Machine Learning–Enhanced Localized Surface Plasmon Resonance Sensing: From Spectra to Intelligent Devices
TL;DR
This review evaluates how machine learning can improve information extraction when the model is matched to the LSPR measurement physics, data structure, and validation regime and connects physics‐first features, dimensionality reduction, supervised learning, convolutional neural networks, attention models, and physics‐informed strategies to preprocessing, classification, concentration prediction, nanostructure inversion, device optimization, and adaptive measurement.
Abstract
Localized surface plasmon resonance (LSPR) sensing converts perturbations of the local dielectric environment, nanostructure coupling, and resonance line shape into optical observables such as extinction or scattering spectra, resonance shifts, intensity changes, linewidth variation, asymmetry, and time‐resolved trajectories. These readouts support many label‐free chemical and biological assays, but their interpretation can be confounded by weak spectral changes, nanostructure heterogeneity, aggregation, drift, and mass‐transport effects. This review evaluates how machine learning (ML) can improve information extraction when the model is matched to the LSPR measurement physics, data structure, and validation regime. We connect physics‐first features, dimensionality reduction, supervised learning, convolutional neural networks (CNNs), attention models, and physics‐informed strategies to preprocessing, classification, concentration prediction, nanostructure inversion, device optimization, and adaptive measurement. Representative studies are organized by plasmonic modality, observable, inference target, data‐splitting unit, and evidentiary scope. Particular attention is given to task‐dependent physical priors, nonmonotonic Fano responses, near‐field and binding saturation, label uncertainty, chip‐to‐chip variation, mass‐transport‐limited kinetics, and mechanism‐specific drift. ML can improve estimation precision and analytical discrimination within validated operating ranges, but it cannot recover information absent from the measurement, resolve a non‐identifiable inverse problem without additional observables or priors, or make extrapolation beyond the calibrated domain inherently safe.