AQCI Dataset and MATLAB Implementation: Monte Carlo Training Data, Mixture-of-Experts Model, and Surrogate Neural Network
Description
Dataset Description (for Mendeley Data) This repository contains the complete dataset and MATLAB implementation supporting the study: “AQCI: A Physics-Informed Mixture-of-Experts Framework for Continuous Indoor Air Quality Comfort Assessment and Pathway Attribution.” The archive enables full reproducibility of the Air Quality Comfort Index (AQCI) framework, including Monte Carlo training data generation, Mixture-of-Experts (MoE) computation, neural network surrogate training, dominance mapping, threshold derivation, and intervention classification. 1. Contents Overview The repository includes: (A) Monte Carlo Dataset Multidimensional sampled environmental input space: CO₂ concentration (ppm) PM₂.₅ concentration (µg/m³) TVOC concentration (µg/m³) Air velocity (m/s) Relative humidity (%) Air temperature (°C) Metabolic rate (met) Corresponding computed outputs: Pathway-specific irritation scores Gating weights Aggregated AQCI severity (continuous, normalized 0–1) (B) Mixture-of-Experts Implementation Softmax gating function Pathway-specific irritation functions: Ventilation-driven pathway Particulate-driven pathway Chemical-driven pathway Aggregation logic Dominance attribution calculation (C) Neural Network Surrogate Trained feedforward neural network model Training scripts Normalization procedures Validation outputs Final constrained AQCI prediction model (D) Threshold Classification Multi-level Otsu implementation Derived thresholds (t₁, t₂, t₃) Severity regime classification scripts (E) Visualization Scripts Dominance maps CO₂–PM₂.₅ design plane plots Gating topology surfaces Severity regime visualizations 2. Reproducibility All scripts are written in MATLAB (version XX or later). No proprietary toolboxes beyond standard Neural Network Toolbox are required. Running the main script file reproduces: Monte Carlo sampling MoE reference computation Surrogate training AQCI prediction Dominance maps Severity classification Random seeds are fixed to ensure deterministic replication of reported results. 3. Intended Use The dataset supports: Reproduction of results in the associated journal article Sensitivity testing under alternative environmental ranges Extension of AQCI to building-specific calibration Integration into building management system (BMS) workflows The repository is intended for academic and research use. 4. License This dataset is released under a CC BY 4.0 license.
Files
Institutions
- United Arab Emirates UniversityAbu Dhabi, Al Ain