Cross-university platform portalShow a desktop notification when the AI TA finishes replying
Notify me when classmates post messages in the forum
Play an alert sound whenever there is a new notification
The hardest parts of a science lesson to teach are often the parts students cannot see. This document explains why the Science Laboratory simulations are designed the way they are, which learning difficulties they address, how they can be brought into the classroom, and which tasks AI assistants should not take over.
The core concepts of the natural sciences can, in most cases, not be directly observed within the time and spatial scale of a single lesson. The Coriolis force that drives typhoon paths, adiabatic cooling that forms clouds, planet-scale three-cell circulation, million-year plate movements, and atomic-scale orbitals are all examples. Students are not unwilling to understand; they simply have nothing they can see or modify with their own hands.
This makes natural science one of the fields in the 12-year national education system that relies most heavily on representational mediation: teaching must first turn inaccessible mechanisms into some visible form before Students can reason about them. Traditionally, this role has been played by static diagrams, formulas and board work, at the cost of Students only being able to accept predetermined representations and being unable to ask “what if it were not like this?”
What the operational simulation changes is not "making the phenomenon look better", but turning parameters into something you can change. When Students can switch gravity to the Moon's, set damping to zero, or turn the Coriolis force off and on again, the mechanism ceases to be a paragraph of description and becomes something that can be questioned.
In a co-design forum with secondary school Earth Science teachers, the scarce resource identified by teachers was not content, but rounds of feedback: a teacher's marking capacity is usually only enough for one round, so students do not revise. In the simulation design we put this upfront; Section 4 explains how it affects the role configuration of the AI TA.
Each simulation page consists of four parts; if any one is missing, inquiry will fall back to mere viewing.
| Component | Function |
|---|---|
| Interface | Expose the key variables of the concept as adjustable parameters, with students setting the initial conditions themselves |
| Live charts | Presenting phenomena and quantitative results together enables students to connect "what they see" with "what they measure" |
| Experimental task | Built-in inquiry prompts on the page point to areas worth following up, rather than giving answers |
| Data record | The Student records the parameters and results of each experiment independently, and can export them to continue analysis in a spreadsheet |
All simulations are performed in the Student's browser using JavaScript numerical integration and do not involve the server. This has three teaching consequences: zero response latency, so Students can try repeatedly without waiting; no login is required, so Teachers can paste the link directly into teaching materials; and the simulation itself does not call a language model or write to the database, so there is no token cost and the whole class can use it simultaneously without generating any expense.
The only exception is the global wind globe. It is not a numerical simulation, but a real data viewer: it periodically captures snapshots of wind fields in global atmospheric patterns, and presents the actual flow field near the surface, in the jet stream layer, and in the upper-air flow layer through particle animation. This exception has a pedagogical purpose in Section 3.
The most direct use is to have students find out when a ruleno longer applies. The pendulum period formula holds at small angles; simulation allows students to increase the angle themselves to the point where it begins to deviate, rather than being told “large angles are not applicable”. Resonance curves work the same way: scan the driving frequency and watch the amplitude surge near the natural frequency, which is closer to physics than memorising the definition of resonance.
The teaching value of tasks like this is that the scope is no longer a footnote in the textbook, but a result measured by the students themselves.
'Change only one variable at a time' is one of the most often stated, and hardest to put into practice, principles in inquiry-based teaching. In a real laboratory, changing a single variable is often constrained by equipment and time; in simulation, it is an action that can be carried out immediately and repeated twenty times.
More importantly, this action leaves a trace. The system records which parameters the Student changed, how many times they ran it, and what data they recorded, so whether they really controlled the variables no longer has to rely only on self-report or Instructor monitoring, but can be seen from the Learning history.
Most teaching simulations provide only idealised models, while most data entry points provide only real data. The Science Laboratory deliberately joins the two on the same route: students first see real, complex flow fields that cannot be reduced to a single formula in the global wind field globe, then move on to an idealised cyclone simulation centred on the Coriolis force, where only a small number of parameters are retained, making it operable, repeatable and derivable.
The gap between the two is itself the teaching content. When students see the irregularity of a real wind field, and then the symmetrical cyclone of an ideal model, 'what the model omits, and why it may be omitted' becomes a question that can be asked. This is precisely the entry point for teaching the nature of science, and also the core concern of model-based learning.
The suggested order for juxtaposition is from the real to the ideal. If students first describe the complexity of the real flow field and then move on to the model, they are more likely to realise that a model is a choice; if they see the model first, they are more likely to mistake idealised symmetry for the way nature really is.
The sidebar on each simulation page has an AI teaching assistant. Its design principle is 'state up, instructions down': the simulation runs entirely in the Student's browser; each time the Student speaks, the system sends the current simulation state to the assistant as well. The assistant can therefore see the experiment in front of the Student, ask about specific parameters, and also return instructions to set parameters, run and reset, which the browser then executes on its behalf.
This solves a common failure mode of conversational TAs: the parameters the TA talks about do not match the simulation on the student's screen.
The TA does not draw charts on behalf of the Student, does not write explanations on behalf of the Student, and does not directly determine the Student's grade. The basis for the first two items comes from the instructor's explicit requirements: generating and interpreting representations themselves are learning objectives, and handing them over to machines is equivalent to outsourcing what needs to be learned. The third item is because grading requires the instructor's judgement of the course context; AI's role is to provide directions for the next round of revisions.
What the TA is truly expected to supplement is the scarce resource mentioned in Section 1: rounds of feedback. After Students have drawn the diagram and written the explanation, the TA can continue to probe this piece of work, giving the Student a second and third chance to revise, which is capacity that a single instructor's marking load is often unable to provide.
Inquiry and practice is a compulsory subject prescribed by the national curriculum in natural sciences, and also the most scarce part of the teaching materials. In the co-design forum, Instructors repeatedly pointed out that there are still few ready-made materials available for this course component, yet it is exactly where simulation and real data can be used to best effect.
Simulation can cover the first half of inquiry: forming questions, planning variables, and generating data. The real-time environmental data provided by the platform can cover the second half: handling raw real-world data, judging data quality, and writing evidence-based explanations. The tasks that the teacher assigned in the Forum without prompts can be used directly as design examples:
The common feature of these designs is: the questions have real answers, but the answers are not in the textbook, and the noise in the data is itself part of the content.
The simulation has no measurement error, no instrument calibration, no poor contact, and no accidents. It trains model reasoning and control of variables, not laboratory skills. Real hands-on work cannot be replaced; the role of simulation is to build expectations before doing, or to explain after doing why the results differ from expectations.
Each simulation made choices: which variables to keep and which to ignore. If teaching does not make this explicit, students may take the model's behaviour as natural phenomena. The juxtaposition design in Section 3.3 is precisely to address this risk, but it requires the Instructor to state it clearly in class.
The science laboratory currently provides teaching design and process records, and has not yet accumulated enough data to support claims of effectiveness. Pre- and post-tests and process analysis are still at an early stage; this document describes the design rationale and possible uses, not verified outcomes.
Generative AI still makes mistakes in fields such as earth science. In the Forum, instructors have made “check first, then use” a prerequisite for adoption, and pointed out that if the platform is handed to students for unsupervised use without checking, students may receive incorrect information. This reminder also applies to the AI TA on the simulation page: instructors are advised to run through it once themselves, confirm that the simulation and TA perform appropriately for the concepts they plan to teach, and only then hand it over to students.
| Discipline | Number of simulations | Examples of scope |
|---|---|---|
| Physics | 46 | Mechanics, waves, electromagnetism, modern physics |
| Earth Science | 31 | Atmosphere, solid Earth, hydrosphere, astronomy |
| Chemistry | 10 | Reaction rate, equilibrium, solution properties |
| Biology | 9 | Physiological regulation, groups, genetics |
All simulations can be used without logging in; the AI Experimental TA requires login and has usage rate limits.