Machine Learning-Assisted Terahertz Metamaterial-Based Sensing
Translated title
Maskinlæring til Terahertz Metamateriale-Baseret Detektion
Author
Bruun, Lasse Westfall
Term
4. semester
Education
Publication year
2026
Submitted on
2026-05-28
Pages
79
Abstract
This thesis investigates a non-referenced terahertz (THz) spectroscopic sensing platform based on resonant antenna (metamaterial) arrays on silicon to classify sugars (fructose, glucose, lactose and sucrose) at concentrations down to 25 mg/dL across 0.3–2.3 THz. The antennas were designed as half-wavelength resonators adapted to material and substrate effects at THz frequencies, fabricated using standard UV lithography and lift-off, and characterized by continuous-wave THz frequency-domain spectroscopy. Because wafer-to-wafer fabrication differences dominated the measured spectra and made manual peak inspection ineffective, machine learning was employed with data normalization, dimensionality reduction, and multiple classifiers (including Bayes, k-nearest neighbors, and support vector machines), combined in classifier ensembles that average class probabilities. Dimensionality reduction produced projections that separated sugar concentrations into distinct clusters; individual models achieved up to 85.19% accuracy, improved to 90.74% with ensembles, while amplitude-only models performed markedly worse (best 67.59%), underscoring the value of including phase alongside amplitude. A limitation was insufficient wafer variety in the training data, causing a spatial offset between training and test sets. The results indicate that combining resonant THz antennas with machine learning can enable robust, non-referenced chemical sensing, with further robustness expected from improved lithography mask layouts that allow each sugar to be measured on a wider variety of wafers.
Dette speciale undersøger en ikke-refereret terahertz (THz) spektroskopisk sensorplatform baseret på resonante antennearrays (metamaterialer) på silicium til at klassificere sukkerarter (fruktose, glukose, laktose og sukrose) ved koncentrationer ned til 25 mg/dL i området 0,3–2,3 THz. Antennerne blev designet som halv-bølgelængde resonatorer, hvor skaleringslove blev tilpasset materialers og substraters indflydelse i THz-området, fremstillet med standard UV-litografi og lift-off og karakteriseret med kontinuert-bølge THz frekvensdomænespektroskopi. Da wafer-til-wafer fabrikationsforskelle dominerede de målte spektra og gjorde manuel peak-inspektion utilstrækkelig, blev maskinlæring anvendt med datanormalisering, dimensionsreduktion og flere klassifikatorer (bl.a. Bayes, k-nærmeste nabo og supportvektormaskiner) samt classifier-ensembler, der gennemsnitter sandsynligheder. Dimensionsreduktion fandt projektioner, der adskilte sukkerkoncentrationer i tydelige klynger; enkeltmodeller opnåede op til 85,19 % nøjagtighed, forbedret til 90,74 % med ensembler, mens amplitude-only modeller klarede sig markant dårligere (bedst 67,59 %), hvilket understreger værdien af at inddrage fase sammen med amplitude. En begrænsning var utilstrækkelig wafer-variation i træningsdata, som gav et rumligt offset mellem trænings- og testdata. Resultaterne indikerer, at kombinationen af resonante THz-antenner og maskinlæring kan muliggøre robust, ikke-refereret kemisk sensing, som kan styrkes yderligere med forbedrede litografimaskelayouts, der muliggør målinger af hver sukkerart på flere wafers.
[This apstract has been generated with the help of AI directly from the project full text]