Machine Learning-Based Predictive Modeling of Infrared Spectroscopic Data from Thermal Conversion of Athabasca...

Area of Science:

• Chemical Engineering
• Spectroscopy
• Machine Learning

Background:

• Traditional physical measurements for analyzing thermal cracking products are time-consuming.
• Fourier-transform infrared (FTIR) spectroscopy provides valuable chemical information but requires optimization for real-time analysis.
• Developing a soft-sensor can bridge the gap between slow physical methods and rapid online monitoring.

Purpose of the Study:

• To develop a reliable soft-sensor for predicting FTIR intensities of thermal cracking products from Athabasca bitumen.
• To enhance the predictive accuracy and efficiency of FTIR spectroscopy using various machine learning (ML) techniques.
• To reduce reliance on slow, traditional physical measurements in process monitoring.

Main Methods:

• Implementation of diverse ML models: Linear Regression (LinR), PLSR, SVR, k-NN, RF, and GBR.
• Training and testing models across four distinct scenarios with varying temperatures (25-420 °C) and reaction times.
• Utilizing Bayesian optimization for hyperparameter tuning to optimize model performance.

Main Results:

• Gradient Boosting Regression (GBR) consistently demonstrated the highest predictive accuracy (R²).
• GBR achieved 99.66% accuracy in Scenario 1 (all data) and robust performance across varied temperature conditions in Scenarios 2-4 (80.40-94% accuracy).
• Ensemble methods, especially GBR, outperformed individual models like RF, k-NN, LinR, and PLSR.

Conclusions:

• GBR models, optimized via Bayesian tuning, are highly effective for predicting FTIR intensities.
• The developed soft-sensor offers a reliable, time-saving alternative to traditional physical experimentation.
• Integrating ML with Bayesian optimization significantly advances online FTIR spectral prediction for chemical analysis.