NMF regularization techniques for unmixing frequency comb data

We use unsupervised machine learning techniques to investigate how optical frequency combs may be used to identify molecular composition in mixed-sample environments. Optical techniques such as dual-comb heterodyne detection with GHz and THz repetition rates are often used as spectroscopic methods for probing ro-vibrational IR spectra of small molecules, proving to be promising in dispersive free-space sensing regimes. We simulate and recover an intensity spectrum from mixed-sample IR absorption in the region spanned by the comb bandwidth. We present several inexpensive, efficient machine learning methods to analyze optimal comb frequencies to maximize the ability to determine the molecular composition of a sample, based only on the interactions at those frequencies with respect to various channels of noise. Using a synthetic dataset, these methods are able to accurately identify molecules using simulated frequency combs, comparable to the ones used in practice. Furthermore, these methods withstand added noise beyond what is expected in real applications. The robustness of these methods suggests the techniques presented in this paper may be integrated into a laboratory setup and even in on-chip comb systems deployed in an atmospheric setting. This approach may be used as a detection framework for future quantum-enhanced systems via squeezed bi-photon comb inputs.