Advanced signal processing

We investigate advanced signal processing methods for optical spectroscopy. Signal processing is key to bringing infrared and terahertz spectroscopy into practical applications. In simple, well-known configurations, model-based signal processing approaches are reliable and consistent.[1,2] However, when moving to more complex systems or unknown environments, model-based methods often fail. For this reason, our research focuses on developing robust signal processing techniques that can adapt to varying conditions and uncertainties inherent in real-world scenarios. By applying data-driven approaches, such as convolutional neural networks, we aim to enhance the accuracy and reliability of optical spectroscopy across a broad spectrum of applications.[3] In particular, we are exploring hybrid machine learning models that combine the prior knowledge of traditional model-based methods with the adaptability of modern machine learning techniques. These hybrid approaches enable us to process optical data with the flexibility of neural networks, while constraining these models to physically feasible networks, resulting in more accurate and resilient signal processing solutions. Our goal is to bridge the gap between theoretical models and practical implementation, ensuring that optical spectroscopy can be effectively utilized in diverse and dynamic environments.

[1] van Mechelen, J. L. M., Kuzmenko, A. B. & Merbold, H. Stratified dispersive model for material characterization using terahertz time-domain spectroscopy. Opt. Lett. 39, 3853–3856 (2014).
[2] van Mechelen, J.L.M., Frank, A. & Maas,D.J.H.C. Thickness sensor for drying paints using THz spectroscopy. Opt. Express 29, 7514 (2021).
[3] Koumans, M., Meulendijks, D., Middeljans, Peeters, D., H., Douma, J.C., and, van Mechelen, D. (2024). Physics‐assisted machine learning for THz time‐domain spectroscopy: sensing leaf wetness, Sci. Reports 14:7034 (2024).