Source-linked AI summary
From Sensor Data to Classroom Inquiry: GenAI-Supported Exploration of School Digital Twin Data
Themistoklis Sarantakos, Dimitrios Amaxilatis, Michail Giannakos, Georgios Mylonas
TL;DR
Digital Twin use in schools remains limited, while GenAI integration raises questions about reliability, privacy, cost, and pedagogical appropriateness. This paper develops and evaluates a chatbot over a school-building Digital Twin using real IoT data, finding perceived advantages for natural-language exploration and sustainability-oriented educational planning. The study also identifies verification, trust, response-speed, dashboard, and long-term-use constraints.
Problem
Digital Twin implementations for educational buildings remain limited, and GenAI integration in schools raises unresolved concerns about cost, energy use, hallucination, privacy, and educational use.
Method
The study builds an LLM-based conversational interface over a Digital Twin for two school buildings, using real environmental and energy IoT data, and evaluates it with 17 educators against a conventional web interface.
Results
Participants generally perceived the chatbot as easier and more intuitive, especially for natural-language questions, measurement comparisons, correlations, inquiry, hypothesis formation, and lesson-plan design.
Takeaways & Limitations
GenAI conversational interfaces may make Digital Twin data more accessible for sustainability-oriented educational activities when interaction is supported by tool-mediated retrieval and pedagogical grounding.
Takeaways & Limitations
The workshop did not assess long-term chatbot use, the educator sample was gender-imbalanced, and the system’s operational energy resources were not considered or communicated.
Abstract
from arXiv · showhide
Digital Twins for educational buildings can support sustainability-oriented learning, but their use in schools remains limited. This paper presents a GenAI-based chatbot built on top of an existing Digital Twin for two school buildings in Greece, using real IoT data from environmental sensors and energy meters. The chatbot enables educators to query live and historical building data, compare spaces, and generate ideas for classroom activities through natural language. The system was evaluated in an 80-minute workshop with 17 secondary-school educators, who compared it with an existing web-based dashboard. Results show strong perceived usability and pedagogical value, particularly for inquiry-based learning, hypothesis formation, and interdisciplinary lesson planning. Participants also highlighted limitations related to response speed, data verification, trust, and the continued value of visual dashboards. Overall, the findings suggest that GenAI interfaces can make Digital Twin data more accessible for educational use, provided they are designed with transparency, verification, and pedagogical grounding.
I. INTRODUCTION
The paper addresses the limited use of Digital Twins in educational buildings and explores whether an LLM-based chatbot can connect real school-building IoT data with sustainability-oriented education. It frames the study around technical feasibility, interface value, and educators’ pedagogical use of the system.
- Motivation: Educational-building Digital Twin implementations remain relatively uncommon despite their relevance within the building stock.Educational buildings average 17% of Europe’s non-residential building stock by area, although country-level figures vary.
- Motivation: GenAI adoption in education remains unsettled because costs, energy use, hallucination, privacy, efficacy, and ethics complicate school deployment.The paper specifically identifies unresolved questions about integrating GenAI into education and highlights financial, energy, reliability, and privacy concerns.
- Study focus: The study combines real environmental and energy IoT data from two school buildings with an LLM-based approach for sustainability-oriented educational activities.The underlying Digital Twin includes sensing infrastructure, 3D building digitization, and web-based interfaces.
- Study focus: An LLM-based chatbot was evaluated with 17 lower-secondary educators to examine access to Digital Twin data and support for educational-activity design.The evaluation used a structured workshop and interviews conducted at the school.
- Research questions: The research questions cover conversational access to school IoT data, chatbot advantages and limitations versus web interfaces, and educators’ lesson-planning perceptions.These correspond to the study’s technical, interface, and pedagogical aims.
II. PREVIOUS RELATED WORK
Prior work has examined Digital Twins, GenAI, IoT-enhanced learning, and multimodal educational sensing, but school-building Digital Twin use for sustainability-oriented classroom activities remains underexplored.
- Digital Twins in education: Educational Digital Twins integrate sensors, BIMs, building-management systems, and space-management systems to support building performance and resource decisions.Existing examples largely emphasize real-time monitoring, visualization, operational decision-making, energy efficiency, and sustainability assessment.
- GenAI in education: GenAI in education may support personalization, creativity, and new interactions, while raising concerns about efficacy, ethics, privacy, hallucination, overreliance, and pedagogical soundness.These concerns are especially relevant in school settings.
- Emerging intersections: Emerging studies connect GenAI, IoT, and data-rich learning environments to interactive learning and interpretation of complex, domain-specific datasets.Related work also examines IoT-based multimodal data collection for learning analytics.
- Research gap: This study combines real-world IoT data, a school-building Digital Twin, and an LLM conversational interface for secondary-school sustainability activities.The authors identify systematic exploration of this combination as underexplored relative to prior work focused on universities, general IoT learning, or urban environmental learning.
III. ARCHITECTURE & IMPLEMENTATION
The system provides a conversational natural-language interface to an IoT deployment through a serverless AWS architecture layered over a Digital Twin for two school buildings.
- System architecture: The system uses three logical tiers: a static web front-end, a serverless back-end with an agentic LLM pipeline, and an IoT platform REST API.The architecture is hosted entirely on AWS.
- System architecture: The underlying Digital Twin supplies data from two school buildings at the University of Patras in Greece.The buildings contain a lower and an upper secondary school, with 46 educators and nearly 390 students.
A. IoT infrastructure and Digital Twin
The Digital Twin combines environmental sensing, power monitoring, building digitization, and controlled conversational access to support educational and research activities.
- IoT infrastructure: The infrastructure includes 12 environmental nodes across 12 classrooms and four power meters across two school buildings.The environmental nodes measure temperature, humidity, noise, motion, and luminosity; the meters monitor overall power consumption.
- Digital Twin: The two school buildings have been digitized and are represented through a detailed 3D model.This digitization forms part of the Digital Twin infrastructure.
- Digital Twin: The implemented infrastructure is intended to support both educational and research activities and is described as the first school-building Digital Twin implementation of its kind.The first-of-its-kind characterization is explicitly qualified by “to the best of our knowledge.”
- Conversational access: The chatbot’s back end authenticates users and uses an agentic LLM pipeline to invoke tools that retrieve IoT information before generating responses.Tool use supports multi-step queries involving room identification, sensor retrieval, and current or historical readings.
D. Agentic Tool Set
The agentic tool set constrains chatbot access to Digital Twin data through controlled IoT-platform tools, supporting conversational exploration of live, historical, and cross-space measurements.
- Agentic data access: The tool set mediates access to the IoT platform, enabling navigation across schools, rooms, and sensors while retrieving live or historical measurements.Constrained access supports grounded responses while hiding internal identifiers and API details from educators.
- Infrastructure: The backend uses cached structural metadata and infrastructure-as-code to support deployment and reduce redundant API calls.Structural school, room, and sensor data are cached in Amazon S3, while AWS resources are provisioned through an AWS SAM CloudFormation template.
- Educational use: Conversational queries make technical building data more accessible to the wider school community.Users can request current readings, sensor overviews, building orientation, and historical summaries in plain language.
- Agentic data access: The system supports complex queries that identify patterns and correlations across building data.Examples include comparing temperatures in similar rooms and investigating changes over time.
V. EVALUATION - RESULTS
The evaluation of model options found that reliable multi-step tool orchestration required a more capable LLM, leading the system to adopt Claude for chained queries over Digital Twin data.
- Model evaluation: Smaller tested models consistently failed on multi-step API interactions by losing context and producing incomplete replies.The use case required sequential calls and reasoning across multiple steps.
- Model evaluation: LangChain tool integration reduced hallucination risk by allowing models to request additional information before forming grounded responses.This orchestration was especially important when queries required multiple API calls.
- Model evaluation: Claude was selected because it handled chained tool use reliably and produced accurate, coherent responses across multi-step queries.The selection reflected the balance between reasoning capability and response requirements.
B. User Evaluation Setup
An 80-minute workshop with 17 lower-secondary educators compared the chatbot with a legacy web dashboard through demonstrations, hands-on tasks, reflection, and questionnaire-based data collection.
- Participants and procedure: 17 educators from various disciplines participated in the evaluation workshop.The workshop lasted 80 minutes and directly compared the chatbot with the conventional web-based dashboard.
- Participants and procedure: Participants provided informed consent under GDPR-compliant procedures.
- Participants and procedure: The session comprised a 20-minute interface walkthrough, 40 minutes of comparative hands-on tasks, and 20 minutes of synthesis and data collection.The tasks covered data retrieval and correlation with the web interface, followed by insight and correlation using the chatbot.
- Measures: The questionnaire assessed usability, effectiveness, critical thinking, and comparative value.These dimensions captured interface functionality, problem-solving support, reflection on the environment, and comparison with other methods.
C. Results
Questionnaire results showed broadly positive evaluations, with usability and pedagogical-value items rated more favorably than limitation-related items.
- Overall ratings: 14 of 18 questionnaire items received mean scores above 4, indicating strong agreement across most participants.Responses were collected from 17 participants using five-point Likert scales.
- Overall ratings: Q2 ease of use received the highest score (M = 4.69, SD = 0.79).
- Overall ratings: Q9 trust in chatbot answers (M = 3.76) and Q7 pattern detection (M = 3.82) were the most modestly positive items, remaining above neutrality.
- Overall ratings: Limitation and readiness items scored approximately 3.5–3.8, reflecting uncertainty rather than outright resistance.These items concerned new challenges, preference for the web interface, and security concerns.
- Overall ratings: No questionnaire item fell below the neutral midpoint, while usability and pedagogical-value items were rated more positively than limitation-related items.
2) Section-Level Composite Scores:
Composite scores were positive for pedagogical value, usability, and trust, while limitations and readiness remained closer to the neutral midpoint. ICT familiarity was associated with significant group differences in usability and pedagogical value, but no item-level familiarity correlation reached significance.
- Pedagogical Value achieved the highest composite score (M = 4.33, p = 0.001 vs. neutral), followed by Usability (M = 4.32, p < 0.001) and Trust & Transparency (M = 4.01, p = 0.001).
- Limitations & Readiness scored near the neutral midpoint (M = 3.63, p = 0.008), indicating mild rather than strong concerns.
- High-familiarity educators rated Usability higher than low-familiarity educators (M = 4.43 vs. 3.83), whereas low-familiarity educators rated Pedagogical Value higher (M = 4.87 vs. 4.21).The groups were high: Q1 ≥4, n = 14, and low: Q1 ≤3, n = 3.
- The familiarity-group differences were statistically significant only for Usability (p = 0.026) and Pedagogical Value (p = 0.032).
- No item-level correlation between ICT familiarity and responses reached significance at p < 0.05.For n = 17, the significance threshold was |r| = 0.48.
4) Inter-Item Correlations:
Inter-item correlations linked pedagogical inquiry, cognitive depth, accessibility, and usability, while chatbot usability was associated with less preference for the legacy web interface. The broader discussion positions these patterns alongside tool orchestration and complementary dashboard use.
- The strongest association was between student research and hypothesis formation (r = 0.825), suggesting a unified pedagogical affordance.
- Deeper thinking and accessibility co-varied strongly (r = 0.776), linking cognitive depth with ease of access to information.
- Ease of use, fewer interaction steps, and pattern detection formed a tight usability cluster with correlations of r = 0.62–0.69.
- Reliable multi-step Digital Twin queries required tool-mediated LLM orchestration, while the web dashboard remained useful for rapid visual inspection, charting, and verification.The chatbot was perceived as easier for natural-language exploration, comparisons, and correlations, but not as a replacement for visual tools.
A. Limitations
The study’s conclusions are bounded by a short workshop evaluation, an unbalanced educator sample, and unmeasured energy resources for chatbot operation. Future work extends evaluation across schools, educational levels, and students while exploring broader Digital Twin integration.
- The workshop did not assess chatbot use beyond the evaluation session, limiting generalizability regarding user retention and long-term effects.
- The educator sample was not balanced by gender, and the energy resources associated with chatbot operation were not considered or communicated.
- The study frames the interface as supporting access to sensing data and sustainability-oriented educational activities, while GenAI raises questions about reliability, transparency, and appropriate educational use.
- Future work will evaluate educators from additional schools and educational levels, include student groups, and investigate tighter integration with 3D building models.