English:AI for science and environmental observation - AI Year 5
Introduction

Can a model learn a simple pattern without learning a reliable weather forecast? Fit one tiny classifier, trace six test cases and compare a baseline. Keep observations, predictions and uncertainty separate. All task data are invented; the credited media provide context.
From observation to a testable question
What was observed and what is only predicted?
A class wants to understand how a prediction can help organise an environmental observation project without replacing evidence. The learning question is: can a very small model use an earlier cloud-card score to predict whether a later invented rainfall record says RAIN or DRY? The intended users are a teacher and learners examining errors, not people deciding whether it is safe to travel or go outdoors.
An observation records what was actually seen or measured under stated conditions. A prediction estimates an outcome that is not yet known. An inference is an interpretation drawn from evidence. The same number can have different roles depending on when it was obtained. A later rain-gauge result is a target for our next-hour task; supplying that result as an input would give the answer away.
The water-cycle diagram supplies physical context: evaporation and transpiration add water vapour to the air, condensation is involved in cloud formation, and precipitation returns water to the surface. The diagram is not a record of six days and does not establish that every cloudy sky produces rain at one particular location in the next hour. Trace its arrows, then distinguish a process diagram from a time-stamped observation table.
Define our invented observations precisely
All F and T records below are original fictional classroom examples. No actual sky was photographed for this dataset, no real gauge readings were collected and no weather service supplied the labels. Imagine a teacher-drawn sky card divided into eight cells. Cloud_cells is the number of completely shaded cells, an integer from zero to eight. This simplified card convention is not a calibrated cloud sensor or a complete meteorological observation.
The target later_rain is RAIN when the separately supplied fictional next-hour record has measurable rain, and DRY when it records no measurable rain. Here those labels are simply part of the invented worksheet. In a real study the measurement method, instrument resolution, location and exact start/end times would need to be defined. A missing measurement must not be renamed DRY.
The Met Office explains that rainfall accumulation is measured over a stated interval with suitable equipment, and that cloud amount can be estimated in eighths. Our eight-cell cards are a deliberately simpler classroom representation. Their scores and labels must not be presented as official Met Office data or a genuine weather forecast.
| Field | Role | Definition | Missing-value rule |
|---|---|---|---|
| record_id | Metadata | Unique invented F or T record name | Required; not a weather feature |
| cloud_cells | Feature | Whole number from 0 to 8 on the earlier fictional card | None means unknown; output ASK |
| later_rain | Target | RAIN or DRY in the separate fictional next-hour record | Unavailable when making the prediction; do not guess it |
| split | Evaluation metadata | F is fitting data; T is held-out test data | Keep the split fixed before checking test results |
A short diagnostic before the experiment
Which is an input available earlier: the cloud-card count or the later rainfall label? Explain why a confident prediction is not an observation of the future. If the distinction is unfamiliar, draw a timeline with an earlier card, a prediction box and a later result. Put the actual target after the prediction box and keep it covered while tracing the method.
A teacher can provide printed cards, a large-text table and read-aloud instructions. The task works without a camera, pupil AI account, location tracking or access to the school grounds. Use paper first; the complete optional program later implements the same finite experiment. Completing this AI enrichment route supports data reasoning but does not certify practical weather observation.
Fit and trace a tiny prediction model
Fit a very small model using six training cards
Our model has one adjustable number, a threshold. If cloud_cells is at least that number, it predicts RAIN; otherwise it predicts DRY. An unknown feature produces ASK so that a person can review it. A model is not useful merely because it is called AI: examine exactly what changes during fitting and what evidence is available.
Before seeing the six test labels, compare only the candidate thresholds three, five and seven on the training table. Count a match when the prediction equals the supplied training label. Select the candidate with the most matches; if candidates tie, our declared tie rule chooses the smaller threshold. This tiny data-fitted classifier illustrates a machine-learning idea. It is not a modern operational weather model.
| Training ID | cloud_cells | later_rain |
|---|---|---|
| F1 | 0 | DRY |
| F2 | 2 | DRY |
| F3 | 4 | DRY |
| F4 | 5 | RAIN |
| F5 | 6 | RAIN |
| F6 | 8 | RAIN |
With threshold three, F3 is wrongly predicted RAIN: five of six labels match. Threshold five matches all six training labels. Threshold seven incorrectly predicts DRY for F4 and F5: four labels match. The fitting procedure therefore selects five. Freeze that choice before uncovering test labels. A perfect fitting score is not evidence that all future skies will behave like the invented cards.
CodeAI: How AI Works: What is Machine Learning? With teacher support, compare learning a pattern from labelled examples with simply following a rule chosen in advance. The wider publisher unit is for older grades; our concrete paper experiment needs no live pupil AI service. The film does not supply or validate these original weather records.
Trace six separate test cases
The test data are separate from the fitting cards. In the classroom, keep the target column covered until predictions have been written. T2 deliberately contradicts the learned pattern and T4 supplies another failure: synthetic test cases should help expose a limitation rather than guarantee success.
| Test ID | cloud_cells | Later supplied label | Frozen-model output | Comparison |
|---|---|---|---|---|
| T1 | 0 | DRY | DRY | Correct |
| T2 | 4 | RAIN | DRY | Wrong |
| T3 | 5 | RAIN | RAIN | Correct; threshold boundary |
| T4 | 8 | DRY | RAIN | Wrong |
| T5 | MISSING | RAIN | ASK | Abstained; not a correct prediction |
| T6 | 6 | RAIN | RAIN | Correct |
For T1, zero is a valid count and falls below five. T2 has four shaded cells, so the frozen model still predicts DRY even though the later label says RAIN. T3 is exactly five and therefore takes the RAIN branch. T4 shows why a fully shaded card cannot guarantee local rainfall. T5 has an unknown feature, which is different from zero; ASK is a request for review, not a weather category. T6 is above the boundary and matches its RAIN label.
Read and run the complete optional reference
The following Python program fits the three candidates and prints the six predictions. Its output was actually checked during course preparation. Learners may trace it on paper; execution requires a teacher-approved Python environment. It calls no online model, collects no personal information and cannot fetch actual weather.
def predict(cloud_cells, threshold):
if type(threshold) is not int or not 0 <= threshold <= 9:
raise ValueError("threshold must be an integer from 0 to 9")
if cloud_cells is None:
return "ASK"
if type(cloud_cells) is not int or not 0 <= cloud_cells <= 8:
raise ValueError("cloud_cells must be an integer from 0 to 8")
return "RAIN" if cloud_cells >= threshold else "DRY"
def fit(training):
if not training:
raise ValueError("training must not be empty")
for cloud_cells, label in training:
if cloud_cells is None or label not in ("RAIN", "DRY"):
raise ValueError("training needs a valid feature and label")
predict(cloud_cells, 5)
candidates = (3, 5, 7)
scores = {
t: sum(predict(x, t) == label for x, label in training)
for t in candidates
}
best = min(candidates, key=lambda t: (-scores[t], t))
return best, scores
training = [(0, "DRY"), (2, "DRY"), (4, "DRY"),
(5, "RAIN"), (6, "RAIN"), (8, "RAIN")]
threshold, scores = fit(training)
test_inputs = [0, 4, 5, 8, None, 6]
print("threshold", threshold, "training_scores", scores)
print("test_predictions", [predict(x, threshold) for x in test_inputs])Expected output: threshold 5 with training scores 3:5, 5:6, 7:4; test predictions DRY, DRY, RAIN, RAIN, ASK, RAIN. The displayed dictionary punctuation follows Python's normal printing. Invalid feature values such as minus one, nine, a word or a decimal are rejected rather than silently rounded; unknown is represented separately by None. For the frozen threshold, four predicts DRY and five predicts RAIN.
Guided practice before changing anything
Cover the output column and trace all six tests independently. Mark correct, wrong or ASK after consulting the supplied targets. Then predict the result of a new card with seven shaded cells without inventing its unknown later label. Your model output is RAIN, but without a later observation you cannot say it was correct.
Partial step: T3 uses the inclusive boundary, so five belongs on the RAIN side. Finish T2, T4 and T5 and explain their different problems. Do not improve your reported test score by secretly changing five after inspecting test labels. Any revised method needs its own new held-out evaluation and a record of why it changed.
Evaluate evidence and keep human control
Compare with a simpler baseline and show every outcome
A baseline is a simpler method used for comparison. Before testing, declare an always-DRY baseline. It matches T1 and T4, but misses T2, T3, T5 and T6: two correct and four wrong. The fitted model gives three correct, two wrong and one ASK. Report all six cases; dropping ASK would hide the need for a human check.
The model therefore has three correct out of six total cases. Among its five actual RAIN/DRY predictions, three are correct. Those two denominators answer different questions and must be named. The baseline has two correct out of six. This small invented comparison does not establish dependable weather skill, a real probability of rain or superiority across different regions and seasons.
If a person later checks T5, report that separate assisted decision and its evidence. Do not retroactively count the model's ASK as a correct rain prediction. A useful human-review route gives the teacher the original card, missing-value flag and observation timeline, plus authority to reject the output. An approval button without evidence is not meaningful oversight.
A weather prediction is not a long-term climate summary

This credited graph summarises historical monthly temperature patterns for south-west England. Its source page describes a 1971–2001 record basis. It does not give the date, local rain label or earlier cloud-card count for T1 to T6. An average July temperature is not a prediction that tomorrow at one school must have that exact temperature or no rain. Use a dataset with the right quantity, place and time scale for the question.
The graph supports a visual reading task: name the horizontal variable, vertical unit and two types of lines; describe the seasonal shape without inventing a particular station value when the station-to-line mapping is not supplied in the legend. A broad regional summary cannot replace a specific, dated observation record. Do not mix its temperature units with the card count's unitless score.
Explain two limitations and propose real checks
First, our feature describes only a simplified earlier card. It omits important weather information and cannot encode a whole changing atmosphere. A proportionate response is to limit the claim to the paper task, examine errors, and consult an appropriate official forecast for a real weather question. Do not invent a physical rule that more shaded cells always causes rain in the next hour.
Second, six fitting cases and six tests cannot represent every place, season, instrument fault or unusual weather event. Plan a new, separately documented evaluation using appropriate, permissioned records and expert supervision before making a broader claim. Merely copying these twelve rows many times would not add new independent conditions. Keep an untested future case, such as an ambiguous partially shaded card, explicitly untested in the evidence log.
CodeAI: AI: Training Data & Bias. Ask whose locations, instruments or conditions an evaluation leaves out. The film is contextual; a better aggregate worksheet score alone does not prove fairness or environmental reliability. Review the actual use and consequences, not only the technology's name.
Transfer responsibly to a biodiversity question
A class could later investigate an original, fictional habitat-card dataset instead of rainfall. State its features and target anew. For example, classifying a card's invented habitat label is different from predicting a later count of organisms. A label guessed from a photograph is not automatically a verified species identification. Record uncertainty and seek an appropriate teacher or expert check rather than assigning a real organism a confident name without evidence.
Neither project requires pupils' faces, home addresses or precise personal locations. Public visibility does not automatically grant reuse rights; check licences, provenance and school approval. Our lesson needs only the supplied fictional records and credited media. A separate outdoor investigation would need staff planning for access, weather and environmental care; this paper model never authorises entry into water, collection of wildlife or decisions about dangerous conditions.
Retain the user question, dictionary, training choices, frozen threshold, six-case comparison, baseline, errors, missing-value route and two limitations. Add a genuinely untested future case and a teacher-review decision. This is editorial AI Year 5 enrichment linked to Computing and scientific reasoning; it does not constitute a separate statutory annual AI requirement or prove real-world forecasting competence.
Interactive Tasks
Quiz: Test Your Knowledge
Which information is an earlier feature in the fictional task? (The shaded-cell count on the earlier cloud card) (!The already known next-hour target label) (!The learner home address) (!The final test score)
What does a later_rain label describe? (The separately supplied fictional next-hour outcome) (!A guaranteed future result for every real school) (!The name of the training program) (!A precise pupil location)
Which candidate threshold is selected from F1 to F6? (Five) (!Three) (!Seven) (!Eight)
How many training matches does threshold three produce? (Five out of six) (!Six out of six) (!Four out of six) (!Zero out of six)
What does the frozen model output at exactly five shaded cells? (RAIN) (!DRY) (!ASK) (!A guaranteed real forecast)
What happens when cloud_cells is unknown? (The model outputs ASK for review) (!The program silently treats it as zero) (!The target is automatically DRY) (!The missing value proves rain)
How many of all six tests does the fitted model get correct? (Three) (!Five) (!Six) (!Two)
What does the always-DRY baseline get correct? (T1 and T4) (!T2 and T3) (!All six tests) (!Only the missing-feature case)
Why cannot the historical temperature graph supply T3 rain truth? (It describes a different quantity and time scale) (!Every line graph is false) (!Degrees Celsius count shaded cloud cells) (!A monthly average guarantees next-hour rainfall)
What would improve the evidence for a revised model? (A new appropriate held-out evaluation with documented sources) (!Copying the same twelve rows many times) (!Hiding every ASK outcome) (!Tuning on test labels and calling them unseen)
Memory Game
| Observation | Evidence recorded under stated conditions |
| Prediction | An estimate of an outcome not yet known |
| Feature | An earlier permitted input to the stated task |
| Target | The separately supplied outcome being predicted |
| Baseline | A simpler method declared for comparison |
| Abstention | A recorded refusal to make the ordinary prediction |
Drag and Drop
| An unknown cloud-card score | Keep it unknown and request review |
| An earlier card count of exactly five | Follow the inclusive RAIN branch |
| A later result already in the input | Remove the leaking answer from prediction inputs |
| A new threshold chosen after reading test labels | Use a new held-out evaluation before a broader claim |
| A monthly temperature summary | Read it at its stated quantity and time scale |
Crossword Puzzle
| RAIN | What is the wet target category in the fictional experiment? |
| FEATURE | What role does cloud_cells have? |
| TARGET | What role does the later outcome have? |
| BASELINE | What is the simpler comparison method called? |
| MISSING | What describes an unknown feature rather than zero? |
| EVIDENCE | What supports a scientific conclusion? |
LearningApps
Optional extra practice. Choose an appropriate activity with your teacher. If the external collection is unavailable, use the native activities and independent tasks in this course.
Cloze Text
Open Tasks
Easy
- Draw the timeline: Place the earlier feature, prediction and later label in order. Explain why the target is covered.
- Trace the boundary: Compare cloud-cell counts four and five with the frozen threshold.
- Read unknown correctly: Explain the difference between zero shaded cells and MISSING.
- Read a real graph: Identify months, temperature units and the minimum/maximum legend without using it as a rain record.
Standard
- Fit the three candidates: Trace F1 to F6 and calculate each candidate score before selecting a threshold.
- Audit six tests: Record predictions and compare with supplied labels, retaining wrong and ASK cases.
- Compare the baseline: Trace always-DRY and name the denominators used in your comparison.
- Keep an evidence log: Record synthetic provenance, field meanings, frozen split, limitations and an untested future case.
Advanced
- Challenge a confident conclusion: Explain why six training matches cannot justify real next-hour forecasting.
- Design meaningful human review: State what the teacher needs to inspect and when they can reject an output.
- Plan a fair next evaluation: Keep new evaluation cases separate from model revision and define a missing-value rule.
- Transfer to habitat cards: Redefine features and target while avoiding unsupported real species identification and private data.
Learning Check
- Observation and prediction: Explain the timeline and target.
- Training scores and choice: Fit candidates three, five and seven.
- Six-case trace: Give outputs and compare with the test labels.
- A baseline and two denominators: Compare the model with always-DRY.
- Limits and an untested case: Name two limits and proportionate responses.
- Transfer and source choice: Apply the evidence process to a habitat-card investigation.
Worked checks and assessment guidance
- Cloud_cells is an earlier fictional card score. Later_rain is a separately supplied next-hour label, hidden while predicting. An observation records evidence under stated conditions; the prediction estimates an unknown outcome. The diagram and historical graph do not supply the invented targets.
- Threshold three misses F3 and matches five of six. Threshold five matches six of six. Threshold seven misses F4 and F5 and matches four of six. Select five using only the training cards, then freeze it. No candidate tie occurs in these six cases.
- T1 DRY correct; T2 DRY wrong; T3 RAIN correct at the inclusive boundary; T4 RAIN wrong; T5 ASK because the feature is unknown; T6 RAIN correct. Totals are three correct, two wrong and one abstention. The unknown input is not zero.
- Always-DRY matches T1 and T4: two correct and four wrong. The model gives three correct among all six records, or three among its five RAIN/DRY predictions. State the denominator and disclose ASK. A later human answer is separate from the model output.
- The simplified card count omits important weather information: keep the classroom claim narrow and use appropriate official sources for real questions. The small invented dataset lacks broad conditions: plan a separate, permissioned evaluation. A partially shaded-card trial is proposed but has not been performed here. Copying records adds no new independent observations.
- Define new features and target, provenance, units and missingness; do not infer verified species identity merely from an attractive picture. Preserve fitting/test separation, baseline, error log and human authority to stop. Check media permissions and school policy, and avoid personal locations. A historical temperature summary is not interchangeable with next-hour rain labels.
Learning evidence
Keep the user question, synthetic source version, dictionary, all three fitting scores, frozen threshold, six-case trace and baseline. Disclose three correct, two wrong and one ASK; name denominators. Explain two limitations with proportionate checks, one genuinely untested future case, media permissions and meaningful teacher review. Optional code output is not practical weather-observation certification.
OER and Wikipedia
DfE Computing anchors the programming objective; DfE Science supplies scientific-enquiry context. This annual AI route and its synthetic experiment are editorial, not a new statutory subject. Google: supervised learning was consulted for fitted models and unseen evaluation. Met Office: measuring weather supports cloud/rain observation context. All classroom records, program and worked cases are original. Reviewed 15 November 2026.
Water cycle.svg, j4p4n (NASA Precipitation Measurement Missions), Public domain. Unchanged original; contextual media, not the source of the fictional data.
Graph of monthly temperature variations in south-west England.png, Jolly Janner, Public domain. Unchanged original; contextual media, not the source of the fictional data.
https://www.youtube.com/watch?v=OeU5m6vRyCk — CodeAI: How AI Works: What is Machine Learning? Original recording embedded unchanged.
https://www.youtube.com/watch?v=x2mRoFNm22g — CodeAI: AI: Training Data & Bias. Original recording embedded unchanged. Teacher-supported viewing; no live pupil AI account required.
Linked learning areas
aiMOOC Projects
Schulfach+


aiMOOCs


aiMOOC Projekte


MOOCwiki · Deutsch
Nach dem Lernen ist vor dem Lernen
Entdecke direkt den nächsten Lernkurs. Weitere Inhalte erscheinen, wenn Du weiter nach unten scrollst.
Zur MOOCwiki-HauptseiteMediathek
Mediathek wird aus dem Wiki geladen ...
Keine passenden Inhalte gefunden. Bitte ändere Suche oder Filter.
NEWSLernweltNOAH fragen