# Machine learning-driven optimization of monolithic gold plasmonic sensors: Achieving ultrahigh sensitivity with interpretable linear models

## TL;DR

This paper combines machine learning and nanophotonic design to create a highly sensitive biosensor using gold-coated photonic crystal fibers.

## Contribution

The novel integration of interpretable linear machine learning models with optimized gold plasmonic sensor structures achieves record sensitivity.

## Key findings

- The sensor achieves a record wavelength sensitivity of 31,846.46 nm/RIU-1 with minimal variation across biological refractive indices.
- Multiple Linear Regression outperforms nonlinear models in predicting optical responses with high accuracy (MAE/RMSE values).
- The design simplifies fabrication while surpassing prior ML-enhanced plasmonic sensors in resolution and noise resilience.

## Abstract

Integrating machine learning (ML) with nanophotonic engineering, this work achieves unprecedented performance in surface plasmon resonance (SPR) biosensing through a co-designed gold-coated photonic crystal fiber (PCF-SPR) sensor and multi-algorithm computational framework. An asymmetric circular PCF structure with concentric air-hole rings (Λ1=3.26 μm, Λ2=2.12 μm) and a 50 nm gold layer maximizes evanescent field-analyte overlap, generating complex spectral signatures ideal for machine learning interpretation. High-fidelity COMSOL Multiphysics simulations produce 1560 synthetic data points across refractive indices (RIs) of 1.33–1.38, capturing confinement loss, wavelength sensitivity, and effective permittivity. Three regression models—Multiple Linear Regression (MLR), Support Vector Regression (SVR), and Random Forest Regression (RFR)—are rigorously evaluated for predicting optical responses. The sensor demonstrates a record wavelength sensitivity of 31 846.46 nm/RIU-1 at RI=1.33, with minimal variation (0.02%) across the biological range, alongside a resolution of 1.57×10−3 RIU. Crucially, MLR outperforms nonlinear counterparts, achieving superior accuracy in confinement loss (MAE = 3.97, RMSE = 5.03) and sensitivity prediction (MAE = 40.18, RMSE = 50.54). This synergy of optimized pure-gold microstructures and interpretable machine learning establishes a robust pipeline for high-sensitivity, noise-resilient biosensing, surpassing prior ML-enhanced plasmonic sensors in critical performance metrics while simplifying fabrication.