Deep Learning-Based Gas Identification Using IR Photodetectors
An IISc research project developing a low-cost, intelligent gas sensing system by combining custom IR photodetector arrays with advanced machine learning.
1. The Challenge & Vision
Reinventing Gas Detection
The Problem
Traditional gas sensors are often bulky, expensive, and limited to specific environments. This creates a research gap for a portable, cost-effective, and intelligent solution for real-world applications like environmental monitoring and industrial safety.
Our Vision
To build a smart, low-cost sensing platform using custom-fabricated Quantum Dot IR photodetectors and an AI backend to accurately identify 15 different gases and their concentrations in real-time.
2. System Architecture
From Photons to Predictions
End-to-End Project Pipeline
Gas Exposure
In a controlled chamber
IR Detector Array
Multi-stack QD sensors
Data Acquisition
Log sensor & environmental data
AI/ML Backend
Classification & Regression
Multi-Stack IR Photodetector
Our system uses a custom array of 4-6 photodetectors. Each detector is sensitive to a different IR wavelength, capturing a unique "spectral fingerprint" for each gas. This design is built using cost-effective Quantum Dot (PbS, HgTe) materials.
3. Performance & Results
A Data-Driven Deep Dive
Gas Classification Performance
By evolving from a baseline Random Forest model to an optimized Voting Classifier (RF + Gradient Boosting), we increased gas identification accuracy dramatically across 15 gases.
Concentration Prediction (Regression)
The regression model accurately predicts gas concentration (in ppm). The final model shows strong generalization on a complex dataset of ~1 million samples.
Visualizing Data Separability with t-SNE
This t-SNE plot visualizes our high-dimensional sensor data in 2D. It shows how distinct gas classes form clusters. Overlapping areas (e.g., CH₄ & H₂S) highlight why certain gases are harder for the AI to distinguish, confirming the challenges of overlapping IR signatures.
4. Conclusion & Future Work
The Path Forward
This project successfully demonstrates a scalable framework for intelligent gas sensing. We've built an end-to-end system from hardware design to a robust AI backend, achieving ~90% classification accuracy.
Future Enhancement Roadmap
Full Device Fabrication
Complete the hardware prototype with a sealed chamber and integrated electronics.
Deep Learning Models
Implement CNNs and MLPs to better distinguish gases with overlapping signals.
Real-Time Integration
Use Arduino/Raspberry Pi for real-time data logging and on-device model inference.
Portable Device
Develop a field-ready sensor for industrial, agricultural, and environmental applications.