Early Approaches in Vibrational Spectroscopy

Dubois and colleagues studied the problem of overlapping fragments. They developed the program DARC-EPIOS, which can retrieve structural formulas from overlapping C-NMR data [29]. The COMBINE software uses similar techniques. 

The problem of the interpretation of vibrational spectra is to calculate all possible combinations of substructures that may be present in a molecule consistent with the characteristic frequencies of a given infrared spectrum. Elyashberg and his team showed with an example that the infrared spectrum-structure correlation, as simply expressed by the characteristic frequency approach, does not allow one to establish the structure unambiguously due to a lack of information in characteristic frequencies [32]. They pointed out that the use of ANNs appears to be particularly promising. 

Artificial neural networks do not require any information about the relationship between spectral features and corresponding substructures in advance. The lack of information about complex effects in a vibrational spectrum (e.g., skeletal and harmonic vibrations, combination bands) does not affect the quality of a prediction or simulation performed by a neural network. 

Great attention has been paid in recent decades to the application of ANNs in vibrational spectroscopy [33,34]. The ANN approach applied to vibrational spectra allows the determination of adequate functional groups that can exist in the sample as well as the complete interpretation of spectra. Elyashberg reported an overall prediction accuracy using ANNs of about 80% that was achieved for general-purpose approaches [35]. Klawun and Wilkins managed to increase this value to about 95% [36]. 

Neural networks have been applied to infrared spectrum interpreting systems in many variations and applications. Anand introduced a neural network approach to analyze the presence of amino acids in protein molecules with a reliability of nearly 90% [37]. Robb used a linear neural network model for interpreting infrared spectra in routine analysis purposes with a similar performance [38]. Ehrentreich et al. used a counterpropagation (CPG) network based on a strategy of Novic and Zupan to model the correlation of structures and infrared spectra [39]. Penchev and colleagues compared three types of spectral features derived from infrared peak tables for their ability to be used in automatic classification of infrared spectra [40]. 

---