A Bayesian Generative Model With Gaussian Process Priors For Thermomechanical Analysis Of Micro-Resonators

Thermal analysis using resonating micro-electromechanical systems shows great promise in characterizing materials in the early stages of research. Through thermal cycles and actuation using a piezoelectric speaker, the resonant behaviour of a model drug, theophylline monohydrate, is measured across the surface whilst using a laser-Doppler vibrometer for readout. Acquired is a sequence of spectra that are strongly correlated in time, temperature and spatial location of the readout. Traditionally, each spectrum is analyzed individually to locate the resonance peak. We propose a Bayesian model using a warped Gaussian process prior taking the correlations into account and demonstrate on both synthetic and experimental data, that it yields better estimates of both location and amplitude of the resonance peak. Thus, the proposed model can give a more precise characterization of drugs, which is important in drug discovery and development.