Subject Area

Computer Science

Abstract

Rotational Spectroscopy is a powerful spectral fingerprinting approach that can be used for identifying different gas molecules in a sample. Gas molecules are free to rotate, with inertia, in fixed states of quantized energy. In Rotational Spectroscopy, radiative beams are shown onto a sample to cause an energy-based transition between quantized rotational states. By sweeping the frequency of the radiative beams and monitoring the absorption with a sensor, one can profile the different rotational states, monitoring for energy based transitions. These transitions are dependent on unique properties of the molecules, thus presenting a unique molecular identification fingerprint (in the form of a spectrum) that facilitates molecular detection. In other words, rotational spectroscopy uses the transition of polar molecules (with permanent electric dipole) between rotational states to identify different molecules in a sample as low as parts-per-trillion (ppt) with absolute sensitivity and specificity. Core spectrometers have the ability to detect gas molecules up to 32 chemical mixtures with absolute specificity in less than 10 minutes. This approach can be used for indoor air quality monitoring, breath analysis, and many more applications that require detection of molecules in a sample. The use of submillimeter and terahertz spectrometers for chemical spectroscopy and imaging has proven to be effective but not practical. Today, spectrometers are made of complex semiconductor devices that are bulky and costly, and therefore, they are not extensively used in the industry and not yet commercialized. Recent improvements in the development of Complementary Metal Oxide Silicon (CMOS) integrated circuits may prove transformational for rotational spectrometers, using far-infrared sensing for detecting rotational states. The use of CMOS would significantly impact the cost-effectiveness of rotational spectrometers, and consequently, the application of such technology will improve immeasurably. The major challenge of using CMOS in rotational spectroscopy fingerprinting is the increased amount of noise in the output spectrum. Various noise sources can reduce the reliability of rotational spectrometers in identifying molecules in the sample accurately. Therefore, as the rotational spectrometers’ hardware feasibility with CMOS sensing is proven, a software application that can detect the fingerprints of molecules even in the presence of noise is essential. This research aims to create such a software application that can prove and establish CMOS sensors’ application in rotational spectroscopy fingerprinting. In this work, we developed a MATLAB-based software suite that serves both as an explanatory and simulation tool for generating CMOS-like sample output files. We characterize stochastic effects gathered from CMOS sensors and develop a library of parameters and ground truth chemicals that can be used and manipulated to generate a database of simulated files. Overall, we created 16,040 simulated gas mixtures with a combination of 50 different ground truth chemicals imported from the NASA Jet Propulsion Laboratory (JPL) molecular library and 9 different parameter selections. Each simulation file contains a spectrum of millions of frequencies. Furthermore, we developed an add-on to the tool for identifying molecules using a spectral matching approach. This add-on serves as a baseline method for detecting known molecules from a library in a simulated CMOS spectrum (i.e., detecting ground truth in a sample file). Imported sample files are tested for each of the molecules (or chemical mixes) avail- able in our library (50 ground truth molecules in various mixtures). Each peak matching simulation report includes the percentage of spectral match between the CMOS output and ground truth. By taking advantage of our comprehensive database, we evaluated this base- line method for matching spectra and detecting molecules in a sample CMOS output file. Our peak matching algorithm accuracy reports area under the curve (AUC) values ranging vi from 0.93 to 0.96 depending on the sample file’s chemical mix and noise parameters. The receiver operating characteristic (ROC) curve and AUC for our peak matching algorithm show a promising future for using a CMOS sensor to find chemical fingerprints in a sample file. However, depending on the application of this software and algorithm, the algorithm’s specificity may need improvement. In addition to this baseline method, we also applied Machine Learning and Deep Learning approaches to improve detection reliability under noisy conditions. Specifically, we developed two models: (1) a convolutional network (ConvNet) for peak-based molecular detection and (2) a contrastive learning (CL) model to enhance molecular identification by focusing on unique spectral fingerprints. These models were evaluated alongside the baseline method to compare the performance of each approach in spectral detection accuracy and specificity. Our results indicate that both the ConvNet peak finding and CL models provided robust molecular identification under certain high-noise conditions, with selective improvements in sensitivity and specificity compared to the baseline spectral matching algorithm. This com- parison will inform the design of future architectures in this field and clarify the trade-offs in each method’s implementation, including calibration requirements and detection reliability. At the conclusion of our research, we aim to detect meaningful molecular spectra in a noisy dataset and accurately match these spectra to compound molecules in a mixture. Our methods demonstrated potential to boost the sensitivity and specificity of molecular identification with CMOS sensors, bringing the process closer to the acceptability standards required for industry.

Degree Date

Winter 12-2024

Document Type

Dissertation

Degree Name

Ph.D.

Department

Computer Science and Engineering

Advisor

Eric C. Larson

Format

.pdf

Share

COinS