FACTS: Adaptive Climate Resilience Training

• FACTS is a dynamic training system that combines IoT atmospheric sensing with AI-driven retrieval and generation to create climate resilience learning challenges.
• The platform integrates edge-deployed sensors, a client application, and a RAG pipeline using models like ChatGPT-4o and Mistral-7B to ensure high retrieval accuracy (Recall@k = 0.9) and evaluation fidelity (BERTScore = 0.89).
• User studies indicate high usability and acceptance, with median scores of 4.5/5 for usefulness and 4.75/5 for ease of use, confirming its effectiveness in adaptive climate education.

The Future Atmospheric Conditions Training System (FACTS) is a platform developed for climate resilience education that leverages real-time atmospheric data collection, retrieval-augmented generation (RAG) architectures, and generative AI to create adaptive, location-aware learning experiences. FACTS integrates Internet-of-Things (IoT) sensor networks, a knowledge base of factual environmental and domain-specific information, embedding-based vector retrieval, and small LLMs (SLMs) to dynamically generate, evaluate, and personalize climate-relevant learning challenges for users based on their current environmental context (Khan et al., 29 Aug 2025).

1. System Architecture and Technical Workflow

FACTS consists of three core architectural modules: (1) a client-facing application, (2) edge-deployed IoT “Atmostate” stations for environmental sensing, and (3) a remote server implementing a RAG-based generative AI backend.

• Atmostate Stations: Each station comprises an ESP32 microcontroller, a DHT20 sensor (for air temperature and humidity), and a BH1750 light sensor, powered by a 5V battery. Atmospheric state vectors are captured at 5-second intervals and transmitted over WiFi via MQTT in JSON-encoded payloads (e.g., { "Temperature": "35℃", "Humidity": "75 %", "Light" : "3000. Lx" }).
• Client Application: The user interface, typically a mobile device, displays a list of proximate Atmostate stations, retrieves the current environmental snapshot, and presents a context-specific learning challenge—consisting of a short, atmospheric-conditional scenario and multiple candidate activities. The user selects one or more actions, after which system-generated feedback is presented.
• Server-Side RAG Pipeline:
  • Knowledge Base (KB): Domain literature (e.g., twenty peer-reviewed viticulture papers in the prototype) is chunked and embedded (“mxbai-embed-large” embedding model).
  • Vector Database: ChromaDB stores document vectors for efficient semantic retrieval (achieving Recall@k = 0.9 for k = 3).
  • Generative Resource Creation: An LLM (ChatGPT-4o) parses and synthesizes possible activities from KB content; DALL·E generates accompanying illustrative images.
  • Personalized Evaluation: Upon user submission, domain-relevant KB chunks are retrieved, and an SLM (Mistral-7B, BERTScore = 0.89 on outputs) evaluates the appropriateness of user choices given local atmospheric inputs, returning personalized justifications and a contextual debriefing.

The pipeline can be summarized as: environmental context sensing → semantic retrieval of knowledge → generative construction of activity options → user interaction/response → retrieval of context-specific explanatory evidence → SLM-based evaluation → adaptive feedback.

2. Adaptation Mechanisms and Personalization Logic

FACTS adaptation operates in two principal stages:

1. Initialization/Resource Generation: Upon KB update, factual resources (activity statements and images) are synthesized from newly embedded document chunks, populating a re-usable activities database.
2. General Per-Task Adaptation: Each learning challenge is dynamically constructed by combining the latest Atmostate vector with a sample of candidate activities. Semantic retrieval tailors the challenge prompt and corresponding reference materials to the current sensed meteorological state and activity set. Evaluation prompts are constructed such that the SLM must, given the union of retrieved KB context and context vector, label each candidate activity as “Good” or “Bad,” accompanied by justifications and an overall assessment.

Personalization occurs both through localization (actual environmental sensing at the user’s site) and action-specific feedback (user selection-dependent evaluation conditioned on authentic, locally measured atmospheric variables).

3. Knowledge Representation, Retrieval, and Verification

The domain knowledge base is segmented into semantically coherent text blocks during system initialization. Each chunk is embedded in high-dimensional vector space for context-sensitive retrieval. The embedding model and vector database are selected to maximize live retrieval accuracy (Recall@k), with empirical validation ensuring robust factual grounding for generative and evaluation tasks.

• Prompt Engineering: Prompts for both activity generation and response evaluation systematically inject environmental variables and retrieved factual evidence. For example, an evaluation prompt integrates: atmospheric vector (temperature, humidity, light, etc.), candidate activities, and top-k retrieved KB chunks.
• Validation: Generative outputs are tested for factual consistency using BERTScore; in the prototype, SLM-generated feedback achieved BERTScore = 0.89.

4. User Study and Effectiveness

FACTS was evaluated using a Technology Acceptance Model-based user study (n=20, ages 21–32, median perceived usefulness = 4.5/5, median ease of use = 4.75/5). All participants described the system as easy to learn and interact with; contextual adaptation and factual feedback were cited as tangible advantages. No formal learning outcomes were assessed, but the data support system usability and user acceptance (Khan et al., 29 Aug 2025).

The evaluation suggests that atmospherically adaptive, generative learning systems can deliver engaging, localized, and actionable climate resilience education with minimal training or domain expertise required from users.

5. Limitations and Areas for Future Research

Several constraints of FACTS are explicitly acknowledged:

• Hallucination Risk: RAG mitigates but does not eliminate generative hallucination; reliance on accurate retrieval and high-quality KBs is essential.
• Domain-Specific KB Construction: System performance is highly contingent on the existence of well-structured, domain-specific corpora that explicitly link activities to environmental contexts.
• Expert Validation and Coverage: Factual and evaluative accuracy ultimately depend on expert-validated datasets, especially in under-resourced domains or languages.
• Linguistic Coverage: Embedding and LLM limitations for non-dominant languages restrict the inclusivity and adaptability of the platform.
• Sample Size: The preliminary user study size does not support conclusions about long-term learning efficacy.

Future work is directed toward large-scale field deployment, KB expansion methodologies (including synthetic or expert-interview-based corpus generation), algorithmic diversification (more robust RAG and SLM pipelines), and application extension to other outdoor, weather-sensitive domains (e.g., forestry, marine biology, agriculture).

6. Broader Context and Comparison with Functional Analogues

While FACTS is an original system for climate resilience education, its technical principles share elements with operational atmospheric forecasting, monitoring, and training platforms in other scientific contexts. For example, mesoscale numerical weather prediction (e.g., Meso-Nh, GDAS) coupled with site-specific validation have been used for atmospheric support in astronomy (Turchi et al., 2017, Turchi et al., 2024) and astroparticle physics (Collaboration et al., 2012, Keilhauer et al., 2012). Unlike these automated forecast or reconstruction frameworks—where the goal is precise environmental state estimation to support observations or simulations—FACTS prioritizes the didactic integration of real-time atmospheric context into learning workflows, mediated by RAG and generative AI.

7. Technical Summary Table

Module	Description	Tools/Models
Sensing (Atmostate)	IoT weather sensors, 5s interval, MQTT over WiFi, ESP32, DHT20, BH1750	Hardware, MQTT
Knowledge Base	Domain-specific document chunks, embedding via mxbai-embed-large	ChromaDB
Activity Generation	Generative LLMs, image synthesis	ChatGPT-4o, DALL·E
Challenge Presentation	Mobile client, activity selection, image display	Android/iOS app
Personalized Evaluation	KB retrieval, SLM-based (Mistral-7B), prompt-based evaluation	SLM, BERTScore=0.89
Feedback Delivery	Justification, Good/Bad labelling, overall assessment	Client-Server

This system architecture enables FACTS to support situationally aware knowledge transfer, where both environmental and semantic context shape adaptive educational challenges in climate resilience domains (Khan et al., 29 Aug 2025).