Teaching "AI in Science" to Architect mode (ages 15–18)

Hook & Warm-Up

No mascot, no gamified framing needed at this level — treat it like the opening of a real lecture. Say: "AI solved protein folding. Now it's transforming weather prediction and drug discovery. Let's examine how AI is becoming essential to scientific research."

Open with a sharper framing question than the younger age bands get: "The scientific method is usually described as: form a hypothesis, run an experiment, analyze the results. As we go through today's examples, I want you tracking which step — or steps — AI is actually automating in each case. It's rarely all of them." This gives the class an analytical lens to apply throughout rather than just collecting impressive facts.

Optional warm-up for a class that's covered machine learning fundamentals already: "If a model is only as good as the data it was trained on, what does that imply about a scientific AI model trained on, say, only the proteins humans have already managed to solve experimentally? What's it likely to be worse at?" This previews the training-data-dependency theme that should run through the whole lesson.

Main Activity

Architect mode should treat this less like a guided tour and more like a seminar — present the mechanism, then open it to scrutiny. Each scene below includes the technical detail this age band can handle, plus a pointed question about limitations.

Scene 1 — Space & Physics. Present Planet Hunter, Dark Matter, Stargazing, and Black Hole Image. "Exoplanet detection via the transit method relies on measuring periodic dimming in a star's light curve as a planet crosses its disk. NASA's Kepler and TESS missions produced enormous volumes of light-curve data, and models like AstroNet — a convolutional neural network — were trained on confirmed and false-positive transit signals to classify new candidates, catching signals too subtle or too numerous for manual vetting to scale to." Push further: "What does a false positive look like in this context, and why would a neural network need to be trained on false positives specifically, not just real transits?" (Good answer: things like eclipsing binary stars or instrument noise can mimic a transit signal — the model has to learn to distinguish the real pattern from lookalikes.)

For dark matter simulation, be precise: "These are N-body and hydrodynamical simulations where AI/ML surrogate models are increasingly used to approximate expensive physics calculations, letting researchers run many more simulated universes under different parameter assumptions than direct simulation alone would allow, then compare statistical properties of the simulated universe to observational data like galaxy clustering." On the black hole image, give the full picture: "The 2019 Event Horizon Telescope image was reconstructed computationally from very-long-baseline interferometry data using algorithms including CHIRP; the 2023 PRIMO reprocessing of the same dataset specifically applied a machine learning model trained on synthetic black hole images to fill in gaps in the sparse telescope coverage and sharpen the reconstruction." Discussion prompt: "PRIMO's neural network was trained on simulated black holes based on general relativity. What's the epistemological risk of using a model trained on simulated data to reconstruct an image of something we're trying to verify actually looks like the simulation predicted?" This is a genuinely open, contested-adjacent question worth sitting with rather than resolving quickly.

Scene 2 — Climate & Earth. Present Ocean Monitoring, Weather AI, Growing Plants, and Wildfire AI. "Satellite-based Earth observation AI applies computer vision models to multispectral or hyperspectral imagery to detect coral bleaching signatures, sea surface temperature anomalies, and active fire hotspots, often at resolutions and update frequencies impossible for manual review." For weather, give the real technical distinction: "Google DeepMind's GraphCast is a graph neural network trained on ERA5 reanalysis data — decades of historical atmospheric state — that learns to predict how the full 3D atmospheric state evolves over time. This is a fundamentally different paradigm from numerical weather prediction, which numerically integrates the primitive equations of fluid dynamics and thermodynamics from current conditions. GraphCast generates a 10-day global forecast in under a minute on a single machine, versus hours on a supercomputer for physics-based NWP — and outperforms the leading physics-based model, ECMWF's HRES, on the majority of the variables and lead times it was evaluated against." Push: "NWP models are built from first-principles physics; GraphCast is built from historical pattern-matching. What kind of event would you predict GraphCast to struggle with, that a first-principles model wouldn't?" (Good answer: conditions with no close historical analog — genuinely unprecedented extremes, which is a real, acknowledged concern with purely data-driven weather models.)

Scene 3 — Biology & Chemistry. Present Protein Folding, Virus Tracking, Basic Microscope, and Material Discovery. "AlphaFold2 uses a Transformer-based architecture (Evoformer) that processes a protein's amino acid sequence alongside evolutionary information from multiple sequence alignments and known structural templates, predicting 3D atomic coordinates directly. It was trained on the Protein Data Bank — the relatively small set of experimentally solved structures — and has since produced predicted structures for over 200 million proteins, essentially the full UniProt catalog, at accuracy that in many cases rivals experimental methods." On materials: "Google DeepMind's GNoME project used graph neural networks trained on known stable crystal structures to screen candidate compositions, proposing roughly 2.2 million new crystal structures, of which about 380,000 were predicted to be thermodynamically stable — all still requiring experimental synthesis and characterization before being considered validated new materials." Push: "AlphaFold's training data is every experimentally solved protein structure humans have ever published. What does that imply about how well it's likely to perform on proteins that are structurally very different from anything in that historical record — say, from an extremophile organism nobody's studied?" (Good answer: performance is likely to degrade on structurally novel proteins underrepresented in training data — this mirrors the exact same generalization concern raised about medical AI and diverse patient populations in other AIQ lessons.)

The Virus Tracking item is worth a fuller technical treatment with this age band, since most of them lived through it: "During COVID-19, AI models analyzed genomic sequencing data, mobility data, and case reports to model transmission dynamics and identify emerging variants. On the vaccine side, computational tools helped researchers rapidly evaluate candidate mRNA sequences and predict aspects of viral protein structure, compressing parts of the traditional development timeline." Push: "Given everything we've said today about AI predictions needing experimental validation, what part of vaccine development do you think AI could NOT compress — and why did clinical trials still take many months even with faster computational tools?" (Good answer: clinical trials test real biological safety and efficacy in real human bodies over time — that's an empirical, not computational, step, and no simulation substitutes for observing an actual immune response and actual safety data over the required follow-up period.)

Close the activity with the framing question from the hook: "Across all three domains, which step of the scientific method — hypothesis generation, experiment execution, or results analysis — is AI actually automating? Is it ever automating all three?" Drive toward the conclusion that today's examples are overwhelmingly about accelerating hypothesis generation and results analysis (finding candidates, finding patterns, generating predictions) — the experimental validation step, where a claim actually gets tested against physical reality, remains almost entirely human-run.

Timing note: at full depth with the pushback questions, this activity runs 18–22 minutes, closer to a full class period alongside discussion. If time is short, the GraphCast paradigm distinction and the AlphaFold training-data-generalization question are the two most valuable threads to protect.

Discussion

• "GraphCast learns from historical weather patterns rather than solving physics equations. As climate change makes historical weather patterns less representative of the future, what does that imply about the long-term reliability of purely data-driven weather models?"
• "AlphaFold's training data is every protein structure humans have ever experimentally solved. Is that dataset representative of all proteins in nature, or biased in some way? What would that bias predict about its blind spots?"
• "PRIMO's black hole reconstruction was trained on simulated data generated from general relativity. If a future observation ever contradicted general relativity, could a model trained this way even represent that possibility?"
• "Self-driving laboratories can design, run, and analyze experiments with minimal human involvement. What scientific or ethical guardrails would you want in place before trusting one with a high-stakes research question?"
• "If AI keeps getting better at generating scientific hypotheses and finding patterns, what does the job of a human scientist look like in twenty years? What does it definitely still require a human for?"
• "GNoME proposed roughly 2.2 million candidate crystal structures, of which about 380,000 were predicted stable. What does the gap between 'predicted stable' and 'confirmed stable material' tell you about how much this technology has actually accelerated materials science, versus how it might be overstated in a headline?"
• "Peer review exists to catch errors and overclaiming in published science. What would peer review need to look like for a paper whose central result came from an AI model, rather than a traditional experiment?"

Quiz Walkthrough

These questions test precise technical understanding, not just general awareness — if a student gets one wrong, it's worth walking through why the other options are specifically incorrect, not just naming the right one.

Wrap-Up & Extension

Close with something like: "AI is becoming embedded in essentially every domain of scientific research — not as a replacement for the scientific method, but as an accelerant for specific steps within it, especially pattern discovery and hypothesis generation. The experimental validation step — where a claim meets physical reality — remains stubbornly human. Whether that stays true as AI-driven robotic labs mature is one of the more interesting open questions in how science itself will change over your careers."

Extension activity — "Design a Self-Driving Lab Constraint." Have students work in pairs to sketch a proposal for what oversight or constraints they would build into a real self-driving laboratory before letting it run autonomously on a research question of their choosing (drug candidate screening, new battery materials, whatever interests them). Require each proposal to address: what the AI is allowed to decide on its own versus what requires human sign-off, how often a human reviews its choices, and what happens if the lab proposes an experiment that's expensive, dangerous, or ethically questionable. Have a few pairs present and let the class critique each other's guardrails. This turns the correlation-vs-causation and autonomy themes of the lesson into an applied design exercise, and comfortably stretches the core lesson into a full 30–40 minute class period.

For students considering a STEM or research career, this is also a natural moment to mention that "AI for science" is an active, fast-growing subfield in its own right — sometimes called scientific machine learning or AI4Science — spanning astronomy, climate modeling, materials science, and computational biology, with dedicated labs and research groups at major universities and companies. If a student expresses interest, the honest framing worth giving them is that the strongest people in this field tend to have depth in an actual scientific domain (physics, chemistry, biology) alongside their machine learning skills, not machine learning skills alone — the domain expertise is what lets someone recognize when a model's output is scientifically plausible versus a plausible-looking artifact of the training data.