Robust and predictive quantitative analysis for mixtures of four dyes through smartphone spectrometer and neural networks

To date, optical spectroscopy has become the most promising, essential, and reliable characterization approach in almost every field of analytical, medical, and food science. However, there is a key challenge to transform current optical spectrometers to miniature devices which also be low-cost and easy-to-use [1,2]. Smart mobile phones with the digital camera associated as an attractive consumer electronic product are capable of being used as analytic optical devices. The analytic method could be colorimetric or spectroscopic, which in second one there is a requirement for proper monochromator. In such way, we considered to use digital versatile disks (DVDs) as monochromator. Spacing between the recording tracks in DVDs are 0.74 µm, making a grating of about 1350 lines/mm, which it is comparable to the gratings in spectrophotometers routinely used in laboratories [3]. A common desk lamp with incandescent light bulb and glass test tube was considered as light source and sample container (cuvette), respectively. Finally, the images captured by smartphone rear camera were analyzed by a home-built MATLAB program. We present the capability of detecting four dyes (i.e. Amaranth, Indigo Carmine, Tartrazine and Violet) in the relevant concentration range. Twenty eight different mixtures of these 4 dyes were prepared and their images and transmittance spectrums captured. The ease and simplicity of colorimetric analysis make operable through non-chemistry personnel. But in this case the results show that estimation of concentration is accurate only in simple mixtures (one or two dyes) and highly controlled optical and lightening situation for imaging, which it caused to being useless in outdoor applications. In the case of transmittance spectrums the common linear methods such as multiple linear regression (MLR) and partial least squares (PLS) were unable to build an accurate and robust model, due to the inherent non-linear relation of transmittance spectrums and concentrations. Therefore, radial basis function (RBF) as a simple type of non-linear modelling of neural network was used [4]. The model was constructed by splitting data to calibration and test sets through random selection of twenty samples as calibration and eight samples as test set. By choosing proper number of centers and sigma value for RBF, prediction performance of model was obtained up to 0.9 in Q2 value for all four dyes and in all concentration range. We demonstrated the comparative accuracy and high sensitivity of an affordable miniature smartphone spectrometer which could be used for education, research and testing portable device.