A major bottleneck in the analysis of broadband chirped-pulse microwave spectra is the identification of unknown molecules. Often, a set of molecular frequencies are fit to an effective Hamiltonian, whereby a set of spectroscopic parameters are used to infer the molecular carrier. These constants are reproduced with electronic structure calculations through a trial and error process, accompanied by chemical intuition involving the precursors used, and the reaction conditions (e.g. electrical discharges). As the size of the molecules increase, the combinatorics of many hundreds to thousands of possible isomers becomes an intractable problem for chemical intuition alone. Since spectroscopic parameters are only weakly informative, we turn to statistical inference to complement conventional spectroscopic analysis. In this talk, I will discuss a new framework for identifying unknown molecules by performing inference on spectroscopic parameters with probabilistic deep learning. Using a series of decoder architectures, we are able to infer the approximate molecular composition/formula and what functional groups are present using only the rotational constants and derived quantities κ (the asymmetry parameter) and ∆ (the inertial defect), and approximate magnitudes and projections of the dipole moments.

With the development of high resolution and high bandwidth microwave spectrometers, we have reached a state where unbiased spectral line surveys of completely unknown mixtures can be systematically deconvolved, given sufficient time and effort. In our recent work on various discharge mixtures of benzene, we have shown that spectra comprising up to 220 distinct spectral carriers-including vibrationally excited states and isotopic species, with  ∼ 60 species unknown prior to this work-can be gradually disentangled following weeks of fervent analysis. As we progress to increasingly complex and unknown mixtures, rotational spectra will become proportionally complicated and congested; conventional spectroscopic analysis cannot match the pace of data collection in astronomical observations and laboratory experiments. In this talk, I will detail some efforts in our group of the recent months that leverage high-speed computation afforded by deep learning models. Inspired by the work of Zaleski and Prozument (2018), we sought to develop probabilistic deep learning models that are able to spectrally deconvolve multi-component spectra in an automated fashion. Within a Bayesian framework, these models are designed to operate with minimal information, requiring only spectral frequencies, and estimates the likelihood of sequences of frequencies originating from the same carrier. In addition to a description of the models, I will discuss how these models act as companions for spectral analysis, converting the typical user workload from a process of trial and error to one of statistical inference.

Accurate spectroscopic modeling is critical when measuring time-resolved, state-specific chemical kinetics of oxygen (O₂) in air. Here, a spectroscopic model was developed to simulate oxygen absorption cross-sections in the Schumann-Runge system under vibrational non-equilibrium conditions. The model calculates the spectral bound-bound O₂ cross-sections using a line-by-line calculation and convolutes this cross-section with the Gaussian lineshape of a pulsed laser system to yield an effective bound-bound cross-section. This effective bound-bound cross-section was then augmented with an effective bound-unbound continuum database to provide a total effective cross-section. Two wavelengths, 211.2 nm and 236.9 nm, were chosen for spectroscopic study, as the effective cross-sections at these wavelengths primarily probe the fourth (v”=4) and sixth (v”=6) vibrational levels of ground state O₂, respectively. Furthermore, the effective cross-sections at both wavelengths demonstrate strong sensitivity to vibrational temperature, and vary only slightly with translational and rotational temperature. Cross-sections in shock-heated O₂ have been measured using a picosecond pulsed ultraviolet (UV) laser, and the results have been compared against the model developed here and the Spectrum Model developed by Bykova et. al. Cross-sections were measured behind reflected shocks in 2% and 5% O₂ in argon (Ar) mixtures around 211.2 nm and 236.9 nm up to initial post-reflected shock temperatures of 10,700 K and subatmospheric pressures. For both wavelength regimes, the experimental cross-sections agree to within 15% of the model developed here, and most experimental values fall within 10% of both spectroscopic models. Using the spectroscopic model developed here to inform appropriate wavelength selection, the pulsed UV laser system becomes a tool for directly tracking both vibrational temperature and populations in specific vibrational modes of O₂ as it undergoes vibrational relaxation and dissociation behind strong shock waves. These population time histories provide important experimental data needed to evaluate current computational models that seek to capture the molecular energy transfer present in high-enthalpy airflows.

Vibrational spectroscopy is known to be a reliable technique for qualitative and quantitative chemical analysis of various biological fluids. Such analysis is rather efficient when molecular compound under study is present in the fluid with high concentration. There are many cases when determination of compounds at extremely low concentration is needed. Such cases include detection pharmaceuticals or their metabolites in blood, detection of cancer cells in biological fluids, etc. In such situations conventional vibrational spectroscopy techniques - infrared absorption and Raman scattering are not suitable due to too low sensitivity, therefore other more sensitive spectroscopic methods have to be used. This work presents some applications of colloidal SERS spectroscopy for the detection of medicine metabolites in blood and for detection of various cancer related chemical substances in intracellular and extracellular fluids. Various gold and silver nanoparticle colloidal solutions were used for the implementation of SERS spectroscopy and the best ones (giving the highest enhancement with the least interference to the spectra) were determined. We conclude that the SERS approach can be applied for screening of pharmaceuticals and drug usage. In case of aspirin consumption, the detection limit was found to be down to mild toxicity. Such approach also allows detection of cancerous tissue areas and, therefore, can be used during surgical operations for exact determination of tumour boundaries.

Triacetone triperoxide (TATP) is a known peroxide-based explosive frequently used by terrorists in improvised devices due to its ease of synthesis and detection difficulty. Here we propose microwave spectroscopy as a highly selective detection method. The main conformer of TATP (95%) has D₃ symmetry and thus it has no dipole moment. However, TATP-water adducts are polar species and their microwave observation could open a way to detect TATP in the gas-phase. In this work we have observed the spectrum of TATP-H₂O and TATP-(H₂O)₂ which, on the other hand, have revealed the interesting properties of TATP. In TATP-H₂O water lies close to the C₃ axis and rotates around it. As a result, the ground state rotational constants average to those of a symmetric top. The complex exhibits a four center trifurcated O_w-H···O hydrogen bond further stabilized by a weak four center C-H···O_w chelated interaction. To our knowledge, those kinds of interactions have not been observed previously in the gas phase. The anomalous large O-H···O distances associated to the main hydrogen bond interaction evidences that TATP oxygen atoms are to some extent encapsulated by their hydrocarbon environment. The dihrydated cluster can be envisaged as formed from the TATP-H₂O adduct with the second water molecule acting as double hydrogen bond donor to the first water molecule and the closest TATP endocyclic oxygen atom.

Laser desorption time-of-flight mass spectrometry (LD-ToF-MS) experiments on pressed-pellet samples of polycyclic aromatic hydrocarbons (PAHs) exhibit the formation of covalently-bonded dimers at masses (m/z) of 2M-2 and 2M-4 (where M is the parent mass). Through replication of these LD-ToF-MS conditions at higher throughput, PAH dimers have been produced and collected in milligram quantities. For collected samples of pyrene, perylene, and coronene, differential sublimation has isolated the dimer sample from residual monomers. After confirmation of the purification using LD-ToF-MS, the samples are analyzed using a variety of spectroscopic methods, including UV-Vis, IR, and Raman spectroscopy. Theoretical calculations of dimer samples have been done using density functional theory with the CAM-B3LYP method at the def2TZV level of theory. Theory calculations were calibrated and checked by comparison with the monomer samples, and used to determine the lowest energy dimer structures. The simulated spectra were compared with collected spectra of isolated dimers to determine the actual structures of the dimerized PAHs.

The isomeric distribution of a popular fluorescent laser dye DCM (4-(Dicyanomethylene)-2-methyl-6-(4-dimethylaminostyryl)-4H-pyran) has been investigated both in methanol solution and gas phase using the multi-dispersive experimental technique involving Ion Mobility Mass Spectrometry (IMMS) and High Performance Liquid Chromatography (HPLC). There could be four rotameric forms, trans1, trans2, cis1, and cis2. Chromatographic separation of the isomers in methanol solution shows two distinctly separated peaks at retention times of 3.73 (peak-I) and 3.87 (peak-II) minutes. In the case of optical probing, the former peak appears much weaker compared to the latter when probed at the absorption maximum of the dye at 467 nm. On the other hand, in mass spectrometric probing corresponding to the molecular mass (DCMH⁺), peak-II is found to be nearly twice as intense as peak-I. The ion mobility measurements have been performed with the two LC separated solution fractions and in each case, three peaks, A, B and C, are observed, whereby the collision cross-section (CCS) values corresponding to these peaks are found to be 174, 185 and 197 Ų, respectively. Interestingly, the relative intensities of these three peaks in the two IM spectra appear in opposite sequences. The minimum energy geometries of the neutral and protonated isomers have been computed by electronic structure theory method and the CCS values of the protonated isomers have been calculated using the MOBCAL program. The analysis reveals that the LC peak-I is associated with the cis2 isomer, whereas the other three isomers contribute collectively to peak-II. By simulating the electronic absorption spectral profiles of the isomers, we have proposed that the distributions of the trans1, trans2, cis1, and cis2 isomers are 33.5, 61.5, 2.0 and 3.0%, respectively, in the solution phase. The IM spectral features also indicate occurrences of isomeric inter-conversions during the ESI process. The collision-induced dissociation of the DCMH⁺ ions corresponding to the LC separated solution fractions reveals that the fragmentation behavior is also isomer dependent.

By means of chirped and cavity microwave spectroscopies, automated double resonance, new high-speed fitting and deep learning algorithms, and large databases of computed structures, the discharge products of three benzene mixtures - alone or with oxygen and nitrogen - have been exhaustively characterized between 6.5 and 26 GHz. In total, more than 3300 spectral features were observed; 88% of these, accounting for roughly 97% of the total intensity, have now been assigned to 160 distinct chemical species and 60 of their variants (i.e. isotopic species and vibrationally excited states). Roughly 50 of the products are entirely new or poorly characterized at high resolution, including many heavier than the precursor benzene. These findings provide evidence for a rich architecture of two- and three-dimensional carbon, and indicate that benzene growth, particularly formation of ring-chain molecules, occurs facilely under our experimental conditions. The present analysis vividly demonstrates the utility of microwave spectroscopy as a precision tool for complex mixture analysis, irrespective of whether the rotational spectrum of a product species is known a priori or not. From this large quantity of data, for example, it is possible to derive a mass spectrum for each discharge that is analogous to a traditional mass spectrum, but with exquisite isomeric resolution.

A high-resolution cavity enhanced absorption spectroscopy (CEAS) technique is presented for infrared absorption studies at temperatures between 77K and 300 K. An optical cavity with a pathlength of several kilometers is attached to a low temperature cryostat. The spectra are obtained with a tunable diode laser with resolution 0.0003 cm⁻¹. The output of the laser is modulated with an electro-optic modulator and coupled to an optical cavity. To illustrate the use of the technique, the first overtone of carbon monoxide at 298 K and 80 K and the first overtone transition of the acetylenic C-H stretch of propyne will be presented. The experiments are performed at pressures from 70 mTorr to 1 Torr. The combination of low pressures and temperatures makes this technique ideal for studies related to astrochemistry. The experimental set up will be used to simulate in the laboratory the kinetics and spectroscopy of molecules in the atmosphere of planets and satellites of the outer solar system.

l0pt Figure Raman Spectroscopy is a great way to distinguish different kinds of API polymorphs.Shijie Zhang, et al. Dynamic Sparse Sampling for Confocal Raman Microscopy, Analytical Chemistry 2018owever,adversarial attacks on spectral classifiers were shown to enable identification of potential vulnerabilities in common dimension reduction analyses of Raman spectra. We tackled this susceptibility with the Generative Adversarial Linear Discriminant Analysis (GALDA) approach. GALDA is analogous to Generative Adversarial Nets (GAN) in the machine learning context. Conceptually in a typical GAN, two models were simultaneously trained, a generative model G that attempts to estimate the sample distribution and a discriminative model D that classifies the output of G.These two models seek to achieve Nash equilibrium during the iterative adversarial training process.Goodfellow, Ian, et al. Generative adversarial nets, Advances in neural information processing systems. 2014e herein incorporated this concept into spectral classification. Spectral classification in the spectroscopic context seems to lag albeit the mass development in the computer science area. As such, the analyzation methods' robustness and susceptibility to malicious attack is considered even less. Therefore, GALDA aims to test against the robustness of LDA classifiers and provide a new framework for considerations for classification strategies less susceptible to spurious misclassification.

---

Footnotes:

Shijie Zhang, et al. Dynamic Sparse Sampling for Confocal Raman Microscopy, Analytical Chemistry 2018H Goodfellow, Ian, et al. Generative adversarial nets, Advances in neural information processing systems. 2014W