Machine learning-assisted design and optimization of a hybrid BK7/Ag/WS2/GaN/ZrN/graphene optical SPR biosensor for label-free brain tumor-related refractive-index sensing

The collection contains 25 data CSV files and 24,831 records covering absorption, resonance response, electric-field profiles, sensing indicators and machine-learning performance for the hybrid BK7/Ag/WS2/GaN/ZrN/graphene SPR sensor examined in the associated numerical study. The development of the hydrogen economy—electrolysers, storage in hydrides and liquid carriers, hydrogen blending into the gas grid, and refuelling infrastructure—is moving hydrogen from laboratory-scale applications into industrial facilities and populated areas. Hydrogen is colourless and odourless, is flammable at concentrations of 4–75 vol.% in air, and has an ignition energy on the order of 0.02 mJ. Reliable leak detection, in-line monitoring of process media, and rapid characterisation of new storage materials are therefore prerequisites for safe scale-up, not optional extras. Established sensing principles have inherent limitations in this environment: catalytic (pellistor) sensors require oxygen and operate with a heated element; metal-oxide semiconductor (MOX) sensors require elevated temperatures and suffer from cross-sensitivity; and electrochemical cells exhibit drift and have a limited service life. Optical plasmonic sensors offer a sensing point that requires no electrical power, immunity to electromagnetic interference, and the possibility of fibre-optic multiplexing with remote interrogation—features that are essential in potentially explosive atmospheres (IEC 60079) and across large industrial facilities. The performance of a surface plasmon resonance (SPR) sensor is not determined by a single “smart” material, but by the interplay between its constituent layers and optical configuration. The plasmonic metal, ultrathin 2D and dielectric interlayers, the top functional layer, the wavelength, and the coupling optics form a coupled design space comprising five or more thickness parameters, alongside material choices. Each point in this space requires an electromagnetic calculation, while experimentally validating every configuration is costly and time-consuming for multilayer structures incorporating 2D materials. A data-driven surrogate—a one-dimensional convolutional neural network (1D CNN) trained on synthetic spectra—makes this space accessible: it enables configuration screening, tolerance analysis accounting for fabrication deviations, and the inverse mapping from spectrum to measurand without manually extracting resonance parameters. For hydrogen sensing, machine learning (ML) offers an additional advantage: the response of palladium and hydride layers is nonlinear, hysteretic, and temperature-dependent—precisely the type of relationship that a model trained on the full spectral shape handles better than linear calibration. Hydrogen sensing introduces additional physics into the general SPR problem: the sensing variable is not the refractive index of an external analyte, but the change in the complex permittivity of a solid layer (PdHₓ, MHₓ) as a function of hydrogen content, accompanied by a phase transition, lattice expansion, surface roughening after cycling, and surface poisoning. Introducing this physics simultaneously with a new computational and ML framework would make it impossible to distinguish model errors from errors in the material data. A paper based on a well-defined model analyte therefore serves as a controlled reference case: (i) it verifies agreement between analytical transfer-matrix method (TMM) and full-wave finite-element method (FEM) solutions for a problem with established literature benchmarks; (ii) it establishes a methodology for data generation, noise augmentation, cross-validation, and out-of-distribution (OOD) testing of the surrogate; (iii) it quantifies the contribution of individual layers and robustness to fabrication tolerances; and (iv) it produces reusable outputs—a spectral dataset, a trained model, and code—which the subsequent hydrogen-focused phase can extend rather than develop from scratch. This follows the same rationale stated in the attached paper: a model analyte provides a controlled numerical framework without presupposing a specific target substance. The model analyte need not be biomedical. For an energy research centre, a more natural choice is a reference series with tabulated refractive indices, such as water–glycerol mixtures or KOH solutions, which also establishes the measurement protocol for subsequent process monitoring of electrolytes. The project association supplied for this deposit is Increasing the resilience of power grids in the context of decarbonisation, decentralisation and sustainable socioeconomic development, award CZ.02.01.01/00/23_021/0008759, supported by European Union funds through the Operational Programme Johannes Amos Comenius.