A Technique for Generating Realistic Synthetic Photoluminescence Time-Series Data with Quenching Events

Abstract

A novel technique is presented for generating realistic synthetic photoluminescence (PL) datasets to support the development of fluorescence-based sensor algorithms for chemical threat detection. Collecting large, well-annotated experimental datasets is costly and slow, often taking months due to complex sample preparation and the handling of toxic analytes. In contrast, the proposed simulation approach requires only 60 to 110 experimental samples to capture system dynamics and generate virtually unlimited synthetic traces. This method replicates the statistical and physical properties of sensor–analyte pairs, allowing customization of analyte concentration, exposure time, and event prevalence. PL traces are modeled in two parts: baseline drift and analyte-induced response. Baseline drift is captured using principal component analysis (PCA), while kernel density estimation (KDE) models the continuous distribution of the PCA scores to enable the stochastic generation of new, realistic baselines. Component trajectories are fitted to constrained exponential functions for temporal extrapolation, and intensity-dependent noise is incorporated based on modelling after wavelet decomposition. Detection events are simulated using a thin-film diffusion model coupled with a Stern-Volmer relation. Validated against experimental data, this scalable pipeline mimics natural variability and significantly reduces reliance on costly experimental data acquisition.