Students Build Factory Digital Twins to Learn Industry 4.0 Skills

A semester-long experiment at Tecnológico de Monterrey has tested a deceptively simple idea: instead of teaching Industry 4.0 as a series of lectures on buzzwords, hand engineering students a real factory problem and make them build the entire digital infrastructure needed to solve it. The result, published in Frontiers in Education, is a framework that walks undergraduates through every layer of a digital twin — from the first process map to an AI-assisted dashboard — and suggests that the act of building the twin, rather than the finished artifact, is where the learning actually happens. Sixty students, organized into eighteen teams, spent a semester constructing virtual replicas of a genuine food manufacturing process, and the majority of them made it through all five technical stages within a single course.

The industrial challenge came from a real partner in the food sector: the manual preparation of individually wrapped chocolate bars used in cake production. Workers were handling batches of roughly 350 units in half-hour cycles with two operators per batch, and the company wanted a semi-automated production cell that could raise productivity without sacrificing food safety or operational flexibility. That brief gave the students something rare in undergraduate coursework — an open-ended problem with real constraints, real economics, and no answer key. Each team had to analyze the existing process, propose an automation concept, and then prove, digitally, that their design would actually work before anyone cut a piece of metal.

The framework’s architecture is what distinguishes it from most educational simulation exercises. Students progressed through five complementary layers. First came process analysis and mapping, where teams documented material flow and identified bottlenecks. Second, conceptual automation design, where components were selected and layouts drafted. Third, discrete-event simulation in Siemens Tecnomatix Plant Simulation, where the proposed line was modeled and stress-tested numerically. Fourth, three-dimensional validation using Siemens Process Simulate and CAD tools to confirm that equipment fit, operators could reach their stations, and moving parts did not collide. The final layer brought the twin alive: SCADA-style dashboards built in Node-RED, cloud data exchange through Google Firebase, and decision-support recommendations generated through the Gemini API.

Completion rates across the cohort tell a story about where the difficulty concentrated. All eighteen teams finished the process-analysis, conceptual-design, simulation, and 3D-validation stages. Fifteen of eighteen completed the SCADA stage, and sixteen of eighteen completed the AI-assisted decision-support layer within the available course timeframe. The attrition at the top of the stack was not attributed to disengagement but to the sheer integration burden — five professional software platforms, none of which students had used before, all needing to talk to each other. The authors note that interoperability troubleshooting consumed time that might otherwise have gone into deeper engineering analysis, and they recommend introducing some tools earlier in the curriculum to flatten that learning curve.

The simulation results themselves were strikingly consistent across independent teams. Cohort-level outputs ranged from 720 to 850 units over the simulated horizon, with a mean of 789 and a standard deviation of just 38. More tellingly, fourteen of the eighteen teams independently identified the same fundamental problem: a capacity mismatch upstream of their proposed cells, visible as severe blocking at the source station. In the illustrative case documented in detail — a team whose members are co-authors of the paper — the source blocking rate reached 89.78 percent, meaning the upstream process was stalled nearly nine-tenths of the time waiting for downstream capacity. That convergence suggests the simulation layer was doing real analytical work, surfacing a genuine bottleneck rather than merely animating a design.

The detailed case study also shows how far students pushed beyond pure simulation. Their three-dimensional validation confirmed a minimum operator clearance of 600 millimeters at the loading position and verified that conveyor transport, cutter descent, and roller extraction motions were geometrically compatible. Component choices were grounded in commercial documentation — a food-grade sanitary belt conveyor, a vibratory bowl feeder for package orientation, a pneumatic cylinder for the cutting mechanism — selected against hygienic design requirements rather than picked at random. The team even connected their technical findings to engineering economics: an estimated automation investment of roughly 19,000 US dollars weighed against projected annual labor savings of about 8,900 dollars yielded an estimated payback period of approximately 2.1 years and a five-year return on investment near 134 percent.

Competency assessment relied on a structured rubric covering six dimensions: systems thinking, simulation modeling, industrial automation, interdisciplinary collaboration, engineering communication, and data-informed decision-making. Two instructors independently scored a subset of five team deliverables, achieving an inter-rater reliability of Cohen’s kappa at 0.78 — respectable agreement for a locally developed instrument. Teams were deliberately mixed, pairing biomedical engineering students, who tended to contribute process analysis and requirements definition, with industrial and mechatronics students, who gravitated toward automation design and controls. The authors flag honestly that this role specialization means they cannot confirm whether both disciplines developed each competency equally, since the rubric was applied at team level rather than to individuals.

The artificial intelligence layer deserves particular attention because of how it was positioned. Rather than letting students treat Gemini’s outputs as authoritative answers, instructors explicitly required them to validate every AI recommendation against simulation evidence and physical constraints. In the documented implementation, the AI layer generated context-appropriate responses to four distinct operational conditions — alarm grouping for repeated cycle-time anomalies, predictive maintenance alerts based on tool-wear estimates, numerical setpoint suggestions such as conveyor speed adjustments, and standard-operating-procedure support for compound abnormal states. The authors frame this as a deliberate design choice aligned with broader research cautioning that over-reliance on AI recommendations can erode independent engineering judgment unless the technology is nested inside human-supervised workflows.

The paper is unusually candid about its own evidentiary limits, which is worth taking seriously when weighing the findings. This was a single course at a single institution, run once, with no pre-intervention baseline and no comparison group — so the study demonstrates that the framework gave students an authentic setting in which to apply and combine competencies, not that those competencies were newly created from zero. The rubric was never externally validated, the numeric scoring was applied only to a subset of teams, and the deep artifact analysis rests on one illustrative team rather than the full cohort. The authors also note that the study sits somewhere between a full multi-cycle design-based research program and a single implementation informed by DBR principles, since the instructional design itself has not yet been revised and re-run based on these results.

Even with those caveats, the framework’s transferability argument is compelling. Because the pedagogy centers on the engineering design process rather than any particular manufacturing application, the authors argue the same five-layer structure could be adapted to logistics, robotics, pharmaceutical production, automotive assembly, or energy systems. Future iterations may incorporate augmented and virtual reality, collaborative robots, digital thread architectures, and edge computing. For now, the study’s core contribution is a proof of concept: that undergraduate teams, given one semester, a real industrial problem, and professional-grade tools, can build a functioning multi-layer digital twin — and that the bottlenecks, blocking rates, and payback calculations they encounter along the way teach them more about modern manufacturing than any slide deck could.

Subject of Research: A design-based research framework integrating multi-layer digital twin development into undergraduate engineering education for Industry 4.0 competency development

Article Title: Integrating multi-layer digital twins into engineering education: a design-based research framework for industry 4.0 competency development

Article References: Bento, A. C., Molina-Leal, A., Silva, J. R., Padrón-Gómez, H. E., Tobías-Rodríguez, S. E., Marroquín-Treviño, E., & Grajeda-Sánchez, J. P. (2026). Integrating multi-layer digital twins into engineering education: a design-based research framework for industry 4.0 competency development. Frontiers in Education, 11, Article 1925782. https://doi.org/10.3389/feduc.2026.1925782

Image Credits: AI Generated

DOI: 10.3389/feduc.2026.1925782

Keywords: digital twin, engineering education, Industry 4.0, project-based learning, design-based research, discrete-event simulation, SCADA, artificial intelligence, manufacturing automation, Tecnomatix Plant Simulation, competency development, Tecnológico de Monterrey