Project duration 3 years, starting May 2021.
Project funding 850,000 TL
Open project personnel positions These are three year positions starting in the first half of 2021. All positions are for scientists with backgrounds in the analytical sciences and/or programming and machine learning. FTIR spectroscopy and TensorFlow/Keras experience would be great but not critical. Doctoral and masters students are expected to complete a thesis based or related to the project based work. Students previously enrolled to a suitable MSc/PhD programs are also welcome to join the team with their supervisors approval.
And also researchers are supported from the following programs.
Undergraduate research assistant (TÜBİTAK STAR Stajyer Araştırmacı programı)
Postdoctoral, Graduate and Undergraduate researchers (TÜBİTAK BİÇABA BİDEB programı)
Contact me via email to organize an interview.
Artificial olfaction methods to determine complex real world mixtures
This project is about the analysis of complex and variable real world mixtures using biomimicry. The mixtures of interest are complex and variable; that is, the sheer number of constituent components and the intricate spectral fingerprint of each component render mixtures’ spectral analysis extremely difficult with classical chemometric methods. And also, these mixtures when found in nature and in industrial applications, manifest considerable molecular, spectral and constitutional differences, while at the same time show common invariant, archetypal characteristics that allow their perceptual/holistic classification. For example, perfumes consisting of many components, petroleum, diesel and jet fuels, coal variants, air samples collected for the determination of air pollution, rock pieces encountered in geology and astro-geology are both complex and variable. All kinds of plant and animal odours, fresh/rotten food odours, biological fluids and gases coming out of living things, human breath, sweat/ skin odour, urine and urine top layer gases are complex and variable in terms of their components and concentrations. The difficulty in the analysis of these samples stem from the competing requirements of high sensitivity and selectivity.
In turn, headspace of such mixtures is usually determined using gas chromatography-mass spectrometry (GC-MS), vibrational and laser-based spectrometers and using various gas sensors. There are two exclusive approaches: In the first spectroscopic method, specific spectral regions of mixtures are sought after and used as markers. However, in practice, sample complexity and variability hinder this approach’s feasibility. The second technique, mimicking biological olfaction, employs a series of redundant cross-responsive sensors that are used to create a fingerprint of the mixture. Rather than detecting individual molecules, a gestalt view is produced and analyzed through advanced statistical and computational methods in order to match the fingerprint patterns to the samples. The success of this approach depends on the pattern recognition back-end algorithms as much as the sensor array itself. The promise of the latter technique, known as the electronic nose research, is that sophisticated chemometrics can be used to identify complex mixtures provided that adequate multidimensional information is acquired through the less than perfect sensors.
This project takes a third way mixing the above two. We plan to use the FTIR spectrometer itself as an electronic nose, by feeding the FTIR spectra into a neural computation grid directly. A typical spectra contains several thousands of absorption intensities each of which is a convolution from many molecular species, thus it is a multidimensional, cross-responsive and redundant fingerprint signal from the mixture. We will start by designing the appropriate deep neural networks to analyze spectra of virtual simulated mixtures. Neural networks design will go beyond standard deep networks and include recurrent and spiking networks to facilitate the use of temporal coding seen in the biological olfaction. In addition, the architecture of the networks will be informed by the design principles observed in the olfactory sensing pathways. Simulations will be validated with FTIR measurements taken from actual complex mixtures prepared in the laboratory. Finally, we will test the ultimate performance of deep neural networks by testing their accuracy with sets of real world natural, industrial and bio-samples. This project will enable scaling up of the proof-of-concept computational studies of FTIR spectra/deep learning, the PI has undertaken in the past few years.
In the literature no comprehensive study has been found in which the potential of the FTIR method is evaluated as an electronic nose. The dazzling progress of deep learning methods in recent years suggest that the technology is now available to train deep neural networks with very large input vectors and multiple hidden layers. With the development of what we call deep spectroscopy, we aim to make a conceptual and technological contribution to complex mixture analytics. This method is expected to become an invaluable asset in diverse areas such as environmental research, biochemistry, biomedicine, industrial process analytics and forensics where complex molecular mixtures are routinely encountered.
General references about spectroscopy, olfaction and neuromorphic computing.
General chemistry Principles of Modern Chemistry, 7th edition, Oxtoby, Gillis Campion, Cengage Learning, 2011.
Instrumental Analysis Principles of Instrumental Analysis, 6th Edition, Skoog, Holler, Crouch, Cengage Learning, 2006.
Neuroscience Neuroscience, Purves et al (editors), 5th edition, Sinauer Associates Inc.,U.S.
Machine olfaction Handbook of Machine Olfaction: Electronic Nose Technology, Editors Pearce, Schiffman, Nagle, Gardner, Wiley, 2002.
Neuromorphic Olfaction, Persaud, Marco, Gutierrez-Galvec, CRC Press, 2013.
Artificial Olfaction in the 21st Century, Covington et al, 2021.
High Selectivity Boolean Olfaction Using Bragg Fibers, Yaman et al, Analytical Chemistry, 2012.
Bioinspired Optoelectronic Nose with Nanostructured Fibers, Yildirim et al, Advanced Materials, 2011.
Koku bilimine doğru elektronik ve fotonik burunlar, Bayındır, Yaman, Yıldırım - Bilim ve Teknik, 2011
Machine learning Deep learning, LeCun, Bengio, Hinton, Nature, 2015
Deep learning with Python, F. Chollet, Manning, 2017.
Elements of statistical learing, Hastie, Tibshirani, Friedman, Springer-Verlag, 2017.
Scikit-learn: Machine Learning in Python, Pedregosa et al., JMLR 12, 011
GPU enhanced Neuronal Networks, pyGeNN.
Scientific Computation Scipy Lecture Notes, Varoquaux, Gouillart, Vahtras, de Buyl, 2020.
Neuromorphic hardware Loihi 2: A New Generation of Neuromorphic Computing