resonances, saturation, behavior the rules don't make
obvious.
● Find where the model breaks. Push it toward an edge case where you expect it to fail.
Document the failure and explain it using the assumptions you listed.
Peer-test round (~20 minutes). Swap simulations with one or two classmates. Use someone
else's as a real user would, then give them structured feedback: Does every control work? Is the
validation readout clear and easy to find? Can you reach a case where you know the answer,
does it match? Can you find where it breaks? Did anything behave wrong or look mislabeled?
Notice where your validation strategies overlapped even though your topics differed.
Revision round (~20 minutes). Fold the feedback you got into a second version (v2).
Re-validate every change. Then capture one insight the rules alone didn't give you, one
sentence.
Student Reflection (~10 minutes)
● When did your method begin to fail, and when did the model itself begin to fail?
● What did building teach that using a finished simulation would not?
● Where were you tempted to accept an AI answer without checking, and what did checking
reveal?
● What changed between your v1 and v2 after peer feedback?
Part 3 - Post-Class Assignment (~2 hours)
Part 3: After Class ~ 2 hours
V1

Scientific Inquiry with AI • Creating AI-Generated Simulations • Student Guide Page 7
A. AI-Generated Report and Review (~45 minutes)
Ask the AI to assemble your specification, validation checks and outcomes, parameter sweep,
failure case, and figures into a report (Appendix A, prompt 7). Then audit every claim against
your own records; the review is the graded part, not the generation. Pay special attention to:
● Numbers: does each value match what you actually measured or calculated?
● Causes: did it explain a discrepancy correctly, or paper over it?
● The failure case: is the point where your model breaks stated honestly, with the assumption
that caused it?
Keep a short note of every correction you make; that list of catches is direct evidence of the
verify, don't trust habit, and it is assessed.
B. Extension Activity: Add Additional Features to Simulation (~45 minutes)
Extend your simulation to answer a new question that came up while you were testing and
discussing it. Aim for a problem that would be impractical or very difficult to investigate
experimentally: changing gravity, speeding up or slowing down the time scale, or testing an
unusual environment. Validate the addition as carefully as the original; the easiest check is to
turn the new setting back to its original value and confirm you recover your already-validated
model.
For the pendulum, workable extensions include: running the sim at lunar gravity (g = 1.62
m/s squared) and predicting the new period before checking it; compressing or stretching the
time scale to observe many decay cycles in seconds; or testing a damping regime meant to
represent an unusual environment, such as a viscous fluid, and checking whether the model
still conserves what it should when damping is set back to zero.
Write down: the new question your extension answers; your prediction versus what it actually
showed; and the one check you used to trust the addition.
C. Retrieval Practice / Course Review (~30 minutes)
Spend about 30 minutes on integrated course review using the retrieval tool, focused on the
validation habits: the two axes, the kinds of independent check, and the two guardrails, so they
carry into future lessons.
Expect a spot-the-mistake item. You will be shown a picture of a simulation or its output
with a planted error, such as energy drifting when it shouldn't, a period that won't change with
amplitude, mislabeled axes, or a total that isn't actually conserved, and asked to say what's
wrong and why.
V1

Scientific Inquiry with AI • Creating AI-Generated Simulations • Student Guide Page 8
Make Your Simulation Better and More Useful
A validated simulation is the floor, not the ceiling. These are the highest-value moves to make
yours genuinely useful, to you now, and to your future self reusing it. The standalone
Simulation_Improvement_Tips.docx has the full set with copy-paste prompts; here are the
essentials.
Make it trustworthy
● Do show a live readout of the quantity your check depends on (period, total energy, a
conserved sum, a key metric) so validation is one glance, not a guess. Don't bury the
science behind animation alone.
● Do add a reset button and let users type exact starting values. Don't hard-code one scenario
you can't reproduce.
Make it explorable
● Do expose the real parameters as sliders with sensible ranges and units on the labels. Don't
lock in magic numbers a user can't change.
● Do let users push to extremes so the model's limits are visible. Don't clamp the inputs to
only the pretty regime.
Make it reusable
● Do keep it one self-contained file you download, label the assumptions on-screen, and add a
one-line what this models and where it breaks. Don't leave it as an in-app preview or an
unlabeled black box you won't understand next month.
Everything You Turn In
● How it was created: your generation prompt plus a change log of the follow-up prompts you
used to debug, correct, and revise it, including changes made after peer feedback.
● Validation record (two axes): technical evidence it runs and responds; scientific record of
your inputs, expected values, results, and the exact point where the simulation breaks down.
● The simulation: a single self-contained .html file (the downloaded file) that runs immediately.
● The extension: the new question, your prediction versus result, and the check that validated
the addition.
● Reflection: where your method failed versus where the model failed; what building taught
that using would not; where you were tempted to over-trust the AI and what checking
revealed; plus your list of report corrections.
Appendix A - Prompt Templates
Adapt the bracketed parts to your phenomenon. Use them in order.
0. Specification critique - Phase 1, before any building
Here is my model specification for [phenomenon]. Do NOT build a simulation. Critique it only:
where are the assumptions incomplete or unstated? Where are units, variables, or governing
V1

Scientific Inquiry with AI • Creating AI-Generated Simulations • Student Guide Page 9
relationships ambiguous? Which of my validation checks are not truly independent of the
model? List what I should add or sharpen before I build.
1. Generate - turn your validated spec into a build
Build a single self-contained HTML file (inline JS/CSS, no external libraries, runs offline in a
browser) that simulates [phenomenon]. Give me the full file to download, not just an in-app
preview. State variables: [...]. Parameters with units: [...]. Governing relationships: [...], use
the full model, do not silently apply simplifying approximations. Controls: sliders for [...] with
ranges [...]. Outputs: [animation/plot of ...] and a live numeric readout of [the quantity my
validation depends on]. Include a reset button and let me type exact starting values.
2. Validate technically
List every control and output you implemented. For each, tell me exactly what it should do so
I can test it. Do not assure me it works, give me the checklist I use to confirm it myself.
3. Validate scientifically - drive it yourself, AI assists
I am checking this simulation against [known case / conserved quantity / hand calculation].
For the inputs [...], the model should give [my predicted value], because [reasoning]. Show
me how to read the corresponding value off the simulation. Do not change the code yet, just
help me locate the number.
4. Refine - one targeted fix at a time
My check found a discrepancy: I expected [...], the simulation shows [...]. I think the likely
cause is [my hypothesis]. Make only the change needed to address that cause, and tell me
what you changed and why. I will re-run my check.
5. Sweep - instrument mode
Add a mode that sweeps [parameter] across [range] and plots [output] versus [parameter], so
I can see how the behavior changes. Keep the single-run view available too.
V1

Scientific Inquiry with AI • Creating AI-Generated Simulations • Student Guide Page 10
6. Find the failure
I want to push this model until it breaks. Given my assumptions [...], at what input values
should the model stop being valid? Let me drive the controls to those values, do not prevent
it, so I can observe and document the failure.
7. Report - Phase 3, you fact-check the output
Assemble a report from material I will paste: my specification, each validation check with its
predicted and observed values, the parameter sweep, and the failure case. Use only what I
give you, do not invent results or fill gaps. Mark anything you are unsure about so I can verify
it.
Appendix B - Phenomenon Menu (a Starting Point)
Any topic, not just science. For this first build, your simulation can model anything you find
interesting: a physics phenomenon, an economic market, an election map, a sports statistic,
a historical what-if, or a board-game probability. The skill is identical: state the model, build it,
and prove it is right. Validation still applies outside STEM because every model makes
assumptions, holds only within a limited range, and can be checked against a known case,
something that must stay fixed, or an independent estimate. Later in the course, you will be
required to build specific scientific simulations; this lesson is the place to build anything as
you learn to be a creator.
Field Modelable phenomena (with a natural check)
Physics Projectile with air drag (no-drag parabola limit); damped pendulum (small-angle
period, energy); RC charging (exponential, time constant); two-body orbit (energy
and angular momentum).
Chemistry Reaction kinetics and rate laws (first-order half-life); equilibrium / Le Chatelier;
titration curves; ideal-gas state changes (gas-law limits).
Biology / Ecology Logistic growth (carrying-capacity limit); predator-prey (oscillations); enzyme
kinetics (Michaelis-Menten saturation); SIR spread (S+I+R conserved).
Math / CS Random walk and diffusion (RMS ~ root-N steps); Monte Carlo pi (converges to pi);
cellular automata; numerical-integration error vs. step size.
V1

Scientific Inquiry with AI • Creating AI-Generated Simulations • Student Guide Page 11
Engineering 1-D heat diffusion (steady-state profile); beam deflection; feedback control of a
setpoint; signal sampling and aliasing.
Earth / Environmental Radioactive decay (half-life, conserved total); one-box carbon cycle; pollutant
dispersion; simple climate energy balance (equilibrium temperature).
Economics / Social Supply-demand equilibrium (market clears at a known price); rumor or disease
spread (population conserved); redistricting / voting (total votes conserved).
History / Civics A scenario what-if (check against the documented outcome); growth of a society or
resource (a known limiting size); spread of an idea or technology (saturation).
Sports / Games Shot or play probability (matches known success rates); dice / card odds
(converges to the exact probability); a game-balance simulation (fairness/symmetry
check).
Appendix C - Example Prompts and Simulations
To see the range before you start, here are example prompt-and-simulation pairs the team has
already built. The damped pendulum (provided with this lesson) comes with its full generation
prompt in Pendulum_Simulation_Prompt_and_Build_Log.docx. Look at one or two, read their
prompts, and notice how specific the prompt has to be to get a correct, checkable result.
● Science: Wave interference; projectile motion; predator-prey (Lotka-Volterra) ecology; a
bouncing-ball coefficient-of-restitution explorer; an energy-sources simulation; a
how-far-can-you-see perception explorer.
● Beyond science: An exam-practice tutor; a sports shot-probability model; an engineering
bridge-load tester; a redistricting / gerrymandering explorer; a historical scenario simulation.
Your instructor can point you to the shared gallery of these files and the prompts that produced
them.
Companion Documents
● Lesson_Overview_Building_to_Understand_Creating_Simulations_V1.docx: the
course-facing summary of the lesson arc.
● Instructor_Guide_Building_to_Understand_Creating_Simulations_V1.docx: your instructor's
facilitation companion.
● Prelab_Tutor_Building_to_Understand_Creating_Simulations_V1.docx: the Socratic AI tutor
document you use before class.
● Student_Guide_Building_to_Understand_Creating_Simulations_V1.docx: this document.
V1

Scientific Inquiry with AI • Creating AI-Generated Simulations • Student Guide Page 12
Additional Support Files
● Simulation_Improvement_Tips.docx: reusable checklist and copy-paste prompts for making
any validated simulation more trustworthy, explorable, and reusable.
● Damped_Pendulum_Simulation.html: self-contained reference simulation used in the live
demo and referenced throughout the materials.
● Pendulum_Simulation_Prompt_and_Build_Log.docx: exact generation, validation,
refinement, and sweep prompts used to build the reference simulation, with recorded
outcomes.
V1