The Data of Breaking Things in Class

A course built on destructive testing turned into a data instrument, and the record of how novices expect failure became the most valuable thing in it.

Framing

The course started as an opportunity rather than a problem. My department had room to grow, and the obvious way to reach students who would never select into a fabrication program on paper was to build something with no prerequisites that demonstrates the discipline rather than describing it.

The premise is destructive testing. We break things in class. We pull them apart, twist them, bend them, vibrate them, and push three times rated current through electrical devices until they fail.

The interest was there and enrollment was strong, which met the original intent. What follows concerns something I did not plan for, which is that the course turned out to be a data generating instrument, and that the data became the most valuable thing in it.

The learning objective

The stated outcome is not a body of material knowledge. It is a disposition I have been calling intuitional engineering: forming a fast, low confidence estimate of how a physical system will behave, treating that estimate as a hypothesis rather than an answer, and then checking it against reality.

Does this look right. Does this sound right. What was your first instinct before you talked yourself out of it. If something feels wrong, that discomfort is signal rather than noise, and it should send you to a test.

Destructive testing suits this because it closes the loop faster and more definitively than anything else available to a first year student. A prediction about failure is unambiguous and it resolves within minutes, in front of the room, with a sound.

Recognizing the dataset

The prediction protocol existed first for pedagogical reasons. Before each test, students commit to a number: at what load, at what deflection, at what current will this fail, and where. The requirement forces commitment, and a student who has not committed to a number learns considerably less from watching something break.

Halfway through the first offering I recognized what that protocol was producing. Every session generates paired records. A distribution of student predictions, and a measured actual. Accumulated across a term and across specimen types, that is not one dataset but two.

The first is a record of how things fail. Useful, and the obvious one.

The second is a record of how a cohort of novices expects things to fail, and therefore of where novice intuition is systematically wrong rather than randomly wrong. That is the one I did not anticipate having, and it is the more valuable of the two, because it is a direct measurement of the thing the course exists to change.

I want to be precise about the sequence, because I think it generalizes. I did not design a data collection instrument. I designed a teaching protocol that happened to be one, and only noticed after the records had piled up. I suspect this is common. Most hands on courses are producing paired prediction and outcome data already, informally, in students' heads, and discarding it at the end of each session.

Making the data usable

Collection alone accomplished nothing. For most of the first term the data existed as raw records, which is functionally the same as not existing, because most of my students are visual learners and a table of paired predictions and outcomes teaches them very little.

The distance between raw records and a usable classroom visual is almost entirely tedium. Cleaning, structuring, plotting, iterating on a chart that does not quite communicate, plotting again. None of it is conceptually difficult and all of it consumes the hours an instructor does not have. That gap is the reason most course generated data stays in a notebook.

I used an AI model to pass through the data and produce charts. This is the least glamorous use of these tools I can describe, and I think it is among the most broadly applicable in teaching. It did not analyze anything I could not have analyzed. It removed the cost of iteration, which changed what was practical to put in front of students, which is the entire point.

What the visualization had to do

The charts required more design attention than I expected, and the requirements were pedagogical rather than aesthetic.

Horizontal bars, sorted strongest to weakest. Sorting does the interpretive work before the student begins reading, and the ranking is the finding in most cases.

High contrast, dark ground, minimal ornament. The visuals get projected in a shop, viewed at distance, often in bad light.

A clear reference line for the predicted value against the measured value. This is the element that matters most and the one I underweighted at first. A chart of failure loads teaches material behavior. A chart of failure loads with the class prediction marked on it teaches the student something about themselves.

The test I settled on is that a student should be able to answer one question in under a second: which of these was I wrong about, and by how much. Every design decision either serves that question or gets removed.

What the data showed

The result that most changed how I talk about design is the performance gap between gussets and fillets on 3D printed parts.

Both are conventional ways to reinforce an interior corner. The fillet is the more intuitive choice and the more common default, and it is what students reach for. In testing, gusseted geometries substantially outperformed filleted ones.

The mechanism is not mysterious once observed. A fillet distributes stress across a curve, which works well in homogeneous material and much less well in a part built from bonded layers with directional weakness at every interface. A gusset places material along the load path in a way that does not depend on the corner geometry itself carrying the load.

But the finding is not the pedagogical event. The pedagogical event is that the room predicted otherwise, and can see that it predicted otherwise, because the predictions were written down before the test and plotted alongside the result. Without the collection protocol, this is a fact students are told. With it, it is an error they watched themselves make.

What the data did to the course

Two effects, both of which I attribute to visualization rather than to collection.

The first is calibration. Once prediction error is plotted rather than merely remembered, the error becomes an object of study in its own right. The question shifts from what broke to why my model of what breaks is wrong in this specific direction. That is the course objective operating on itself, and no mechanism for it existed before the charts did.

The second is that students left believing data is not boring. They arrived with data framed as the tedious part, the thing that comes after the interesting part. Having generated it themselves by breaking things, with their own predictions embedded in it, they met it as the record of an argument they had personally won or lost.

I suspect that reframing transfers further than any specific material result in the course. A student who finds data interesting will keep collecting it.

Limitations

The measurements are teaching grade. Specimens are not statistically controlled, sample sizes are small, conditions vary between sessions, and nothing here should be used to select material for a load bearing application without independent verification. I say this to students explicitly and I say it here for the same reason.

The prediction data has a contamination problem. Predictions are made in a room, in proximity, and students influence each other. The distribution is therefore not a clean measure of independent novice intuition, and I would not make a strong claim from its shape.

The course has run a limited number of times. The gusset result is consistent across the specimens tested and consistent with the stated mechanism, but it is not a controlled study of print geometry.

Three things seem worth pursuing.

Whether prediction accuracy measurably improves across a term, and whether improvement generalizes to specimen types a student has not seen before. The data structure to test this already exists, and it is the direct measure of whether intuitional engineering is being taught or merely described.

Whether the visualization is doing the work I am attributing to it. I introduced charts and observed changes in engagement and calibration, but I introduced them to everyone at once and have no comparison group.

Whether this generalizes past destruction. Destruction is vivid and it commits the student. It may be that any sufficiently unambiguous fast resolving test produces the same paired data and the same effect, in which case the approach extends well beyond a shop.