Deep Learning Algorithm for Identification of Combustion Relevant Species in FTIR Spectra
Conference Proceeding
Overview
abstract
Fourier Transform Infrared (FTIR) Spectroscopy is a widely used analytical technique for identifying the composition of materials by analyzing their spectra and is a diagnostic used in combustion and fire experiments as an accurate chemical detection method for speciation. These spectra provide crucial information about molecular structures by revealing the presence of specific functional groups within a molecule. Traditional analysis of FTIR spectra involves interpreting absorption peaks corresponding to various bond vibrations, which are characteristic of specific functional groups. This process requires expert knowledge and a meticulous examination of the spectra to accurately associate peaks with functional groups. However, such traditional methods are not only time-consuming but also prone to inaccuracies, particularly when dealing with complex molecules where overlapping peaks complicate peak identification. Recent advances in deep learning have shown promise in automating the interpretation of FTIR spectra. Several studies have demonstrated the potential of machine learning models to identify functional groups with high accuracy. However, these models have typically been limited in scope, focusing only on a narrow range of the most common functional groups found in organic compounds. This limitation restricts their broader application, especially in analyzing more diverse or less studied chemical structures. In this work, we introduce a new deep learning framework designed to achieve accurate identification of functional groups in a wide variety of compounds, including those with complex or rare substructures. Our approach employs a combination of an autoencoder for feature extraction, a feature selection process to isolate key spectral features, and an optimized neural network for precise classification. These components work together to create a system capable of processing diverse FTIR spectra. To train and validate our model, we compiled a large dataset of infrared spectra by web scraping publicly available databases, including the National Institute of Standards and Technology (NIST) and Spectral Database (SDBS) repositories. These datasets encompass a broad range of functional groups, offering a diverse foundation for model development. The training process involved over 50,000 spectra, allowing the model to learn subtle patterns and relationships within the data. A key feature of our trained model is its ability to identify spectral patterns that human chemists typically use to recognize functional groups. This capability allows the model to achieve high accuracy over a wide range of functional groups, making it valuable tool for quick identification in combustion analysis.
