Molecular Screening for Terahertz Detection with Machine-Learning-Based Methods

The molecular requirements are explored for achieving efficient signal up-conversion in a recently developed technique for terahertz (THz) detection based on molecular optomechanics. We discuss which molecular and spectroscopic properties are most important for predicting efficient THz detection and outline a computational approach based on quantum-chemistry and machine-learning methods for calculating these properties. We validate this approach by bulk and surface-enhanced Raman scattering and infrared absorption measurements. We develop a virtual screening methodology performed on databases of millions of commercially available compounds. Quantum-chemistry calculations for about 3000 compounds are complemented by machine-learning methods to predict applicability of 93 000 organic molecules for detection. Training is performed on vibrational spectroscopic properties based on absorption and Raman scattering intensities. Our top molecules have conversion intensity two orders of magnitude higher than an average molecule from the database. We also discuss how other properties like molecular shape and self-assembling properties influence the detection efficiency. We identify molecular moieties whose presence in the molecules indicates high activity for THz detection and show an example where a simple modification of a frequently used self-assembling compound can enhance activity 85-fold. The capabilities of our screening method are demonstrated on narrow-band and broadband detection examples, and its possible applications in surface-enhanced spectroscopy are also discussed.

Popular Summary

Novel techniques for detecting terahertz radiation have great potential for applications in medicine, communication, and astronomy. A recently proposed device for terahertz detection would harness the ability of molecular vibrations to convert terahertz to visible light and use plasmonic nanoantennas for enhancing signals. A key element for realizing this detection technique is the identification of molecules capable of efficient terahertz-to-visible conversion when placed in plasmonic nanocavities. We present the first detailed investigation of a wide range of chemical compounds to detect terahertz radiation. We propose key molecular properties to be considered, and we identify the most suitable molecules for this application.

Our computational approach is based on accurate quantum chemistry and machine-learning calculations that efficiently and accurately predict relevant molecular and spectroscopic properties. We select the most suitable molecules from commercially available compound databases containing millions of compounds. Absorption and Raman-scattering experiments for a selection of compounds confirm the accuracy of our computational methodology and validate the presence of normal modes that can be harnessed in the terahertz detector.

Our top-ranking molecules have more than 2 orders of magnitude higher conversion intensity than molecules commonly used in plasmonic experiments. The machine-learning model also provides insight into the nature of the conversion process and identifies functional groups that enhance the performance of molecules.

This work presents a key step in achieving efficient terahertz detection with molecular nanocavities and our top-ranking molecules provide the first proposal for customized detector architectures enabling specialized applications in diverse areas.

Figure 1

The THz detection process. Absorption of THz radiation by a molecular vibrational mode in a molecular nanocavity increases population of excited vibrational levels, which is detected through the increased intensity of the SERS anti-Stokes signal of the vibrational mode in question.

Figure 2

Comparison of calculated and experimental Raman Stokes properties of a set of test molecules. (a) Corrected solution-phase (${\tilde{R}}_{\exp }/c$) vs calculated (${\tilde{R}}_{\mathrm{calc}}$) integrated intensities. (b) NPOM measured (${\tilde{R}}_{\exp }$) vs corrected calculated (${\tilde{R}}_{\mathrm{calc}}E{S}^{-1}$) integrated intensities. (c) Measured molecules. (d) Mean experimental and calculated spectra for NPOM systems along with spectral distances ($d$) between experiment and calculation.

Figure 3

Distribution of $P$ (standardized logarithm of summed conversion intensity) among (blue) known self-assembly materials, (orange) randomly chosen molecules, and (green) the QC database. Some molecules with (a) low, (b) moderate, and (c)–(g) high potential for THz detection are also depicted.

Figure 4

(a) ML results for target $P$ with the EN model. Features with largest weights in the trained EN model for (b) THz conversion $P$, (c) absorption $A$, and (d) Raman $R$ targets.