|
|||||
|
|
Deep chemometrics using one-dimensional convolutional neural networks for predicting crude oil properties from FTIR spectral data |
|
| Authors: | Souvik Ta Shahla Alizadeh Lakshminarayanan Samavedham Ajay K Ray |
| Affiliation: | 1. Department of Chemical and Biochemical Engineering, Western University, London, Ontario, Canada
Contribution: Conceptualization, Data curation, Formal analysis, Methodology, Software, Validation, Visualization, Writing - original draft;2. Department of Chemical and Biochemical Engineering, Western University, London, Ontario, Canada Contribution: Conceptualization, Data curation, Formal analysis, ?Investigation, Methodology, Software, Validation, Visualization, Writing - original draft;3. Department of Chemical and Biochemical Engineering, Western University, London, Ontario, Canada Department of Chemical and Biomolecular Engineering, National University of Singapore, Singapore, Singapore Contribution: Conceptualization, Formal analysis, Funding acquisition, Methodology, Project administration, Resources, Supervision, Writing - review & editing;4. Department of Chemical and Biochemical Engineering, Western University, London, Ontario, Canada |
| Abstract: | The determination of physicochemical properties of crude oils is a very important and time-intensive process that needs elaborate laboratory procedures. Over the last few decades, several correlations have been developed to estimate these properties, but they have been very limited in their scope and range. In recent years, methods based on spectral data analysis have been shown to be very promising in characterizing petroleum crude. In this work, the physicochemical properties of crude oils using Fourier transform infrared (FTIR) spectrums are predicted. A total of 107 samples of FTIR spectral data consisting of 6840 wavenumbers is used. One dimensional convolutional neural networks (CNNs) were used employing FTIR spectral data as the one-dimensional input and Keras and TensorFlow were used for model building. The Root Mean Square Error decreased from 160 to around 60 for viscosity when compared to previous machine learning methods like partial least squares (PLS), principal component regression (PCR), and partial least squares regression with genetic algorithm (PLS-GA) on the same data. The important hyper-parameters of the CNN were optimized. In addition, a comparison of results obtained with different neural network architectures is presented. Some common preprocessing techniques were also tested on the spectral data to determine their impact on model performance. To increase interpretability, the intermediate neural network layers were analyzed to reveal what the convolutions represented, and sensitivity analysis was done to gather key insights about the wavenumbers that were the most important for prediction of the crude oil properties using the neural network. |
| Keywords: | deep learning in spectra hyperparameter search neural architecture search one dimensional convolutional neural networks property prediction using deep learning |