Predicting Indoor-Outdoor PM Differentials via XGBoost

This repository contains the data, source code, and analysis for a machine learning project focused on predicting Indoor Air Quality (IAQ) based on high-resolution outdoor environmental data. The study is conducted at the Richardson IQHQ in Richardson, Texas—a facility equipped with a specialized pathogen defense system and HEPA-filtered HVAC.

Project Overview

The objective is to predict the difference between outdoor and indoor Particulate Matter (PM) concentrations ($\Delta$ PM = PM_Outdoor - PM_Indoor). By leveraging high-fidelity outdoor sensor arrays and machine learning (XGBoost), we aim to understand how outdoor pollutants penetrate or are mitigated by advanced building filtration systems.

Data Source & Sensor Fusion

1. Outdoor Sensors (MINTS-AI Lab, UT Dallas)

• Pierra IPS7100: Measures PM density ($\mu$ g/m³) and particle count (counts/Litre) for 7 size bins: PM_0.1 , PM_0.3 , PM_0.3 , PM_1,0 , PM_2.5 , PM_5.0 , and PM_10.0 .
• BME280: Air temperature, pressure, humidity, and dewpoint.
• RG15: Rainfall intensity.
• AIRMAR: Ultrasonic wind speed, direction, and secondary meteorological data.
• AS7265x Multispectral Sensor: Measures solar irradiance across 18 wavelength bins (410nm to 940nm).
• Calculated Features: Solar Zenith Angle (derived from lat/long and local time).

2. Indoor Sensors

• Network: 20 indoor sensors distributed throughout the IQHQ building.
• Resolution: 10-minute intervals.
• Metrics: PM_1.0, PM_2.5, and PM_10.0 concentrations(($\mu$ g/m³))

Methodology

Data Processing Pipeline

1. Downsampling: Outdoor data (1-second resolution) was time-averaged to a 10-minute resolution to match the indoor sensor frequency.
2. Lag Analysis: To account for the time it takes for outdoor air to affect indoor environments, a sensor-by-sensor lag analysis was performed for each PM bin (PM_2.5, PM_2.5, PM_10.0).
3. Feature Engineering: Integration of spectral solar data, meteorological conditions, and calculated solar angles as predictors.

Machine Learning

• Model: XGBoost (Extreme Gradient Boosting) in Python.
• Target Variables:
  • $\Delta$ PM_1.0
  • $\Delta$ PM_2.5
  • $\Delta$ PM_10.0

Repository Structure

IndoorOutdoorStudy/
├── Code/
│   └── googleColab_GPUoptimized/
│       └── Indoor_Outdoor.ipynb          # Main GPU-optimized Google Colab notebook
├── Data/
│   └── MintsData/
│       └── Indoor/
│           ├── 001e064a1520/             # ⚠️ Outdoor sensor data (MINTS-AI Lab node)
│           │   └── combinedValo01Data.csv
│           ├── 70b3d540f40ce420/         # Indoor sensor 01
│           │   └── 70b3d540f40ce420_combined.csv
│           ├── 70b3d540f40ce421/         # Indoor sensor 02
│           ├── 70b3d540f40ce422/         # Indoor sensor 03
│           ├── 70b3d540f40ce423/         # Indoor sensor 04
│           ├── 70b3d540f40ce424/         # Indoor sensor 06
│           ├── 70b3d540f40ce425/         # Indoor sensor 07
│           ├── 70b3d540f40ce426/         # Indoor sensor 08
│           ├── 70b3d540f40ce427/         # Indoor sensor 09
│           ├── 70b3d540f40ce429/         # Indoor sensor 10
│           ├── 70b3d540f40ce42d/         # Indoor sensor 11
│           ├── 70b3d540f40ce42f/         # Indoor sensor 12
│           ├── 70b3d540f40ce430/         # Indoor sensor 13
│           ├── 70b3d540f40ce433/         # Indoor sensor 14
│           ├── 70b3d540f40ce434/         # Indoor sensor 15
│           ├── 70b3d540f40ce435/         # Indoor sensor 16
│           ├── 70b3d540f40ce436/         # Indoor sensor 17
│           ├── 70b3d540f40ce438/         # Indoor sensor 18
│           ├── 70b3d540f40ce43a/         # Indoor sensor 19
│           ├── 70b3d540f40ce43b/         # Indoor sensor 20
│           ├── 70b3d540f40ce43c/         # Indoor sensor 21
│           └── valo_cross_correlation_results.csv
└── Results/
    ├── PM1.0/
    │   ├── allParameters_PM1/            # Heatmaps & CSVs using all outdoor features
    │   └── top10_PM1/                    # Per-sensor plots using top-10 features
    ├── PM2.5/
    │   ├── allParameters_PM2.5/
    │   └── top10_PM2.5/
    └── PM10.0/
        ├── allParameters_PM10.0/
        └── top10_PM10.0/

Note on data placement: The folder 001e064a1520 resides inside Data/MintsData/Indoor/ for organizational convenience, but it contains outdoor sensor data collected by the MINTS-AI Lab node. All 70b3d540f40ce4xx folders contain indoor sensor data.

Requirements

• Python 3.x (Google Colab environment recommended)
• xgboost
• pandas
• numpy
• matplotlib / seaborn
• scipy (for lag/cross-correlation analysis)

Usage

1. Open the notebook: Load Code/googleColab_GPUoptimized/Indoor_Outdoor.ipynb in Google Colab (GPU runtime recommended).
2. Point to data: Ensure Data/MintsData/Indoor/ is accessible (e.g., mount Google Drive or upload the folder).
3. Run all cells: The notebook handles data loading, downsampling, lag analysis, XGBoost training, and result export in sequence.
4. Outputs: Model metrics, feature-importance CSVs, and visualization plots are saved to the corresponding Results/PM*/ subfolders.

Acknowledgments

Special thanks to the MINTS-AI Lab at UT Dallas for providing the outdoor sensor infrastructure and the management at Richardson IQHQ for facility access.