The Japanese version of this article is available here: AIによるスペクトル解釈支援：フルスペクトル・データ駆動型アプローチによるIR分析の探求
Fundamentals of FTIR Analysis and Its Challenge
Fourier Transform Infrared Spectroscopy (FTIR) stands as a cornerstone analytical technique in modern chemistry and materials science, offering multiple significant advantages. As a powerful tool for Chemical Identification, it provides detailed insights into molecular structures through the analysis of vibrational and rotational states. Its non-destructive analysis capability ensures sample preservation, while its quantitative capability enables precise concentration measurements. The technique's Versatility allows for analysis across various sample types — solids, liquids, and gases — making it invaluable across multiple industries. Furthermore, its real-time monitoring capabilities and cost-effectiveness make it an essential tool in both research and industrial applications.
Fig.1 Key Advantages of Infrared (IR) Spectral Analysis
FTIR spectra comprise two primary regions, each offering distinct analytical value. The functional group region, spanning from 4000 to 1500 cm⁻¹, contains characteristic absorption bands that correspond to specific molecular moieties. This region typically features well-separated peaks that enable direct identification of functional groups such as hydroxyls, carbonyls, and amines. The fingerprint region, extending from 1500 to 400 cm⁻¹, complements the functional group region by providing a unique molecular signature. This region contains complex patterns of overlapping bands that reflect the entire molecular skeleton.
The complexity of FTIR interpretation extends beyond simple peak identification. Peak characteristics such as shape, width, and intensity all carry significant information. A broad peak might indicate different molecular environments compared to a sharp peak at the same position. Furthermore, molecular vibrations can couple and interact, leading to complex spectral patterns that defy straightforward interpretation.
Fig.2 Challenges in FTIR Spectral Interpretation of Functional Group and Fingerprint Regions
Machine Learning Revolution in IR Analysis using Full-Spectrum Data-Driven Approach
Modern machine learning for IR analysis using a full-spectrum data-driven approach has dramatically expanded analytical capabilities. Current systems can analyze molecular structures directly, without requiring extensive pre-processing, manual feature extraction, peak detection, and peak clustering processes. These approaches treat the spectrum as a continuous data pattern, considering all spectral features simultaneously.
The development of various neural network architectures has been crucial to this progress. These systems can now process complex molecular information and identify subtle patterns in spectral data that might be missed by traditional analysis methods. These patterns include variations in peak shapes, baseline changes, and complex interactions between different spectral regions.
Key applications of Full-Spectrum Data-Driven Approach for IR spectra include:
- Classification of Organic Compounds using SVM, Random Forests, and Neural Networks
- Quantitative Analysis of Mixtures through PLSR combined with machine learning methods
- Functional Group Identification using CNN models
- Structure Identification using Transformer model
The application of IR analysis using Full-Spectrum Data-Driven Approach
At MI-6, we explore an integrated approach that combines multiple specialized applications for spectral data analysis. The core methodology utilizes a transformer architecture, implemented for both encoder and decoder parts using the OpenNMT-py library.
Fig.3 Schematic diagram of machine learning model for chemical properties prediction from IR spectra
The study utilized a dataset of 317,292 spectra simulated from molecular dynamic simulations. The simulated molecules contained C, H, N, O, S, P, and halogens, with heavy atom counts ranging from 6 to 13. Each spectrum was recorded with 2 cm⁻¹ resolution spanning 400–3982 cm⁻¹. The dataset was divided into training (85%), testing (10%), and validation (5%) sets.
Case study 1: Functional Groups and Molecular Fragment Identification
Traditional IR analysis provides valuable bonding and functional group information, yet this interpretation traditionally requires deep expertise for accurate analysis. The machine learning-enhanced approach transforms this landscape significantly. In our case study, machine learning technology not only matches expert-level analysis but excels in identifying over 80 organic functional groups with high accuracy.
The outstanding capability of machine learning lies in its ability to discriminate between structurally similar compounds, maintaining exceptional classification performance even when structural differences are minimal. Beyond basic functional group analysis, machine learning technology offers unprecedented flexibility - it can be customized to detect and identify specific structural features of interest based on client needs.
This machine learning-driven approach delivers multiple benefits: it streamlines the analysis process, reduces time requirements, maintains high accuracy in IR spectral interpretation, and most importantly, bridges the expertise gap between specialists and non-specialists in IR analysis. This democratization of IR analysis capability ensures consistent, expert-level results regardless of the user's experience level.
Fig.4 Performance of functional group and molecular fragment identification using IR spectra and machine learning
Case study 2: Chemical Properties Prediction
While traditional IR spectral analysis primarily reveals bonding and functional group information, our integration of machine learning with IR spectroscopy has revolutionized chemical property prediction by offering several key advantages. First, it enables rapid and automated analysis that surpasses human capability in processing complex spectral patterns. Second, it uncovers hidden correlations between spectral features and chemical properties that would be impossible to detect through conventional analysis. Third, it provides quantitative predictions of multiple chemical properties simultaneously from a single IR spectrum - a task that would be extremely challenging through conventional human analysis.
Fig.5 Performance of chemical properties prediction using IR spectra and machine learning
Our case study demonstrates these advantages, with a machine learning model successfully predicting chemical properties from the RDKit Molecule Descriptor Module. These predictions include crucial molecular characteristics such as Molecular Refractivity (MR), molecular log partition coefficient (LogP), molecular weight (MolWt), topological polar surface area (TPSA), and VSA descriptors that capture surface area properties. This comprehensive prediction capability from a single IR spectrum showcases how machine learning transforms a basic analytical tool into a powerful platform for advanced chemical characterization.
Case study 3: Elemental Composition Prediction
Elemental Composition (CHON%) is crucial information in biomass and energy conversion fields, as these elements directly influence key properties such as heating value, combustion behavior, and environmental impact. Traditionally, determining CHON% requires destructive analytical techniques like CHN/O Elemental Analyzer, which consumes samples, demands significant time, and requires specialized equipment.
Fig.6 Performance of elemental composition prediction using IR spectra and machine learning
While IR spectroscopy typically doesn't directly provide elemental composition information, integrating IR spectral data with machine learning has opened new possibilities. This innovative approach enables the prediction of elemental compositions directly from IR spectra. The machine learning model demonstrates remarkable accuracy in predicting the content of C, H, N, and O - the most prevalent elements in organic molecules. The predictions show impressive precision with average errors of less than 1 atom for carbon, nitrogen, and oxygen (C, N, O) and less than 2 atoms for hydrogen (H) compared to conventional analysis methods. The high average R² value of 0.79 across all elements confirms the robust correlation between predicted and actual values.
This advancement transforms traditional IR analysis capabilities, offering a non-destructive, rapid, and accurate method for determining elemental composition, which is particularly valuable for biomass and energy conversion applications.
Summary
Advanced IR Spectral Analysis
The Full-Spectrum Data-Driven Approach revolutionizes traditional IR analysis by analyzing complete spectral patterns without peak detection or manual feature extraction processes. Our methods treat the spectrum as a continuous data pattern, considering all spectral features simultaneously. Through advanced model selection utilizing machine learning, we achieve comprehensive spectral analysis that leverages modern AI capabilities to uncover complex patterns in spectral data.
Value for R&D and Industry
Our solutions dramatically reduce analysis time and resource requirements while ensuring cost optimization through minimized sample preparation and handling. We maintain quality and reliability through consistent, reproducible results, making advanced analysis capabilities accessible to both specialists and non-specialists. This accelerates material development cycles and enables data-driven decision making for faster market entry.
References
- R. Almalih, "Introduction to Fourier Transform Infrared Spectroscopy (FTIR)", 2024.
- B. Stuart, "Infrared Spectroscopy", Fundamentals and Applications, 2005, DOI: 10.1002/0471238961.0914061810151405.a01.pub2.
- A. Kassem et al., "Applications of Fourier Transform-Infrared Spectroscopy in Microbial Cell Biology and Environmental Microbiology: Advances, Challenges, and Future Perspectives", Frontiers in Microbiology, vol. 14, 2023, DOI: 10.3389/fmicb.2023.1304081.
- A. Argyris, J.-J. Filippi, and D. Syvridis, "Support vector machine classification of volatile organic compounds based on narrow-band spectroscopic data", Journal of Chemometrics, vol. 29, 2015, DOI: 10.1002/cem.2660.
- T. Bikku, R. Fritz, Y. Colón, and F. Herrera, "machine learning identification of organic compounds using visible light", 2022, DOI: 10.48550/arXiv.2204.11832.
- L. H. Rieger, M. Wilson, T. Vegge, and E. Flores, "Understanding the patterns that neural networks learn from chemical spectra", Digital Discovery, vol. 2, no. 6, pp. 1957–1968, 2023, DOI: 10.1039/D3DD00203A.
- M. Madden and A. Ryder, "machine learning Methods for Quantitative Analysis of Raman Spectroscopy Data", Proceedings of SPIE - The International Society for Optical Engineering, vol. 4876, 2002, DOI: 10.1117/12.464039.
- G. Jung, S. G. Jung, and J. M. Cole, "Automatic materials characterization from infrared spectra using convolutional neural networks", Chemical Science, vol. 14, pp. 3600–3609, 2023, DOI: 10.1039/D2SC05892H.
- M. A. Z. Chowdhury and M. A. Oehlschlaeger, "Deep Learning for Gas Sensing via Infrared Spectroscopy", Sensors, vol. 24, no. 6, article 1873, 2024, DOI: 10.3390/s24061873.
- N. Saquer, R. Iqbal, J. D. Ellis and K. Yoshimatsu, “Infrared spectra prediction using attention-based graph neural networks”, Digital Discovery, 3, 602-609, 2024, https://doi.org/10.1039/D3DD00254C
- V. H. M. Doan, C. D. Ly, S. Mondal, T. T. Truong, T. D. Nguyen, J. Choi, B. Lee, and J. Oh, "Fcg-Former: Identification of Functional Groups in FTIR Spectra Using Enhanced Transformer-Based Model", Analytical Chemistry, 96 (30), 12358-12369, 2024, DOI: 10.1021/acs.analchem.4c01622.
- G. Klein et al., "OpenNMT: Open-Source Toolkit for Neural Machine Translation," in Proceedings of ACL 2017, System Demonstrations, Vancouver, Canada, 2017, pp. 67–72.
- X.F. Cadet, O. Lo-Thong, S. Bureau et al., “Use of Machine Learning and Infrared Spectra for Rheological Characterization and Application to the Apricot”, Sci Rep 9, 19197 ,2019, https://doi.org/10.1038/s41598-019-55543-7
