Teaching Engineers to Act on Uncertain Recycled-Concrete Predictions

Concrete made from crushed demolition rubble is weaker and less predictable than concrete made from quarried stone. The main reason is a porous crust of old cement paste stuck to each recycled particle. That crust soaks up mix water and leaves a weak boundary layer around the aggregate. Machine-learning models can now predict the strength of such concrete with prediction intervals, meaning error bars with a stated coverage rate. Those models also carry known traps. Their training data contain spurious correlations, and their uncertainty can be made to look smaller than it really is. Engineers who must decide whether a recycled mix is safe for a structural member face two opposite ways to fail. They can trust a confident but wrong prediction and build something unsafe. Or they can reflexively specify virgin aggregate, which wastes resources and keeps recycled material out of construction. We propose a teaching module that places the ethics of that decision inside the technical work. Students measure how much water real recycled aggregate absorbs. They run a strength predictor that reports intervals and audit it for planted flaws. Then they make and defend accept-or-reject specifications. We grade those specifications against measured strengths that the students never see.

The opportunity is that this setting gives ethics education something it rarely has: an objective score for the decisions themselves. Most integrated-ethics studies measure attitudes or immediate learning gains in computing or mechanical courses. Here, every student decision about a held-out concrete mix can be checked against a measured laboratory strength. That lets us count how often students accept an unsafe mix and how often they reject a mix that would have been adequate. Both counts are concrete ethical outcomes. The first is a public-safety failure. The second is the excess conservatism that industry experts name as a main barrier to recycled materials.

Core Claim

The ethical substance of specifying recycled aggregate lies in how an engineer acts on porosity-driven uncertainty. Abstract professional codes and deterministic carbon-versus-cost trade-offs do not reach it. Our claim is that teaching this through an audit-and-specify workflow will improve students' two-sided decision quality more than three alternatives would: a standalone ethics unit, a direct port of sustainability problem-based learning with point predictions, or a purely technical module on uncertainty. Two-sided decision quality means fewer unsafe acceptances without more unwarranted rejections. The new element is the combination of a decision-scored assessment with deliberately planted model flaws drawn from the recycled-concrete literature. The study design separates the contribution of the ethics framing from the contribution of uncertainty literacy alone.

Grounding in Prior Work

The materials side supplies the mechanism and the data. Gu et al. built a random-forest model for flexural strength of paving concrete, tuned with the firefly algorithm. They found that recycled-aggregate water absorption ranked among the most important inputs (importance score 0.7835, second only to water-cement ratio) (Potential of Repurposing Recycled Concrete for Road Paving). This ties prediction uncertainty directly to porosity. It is why students in our module measure absorption themselves rather than receive it as a given number. Dong and Zhang predicted compressive strength of recycled aggregate self-compacting concrete on 289 literature records. They used a stacked combination of a transformer for tables and a masked neural network, plus split conformal prediction, a method that turns held-out errors into intervals with guaranteed coverage (Hybrid Deep Learning with Conformal Prediction for RASCC Strength Prediction). Their attribution analysis also exposed non-physical associations involving aggregate density. We reuse their dataset and model class as the student-facing tool, and we turn their confounding finding into a planted flaw students must find.

Treatment options define the design space students choose from. One is carbide-slag slurry soaking followed by carbonation, which fills pores with calcium carbonate (Mechanical Properties of Recycled Concrete with Carbide Slag Slurry Pre-Immersed and Carbonated Recycled Aggregate). Another is adding 10% zeolite powder, which improved freeze-thaw durability (Freeze–Thaw Cycle Durability and Mechanism Analysis of Zeolite Powder-Modified Recycled Concrete). Students can therefore answer "improve the aggregate" rather than only "accept" or "reject".

Two machine-learning papers constrain how the tool must be built. Min et al. showed that a conformal predictor can shrink its average interval width while keeping nominal coverage, by randomly returning empty sets on some inputs. They call this the "Prejudicial Trick", and they tested it on benchmarks including the standard CONCRETE dataset (Questioning the Coverage-Length Metric in Conformal Prediction). Their code (PT-Conformal-Prediction) lets us build a deceptively "efficient" competitor model for students to audit. Yeh et al. showed that calibrating uncertainty with awareness of downstream decisions beats estimating first and optimizing second. The gains were up to 42% in battery arbitrage (End-to-End Conformal Calibration). We take from them the principle that costs are asymmetric, and our scoring weights an unsafe acceptance more heavily than an unnecessary rejection.

The education side supplies the method and motivates the gap. Zhao et al. embedded sustainability criteria into a mechanical-vibration design project and reported gains over lecture instruction (Fostering Technical and Sustainability Competencies Through an Integrated PBL Approach). Their criteria were deterministic, they ran in a single discipline, and they measured outcomes only immediately after the course. Krakowski et al.'s "4D" problem-based curriculum (Determination, Data, Design, Decision) led high-school students to see that some tensions have no technical fix (Authentic Integration of Ethics and AI). Bender, Hovy, and Schofield taught ethics through hands-on exercises in which students directly observe harms that a technical pipeline produces (Integrating Ethics into the NLP Curriculum). We borrow that "observe the harm in your own pipeline" design and the 4D question structure, and we move both onto physical, safety-critical material data. Geng et al.'s expert interviews identify liability-driven defaults to virgin aggregate, public safety skepticism, and fragmented standards as adoption barriers (Overcoming Barriers to the Adoption of Recycled Construction Materials). This is why we score unwarranted rejection as a failure and not as a safe default.

Research Program

The module runs over roughly three weeks of an undergraduate concrete-materials or construction-materials course.

In the first week, students measure water absorption and bulk density on recycled aggregate batches. Some batches are untreated and some have received a simple pre-soak and carbonation treatment, when lab capacity allows. Students then see how these porosity measurements move a strength predictor's interval.

In the second week, students receive two predictors trained on the RASCC records. One is honest split conformal. The other is the Prejudicial Trick variant, which looks better on the average-width metric. Students also receive attribution plots in which the density effect contradicts concrete science. The task is to audit both predictors. Students are guided by 4D-style questions: should a model be used for this member at all? Which mixes and absorption ranges are represented in the data? What does the interval actually guarantee? Who bears the cost of each kind of error?

In the third week, each team specifies aggregate for a set of structural cases. For each case they accept, reject, or require treatment and testing, and they write a justification addressed to a skeptical client and a building official.

We will build three artifacts. The first is the open notebook set with both predictors and the planted flaws. The second is a pool of specification cases drawn from held-out RASCC records. The third is a grading rubric validated by inter-rater agreement.

The specification cases work as follows. Each case pairs a held-out mixture with a required strength set near its measured value. That way the correct answer is not obvious from the point prediction alone. Because the measured strength is known, each decision can be scored automatically as a safe acceptance, an unsafe acceptance, a necessary rejection, or an unnecessary rejection. A separate rubric scores the justification text. It checks whether the student named the interval's guarantee, flagged the planted flaws, weighed resource and carbon consequences, and identified who carries the risk. Asking for treatment and retesting on a borderline case is scored as correct.

Baselines

The naive reference point is the course as currently taught plus a one-session standalone engineering-ethics lecture on codes of conduct and recycled materials. Students in this arm complete the same specification test with only the point predictor and the code-style replacement guidance they would normally receive. This arm tells us how often students accept unsafe mixes, and how often they reflexively reject adequate ones, without intervention. There is no reported result found on Lacuna for decision error rates of civil engineering students on recycled-aggregate specification.

The strongest existing method is a faithful port of Zhao et al.'s sustainability-integrated problem-based learning to this course. It keeps their four pillars: authentic trade-offs, decision matrices, iteration with reflection, and collaboration. The design criteria become embodied carbon, cost, and point-predicted strength, with no uncertainty audit. Zhao et al. report Cohen's d of 0.61 for sustainability-oriented decision-making, plus d = 0.74 for conceptual understanding and d = 0.77 for mathematical proficiency, against lecture instruction. To beat this baseline, our module must reduce the unsafe-acceptance rate relative to the port without raising the unnecessary-rejection rate. It must also do so while matching the port on a shared conceptual concrete-materials test, so that the gain is not bought with lost technical learning. Both arms get the same contact hours, the same instructor-preparation budget, the same lab sessions, and the same cases.

The key ablation is a technical-only uncertainty module. It has identical notebooks, the same absorption lab, and the same planted flaws, and it teaches conformal prediction and model auditing as pure technique. It drops the stakeholder, Determination, and Decision questions and the client-facing justification. A second ablation removes the planted flaws from the full module. This tests whether students learn to audit only when they encounter a broken tool. For the predictor, the reference performance is Dong and Zhang's reported test R² of 0.940, RMSE of 4.219 MPa, and 90% coverage with an average interval width of 8.32 MPa. Our honest predictor should reproduce numbers in this range. Otherwise, students would be auditing a strawman.

Why not just…?

Why not just keep ethics in a standalone course? The integrated-ethics literature argues that bifurcated curricula produce students who do not apply judgment inside technical work. However, Zhao et al. compared integration against lecture instruction, not against a matched standalone ethics course. That comparison has not been run. Our naive arm runs it, on a decision-level outcome rather than attitudes.

Why not just port Zhao et al.'s problem-based design as is? This is a reasonable proposal, and it is our strongest baseline. Its weakness for this material is that its criteria are deterministic. In recycled concrete, the safety question depends on porosity-driven variability: absorption is a dominant feature in Gu et al., and Dong and Zhang's intervals are about 8 MPa wide. If the port matches our module on unsafe acceptances, the uncertainty component is unnecessary, and we will report that result.

Why not just teach conformal prediction and model auditing as a technical skill? This is the most serious threat. It may be that uncertainty literacy alone produces the decision gains and the ethics framing adds nothing measurable. The technical-only ablation settles the question. Our specific prediction is that the two arms will differ mainly on unnecessary rejections and on justifications. Students taught only technique may treat wide intervals as reasons to reject. Students who also weigh resource and carbon costs, and who consider treatment options, should more often prescribe treatment and testing. If the two arms do not differ, the honest conclusion is that the contribution reduces to a materials-specific uncertainty curriculum.

Why not just teach the code replacement limits? Prescriptive caps do not respond to measured absorption or to remediation such as carbonation. Geng et al.'s interviewees identify code bias toward virgin aggregate as itself a barrier. The specification cases include treated, high-replacement mixes that measured adequately. A limits-only approach would reject these, and the naive arm will show how often that happens.

Proposed Experiments and Decisive Checks

We will run a cluster-randomized study across course sections at two or more institutions, with four arms: the naive arm, the problem-based port, the technical-only ablation, and the full module. A no-planted-flaw variant of the full module will run where an extra section is available. Individual-level sample needs are about 64 students per arm to detect d ≈ 0.5 at 80% power and α = 0.05; this figure follows the standard two-sample calculation. Clustering will inflate the requirement, so section-level random effects will be modeled and the analysis pre-registered.

The primary outcome is a cost-weighted decision error on 20 held-out specification cases, with unsafe acceptances weighted more heavily, in the spirit of Yeh et al. The two error rates will also be reported separately. Secondary outcomes include three measures. The first is the detection rate for the Prejudicial Trick predictor and for the density confound. The second is rubric-scored justification quality. The third is the shared conceptual test, which guards against technical dilution.

A delayed retest, given eight to ten weeks later with new cases, addresses the retention gap that Zhao et al. leave open.

The decisive result has two parts. The full module must lower unsafe acceptances relative to the problem-based port, and it must lower unnecessary rejections or treatment-blind rejections relative to the technical-only ablation, with both effects still present at the delayed retest. If only the first holds, uncertainty literacy is doing the work. If neither holds, integrated ethics framing does not transfer to this material domain in a way that changes decisions. That would also be a useful negative result for a literature that so far reports mostly positive results measured immediately after a single course.