Inverse-designed plasmonic biosensors with LSPR-SPP-wood anomaly coupling enhanced for biomolecular analysis Authors: Yihui Yang, Youqian Chen, Rui Li, Wen Li, Mingqian Chen, Yanan Li, Wenjun Hu, Gang Logan Liu, Liping Huang Published: 2025 ABSTRACT Investigating molecular interactions is crucial for advancing biological research and therapeutic discovery. Traditional analytical techniques often face limitations in sensitivity, quantification accuracy, and simplicity. Metasurfaces support resonances that are widely explored both for far-field wavefront shaping and for near-field sensing. Here, we introduce an innovative inverse-designed multilayer metasurface plasmon resonance (IDMM-SPR) sensor that overcomes these challenges. Using a double-objective optimization method that combines numerical simulations and machine learning, we developed an IDMM-SPR sensor featuring an optimized periodic nanocup array. This design yields unparalleled sensitivity and stability by harnessing collective resonances—including localized SPR, Wood’s anomalies, and the Bloch wave SPR—which collectively enhance the sensing performance to enable the analysis of ultra-high affinity and low molecular weight interactions. The sensor achieves a figure of merit (FoM) of 26.3 and a detection limit (LOD) for C-reactive protein (CRP) as low as 0.84 ng/mL. Its compatibility with microplate absorbance readers and imaging detection makes it highly practical. The IDMM-SPR sensor shows exceptional promise for high-throughput direct molecular fishing, drug development, and disease diagnosis, offering a powerful tool for real-time, label-free affinity detection and quantitative analysis of biomolecular interactions.Graphical Abstract Inverse-Designed Plasmonic Biosensors with LSPR-SPP-Wood Anomaly Coupling Enhanced for Biomolecular Analysis. Supplementary InformationThe online version contains supplementary material available at 10.1186/s12951-025-03930-w. Introduction Molecular interactions underpin every aspect of cellular life and higher-order biological processes [1]. Detection of molecular interactions is the foundation for many important biotechnology applications in society and industry, such as drug discovery, diagnostics, and DNA sequencing [2]. The above problems usually need to be addressed at different levels of molecular complexity. Conventional assays such as ELISA [3], Western blot [4], FRET [5] or HPLC [6] can reveal tight interactions, but cannot resolve individual association or dissociation rates, are labour-intensive, low-throughput and require multiple washing and labelling steps [7]. The label-free, real-time, and high-throughput molecular interaction analysis methods include biolayer interferometry (BLI) [8], surface plasmon resonance (SPR) [9], and localized surface plasmon resonance (LSPR) [10]. Among them, SPR is the most widely adopted platform for studying antigen–antibody, drug–target, protein–nucleic acid and protein–protein interactions. Nevertheless, existing label-free techniques still face major challenges when coping with ultra-high-affinity interactions (pM range, demanding exceptional sensor stability) or with small-molecule analytes (requiring superior sensitivity) [11, 12]. There is thus an urgent need for next-generation technologies that combine higher sensitivity, specificity and instrument compatibility to satisfy the evolving requirements of modern biology. The SPR technology relies on attenuated total internal reflection to measure changes in the refractive index (RI) at the sensor surface, which are induced by variations in surface loading [13]. With the advancement and refinement of micro/nanofabrication technologies, materials science, and related theories, the field of micro-nano optics has witnessed numerous intriguing discoveries and applications. A growing number of novel devices based on optical and micro-nano technologies have been developed for various fields. Examples include plasmonic logic gates with dual optical communication bands designed on a dielectric–metal–dielectric structure [14], and directional coupler switches realized on monolithic lithium niobate for high-efficiency optical communication applications [15]. Additionally, a range of high-performance optical sensors with unique designs have been developed, such as resonator sensors based on low-cost silica–titania [16], nanosensors based on metal–insulator–metal (MIM) waveguides [17], and dielectric nanohole array metasurfaces [18], among others. Metasurface plasmon resonance (Meta-SPR) technology is an innovative technique compared to SPR, leveraging metasurface structures and metamaterials to induce resonance of free electrons and manipulate light transmission. Thus, there is no need for special angles to excite surface plasmon resonance. Metasurfaces are sub-wavelength-spaced arrays of optical scatterers, which have been widely used to overcome the limitations of traditional optical devices [19]. The constituent materials of metasurfaces can be metals or dielectrics, which are classified as plasmonic or dielectric metasurfaces, respectively. The optical response of metal metasurfaces is typically controlled by local plasmon polariton resonances induced by incident electromagnetic waves [20]. Our previous research [21] shows that by transforming two-dimensional planar thin films into uniform nano cup structures with three-dimensional structures, biosensors are endowed with unique optical properties, thus overcoming the problems of traditional SPR requiring bulky and precise optical instruments. This makes it possible to create desktop, easy-to-use MetaSPR molecular interaction instruments to meet the requirements of individuals and laboratories. However, due to the sensor’s sensitivity, many of its application ranges are still limited. In traditional sensor development, it is often necessary to repeatedly fabricate and test various sensor parameters to determine the optimal sensor specifications. To address this inefficiency, several machine learning-based approaches have been introduced to accelerate the design process [22–24]. Moreover, conventional surface plasmon resonance (SPR) sensors typically rely on detecting peak shifts and resonance angles to collect signal responses [25], which necessitates the use of specialized detection equipment. This significantly increases the development costs and time. In this work, we designed and defined the target function to inversely design a plasmonic metasurface with high optical density (OD) response and peak shift response, based on multilayer metal materials and a periodic nanocup array using machine learning. The resulting high-performance metasurface sensing chip is compatible with virtually all commercial microplate absorbance readers and spectroscopic instruments. Moreover, the metasurface exhibits high sensitivity to resonance peak shifts, which also induces color changes. Therefore, it is also highly suitable for colorimetric sensing applications. As shown in Fig. 1a, we considerate the both response of OD response and peak shift response. Subsequently, we employed a combination of three-dimensional finite-difference time-domain (FDTD) simulations and machine learning algorithms to search for the optimal parameters within the defined parameter space. This optimization process focused on key metasurface parameters, including the period, radius, height, and thickness of the multilayer metal materials (Fig. 1b). The excellent performance of the designed sensor was verified through manufacturing and experiments, and the optimal parameters were determined (Fig. 1c). Afterward, photolithography, nanoimprint lithography, electron beam evaporation, and chip functionalization techniques can be used on a large scale, quickly, and inexpensively to produce low-cost wafer-level IDMM-SPR chip sensors. By assembling with 96-well or 384-well plates (Fig. 1d), a detection device suitable for almost all microplate absorbance reader instruments can be produced, which can conveniently, quickly, high-throughput, and accurately detect molecular interactions and obtain real-time detection data (Fig. 1e). The protocol is validated through three ultra-high affinity antibodies and Quercetin affinity evaluation. Concurrently, the rapid, label-free detection of low concentrations of CRP has showcased the quantitative analysis capability of the developed sensor. Overall, these biosensors offer several advantages, including high selectivity and sensitivity to analytes, fast and in situ real-time monitoring, label-free operation, and reduced costs for target analysis. The research conducted by this institute will significantly enhance the application of metasurface sensors in areas such as drug development and disease diagnosis. Fig. 1Development schematic of high-throughput molecular interaction platform based on IDMM-SPR sensor design. (a) Definition of FoM for the IDMM-SPR sensor. (b) Main design parameters and their boundary ranges for the metasurface design. (c) Schematic illustration of the nanostructured architecture of the inverse-designed IDMM-SPR sensor. (d) The IDMM-SPR sensor chip is integrated with a universal 96-well plate to form a portable high-throughput detection device compatible with commercial microplate readers. (e) Schematic diagram of the developed sensing platform, enabling real-time monitoring of biomolecules’ rapid and high-throughput binding and dissociation processes, with the capability for regeneration and multiple detections Development schematic of high-throughput molecular interaction platform based on IDMM-SPR sensor design. (a) Definition of FoM for the IDMM-SPR sensor. (b) Main design parameters and their boundary ranges for the metasurface design. (c) Schematic illustration of the nanostructured architecture of the inverse-designed IDMM-SPR sensor. (d) The IDMM-SPR sensor chip is integrated with a universal 96-well plate to form a portable high-throughput detection device compatible with commercial microplate readers. (e) Schematic diagram of the developed sensing platform, enabling real-time monitoring of biomolecules’ rapid and high-throughput binding and dissociation processes, with the capability for regeneration and multiple detections Materials and methods Software Python3.6, FDTD Solutions (Lumerical Inc., Canada). Materials Bovine serum albumin (BSA), Phosphate-buffered saline (PBST), 11-mercaptoundecanoic acid (MUA), 1-ethyl3-(dimethylaminopropyl) carbodiimide (EDC), N-hydroxysuccinimide (NHS), Dimethyl Sulfoxide (DMSO), Fish gelatin,, ethanolamine, and phosphate-buffered saline (PBS) buffer, Polyethylene glycol 2000 (PEG) were purchased from Sigma-Aldrich; Protein A, IgG, CRP, and CRP ab, hTGFBR1 were purchased from monoclonal anti-CRP capture and detection antibodies were purchased from Beijing Sino Biological, Inc. (Beijing, China); Sintilimab, Adalimumab, Polatuzumab, Sacituzumab, PD-1, TNF-α, CD79B, TROP2, Quercetin were purchased from MedChemExpress.; Sucrose was purchased from Shanghai Macklin Biochemical Technology Co., Ltd, shanghai, China; Photocrosslinking agent was purchased from Bidepharm; Polyethylene terephthalate (PET) films were purchased from Yixing Yangguang Plastic Co. Ltd. (Yixing, China); Biotek Epoch Microporous plate spectrophotometer was purchased from Beijing agency of Biotek Instruments, Inc., Beijing, China; WeSPR 200 and WeSPR HT8 was purchased from Liangzhun (Wuhan) Industrial Co. Ltd., Wuhan, China. Definition of sensitivity and FoM of inverse design metasurface The sensitivity of an SPR sensor is the ratio of the response change of the sensor output to the change in the measured (such as RI). For instance, the sensitivity S of an SPR sensor to the RI can be written as: [26]1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:S=\frac{\delta\:y}{\delta\:{n}_{ef}}\frac{\delta\:{n}_{ef}}{\delta\:n}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\delta\:y$$\end{document} represents the sensor output. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\delta\:{n}_{ef}$$\end{document} is the effective RI on the surface, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\delta\:n$$\end{document} is the RI of the medium. The first item describes the effective sensitivity of the sensor output to surface plasma, which depends on which excitation method and modulation method the surface plasma uses [27]. The second item describes the sensitivity of the effective RI of the surface plasma to the RI. Traditional SPR almost always achieves detection by measuring the shift of the response wavelength, so sensitivity can be defined as:2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{S}_{shift}=\frac{{\mathrm{p}\mathrm{e}\mathrm{a}\mathrm{k}}_{det}-{\mathrm{p}\mathrm{e}\mathrm{a}\mathrm{k}}_{ref}}{\delta\:n}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathrm{p}\mathrm{e}\mathrm{a}\mathrm{k}}_{det}$$\end{document} is detection RI, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathrm{p}\mathrm{e}\mathrm{a}\mathrm{k}}_{ref}$$\end{document} in refference RI. To adapt to the commonly used optical instruments such as microplate absorbance readers in the market, we hope that the designed sensor chip will respond in the form of optical density. Therefore, we convert the change in peak shift into the change in light intensity by subtracting it, so the response OD can be defined as:3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{R}_{OD}=({\mathrm{O}\mathrm{D}}_{det}-{\mathrm{O}\mathrm{D}}_{ref})$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathrm{O}\mathrm{D}}_{det}$$\end{document} is the signal from the microplate reader after detection, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathrm{O}\mathrm{D}}_{ref}$$\end{document} is the signal of reference.So the sensitivity in this study can be defined as:4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{S}_{OD}=\frac{\mathrm{max}\left({\mathrm{O}\mathrm{D}}_{det}-{\mathrm{O}\mathrm{D}}_{ref}\right)-\mathrm{m}\mathrm{i}\mathrm{n}({\mathrm{O}\mathrm{D}}_{det}-{\mathrm{O}\mathrm{D}}_{ref})}{\delta\:n}$$\end{document} Another parameter is commonly used in SPR sensor design to evaluate sensor performance. Figure of merit (FoM), defined as the ratio of the sensitivity (S) to the width of the spectral peak (the full-width at half maximum (FWHM) at the resonant peak). \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:FOM$$\end{document} can be written as: [28]5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:FoM=\frac{S}{FWHM}$$\end{document} Simulation of multilayer metasurface nanocup array sensor chip FDTD Solutions (Lumerical Inc., FTDT solutions, Canada) was used to conduct simulations of the optical experiments of our nanocup biosensor chip to explore the optimal structures and illustrate the experiment results. Cr, Ag, and Au three-layer metal layers were placed on the PET film to build the multilayer metasurface modeling. An x-direction polarization plane-wave light source placed on the side of the metal surface was used to excite surface plasmon polarization (SPP) of the metasurface nanocup array. Perfectly matched layer (PML) boundary conditions along the perpendicular z-axis and x- and y-axes (each parallel to the array edge) were used with periodic boundary conditions. The mesh grid size was set at 3 nm in all axes, which was verified by convergence tests and adequate accuracy for the model of the plasmonic effects of the metal. The simulation time was set to 1000 fs. The transmission intensity was monitored at a plane 800 nm from the surface of the metal film on the PET side. Then, we change the surface RI and calculate the sensitivity we defined to evaluate the sensing performance in different RI solutions. Additional model-validation steps are detailed in Supplementary Information S2 (Supplementary Fig. S1). Particle swarm optimization calculation To design an ultrasensitive metasurface sensor, we employed the finite-difference time-domain (FDTD) method in conjunction with the particle swarm optimization (PSO) algorithm. PSO is a robust population-based stochastic evolutionary computation technique inspired by the social behavior and dynamic movements of animal groups (referred to as “particles”) as they search for optimal solutions in a given space. In this study, PSO was utilized to optimize the structural parameters of the multilayer nanocup metasurface sensor to achieve the highest figure of merit (FoM) as defined in Sect. "Materials". To simulate different refractive index environments, we selected two refractive indices, 1.3333 and 1.3397, for the optimization process. The optimization began by initializing 12 multilayer nanocup arrays with randomly distributed parameters within their respective parameter spaces. These arrays were then sent to the Lumerical FDTD platform to compute the optical density (OD) values for each structure. The FDTD platform returned the computed optical values to the PSO algorithm, which calculated the FoM and generated parameters for the next generation. After every FDTD call the returned transmission spectrum was post-processed by four hard filters that mitigate the numerical pitfalls of high-dimensional electrodynamic optimisation: Moving-average smoothing removes ultra-sharp spikes that are often artefacts of under-meshing.5 nm < FWHM < 300 nm prevents the algorithm from chasing non-physical, Dirac-like resonances that can arise when the solver has not fully converged.0 nm < ΔPeak < 10 nm avoids mis-interpreting two adjacent modes as a single shifted peak.0 < T < 1 guarantees that the transmitted energy remains physically meaningful. Moving-average smoothing removes ultra-sharp spikes that are often artefacts of under-meshing. Moving-average smoothing removes ultra-sharp spikes that are often artefacts of under-meshing. 5 nm < FWHM < 300 nm prevents the algorithm from chasing non-physical, Dirac-like resonances that can arise when the solver has not fully converged. 5 nm < FWHM < 300 nm prevents the algorithm from chasing non-physical, Dirac-like resonances that can arise when the solver has not fully converged. 0 nm < ΔPeak < 10 nm avoids mis-interpreting two adjacent modes as a single shifted peak. 0 nm < ΔPeak < 10 nm avoids mis-interpreting two adjacent modes as a single shifted peak. 0 < T < 1 guarantees that the transmitted energy remains physically meaningful. 0 < T < 1 guarantees that the transmitted energy remains physically meaningful. Particles that violate any filter receive a penalised fitness, steering the swarm away from local minima and non-optimal convergence zones. A detailed technical description of the PSO algorithm used here is provided in Supplementary Information S4. Ideally, PSO should continue iterating until all particles converge to the global optimal solution. Fabrication of metasurface nanocup biosensor chip First, Electron beam lithography processes the different parameters mask templates and then obtains the nanopillar structure templates on a silicon substrate through dry etching. Then, the periodic Ti-Ag-Au multilayer nanocup array device was prepared using a replica molding process. We evenly applied a thin layer of ultraviolet (UV)-curable polymer (NOA-61, Sigma Aldrich) onto the mold and covered it with a layer of polyethylene terephthalate (PET) film. PET, also known as poly(ethylene terephthalate), is superior in heat resistance, chemical resistance, dimensional stability, and transparency in the visible and near-infrared spectral range. This PET film serves as the substrate for the chip, providing a robust and flat surface for the deposition of the metasurface structures without significantly affecting the optical properties of the sensor. After curing with ultraviolet light (105 mW/cm2) for 3 min, the periodic nanocup pattern substrate was carefully peeled off from the mold. Finally, a chromium layer with good adhesion, a Silver layer with good plasma resonance and low cost, and a gold layer with excellent stability and biocompatibility were deposited on the nanocup substrate using electron beam evaporation technology to fabricate the nanoplasma sensors. The nanocup chip will be assembled with 96-well plates, which was fabricated by a 3D printer for convenient use in the future. Surface refractive index measurement The Nanocup chip was washed with DI water twice and then exposed to varying concentrations of aqueous sucrose solutions with 80 uL per hole. Using Biotek Epoch Microporous plate spectrophotometer (purchased from Beijing agency of Biotek Instruments, Inc.) to measure the transmission intensity spectra. The scan rate was set in normal mode, Each collected spectrum was the average of 8 individual measurements. Measurements were repeated three times under identical experimental conditions. Surface binding experiment measurement Protein A binding experiment: the chips were washed with DI water twice, and 40 µg/mL (in CBS) Protein A was immobilized to the chips for 12 h at 4 ℃. Rinse the chips with DI water after incubating the chips in 60 µg/mL BSA blocking solution last for 30 min h at 37 ℃. After surface modification, functionalized chips were gently washed with PBS twice. Through the above program, the functional chip is ready, and we can store it in a 4 ℃ refrigerator for future use. The 50 µL solutions of IgG at different concentrations were added to a 96-well plate to detect the association process for 15 min using WeSPR 200 (Xlement, China). Then, buffer solutions were added to detect the dissociation process for 30 min. Non-specific absorption experiment: Control groups of 8 µg/mL and 0 µg/mL are previously incubated with 40 µg/mL Protein A in the sensor chip surface for 12 h 60 µg/mL BSA blocking solution last for 30 min h at 37 ℃, and the experiment groups don’t incubate Protein A, but still blocked with 60 µg/mL BSA solution last 30 min h at 37 ℃. During the experiment, groups of 8 µg/mL are added 50 ul 8 µg/mL IgG and 0 µg/mL group is added buffer solution. Regeneration experiment: the functionalization process of the sensor chip is consistent with the Protein A binding experiment mentioned earlier. Next, use WeSPR One Auto™ to measure the association and dissociation processes. WeSPR One Auto™ is a microfluidic instrument that can easily switch solutions. Wash the chip with a pH 2.0 glycine solution for 10 s, followed by rinsing with a buffer solution for 20 s for regeneration. Antibodies drugs affinity measurement experiment: The functionalization process of the sensor chip is consistent with the Protein A binding experiment mentioned earlier. However, we incubated 40 µg/mL antibodies drugs of antigens instead of Protein A. Then three different antibodies drugs with different concentrations were added to measure the association and buffer solutions were added to detect the dissociation process. Small molecule (Quercetin) affinity measurement experiment: Wash the sensor once in 75 % ethanol, then rinse twice with ultrapure water, and dry in nitrogen (N2). Incubate the chip in 1 mM MUA solution (ethanol) at room temperature for 12 h. Then wash the chip twice in 75 % ethanol and ultrapure water, and incubate for 30 min at room temperature in a 1:1 mixture of 400 mM EDC and 100 mM NHS (diluted in ultrapure water). Then, Quercetin was spotted in the middle of the chip in the form of DMSO solution. After vacuum drying, the solid Quercetin small molecules were fixed on the surface of the chip under 365 nm ultraviolet light irradiation. Finally, the chip was sealed with fish gelatin solution and dried with nitrogen gas for later use. During the experiment, different concentrations of hTGFBR1 solution were added to the detection well to detect the binding process, and then buffer solution was added to detect the dissociation process. CRP detection experiment: wash once in 75 % ethanol, then rinse twice with ultrapure water, and dry in N2 Incubate the chip in 1 mmM MUA solution (ethanol) at room temperature for 12 h. Then wash the chip twice in 75 % ethanol and ultrapure water, and incubate for 30 min at room temperature in a 1:1 mixture of 400 mM EDC and 100 mM NHS (diluted in ultrapure water). Then, incubate the solution with 40 µg/mL monoclonal anti-CRP antibody at 4 ℃ for 12 h. After washing the device with PBS, incubate the chip in a 60 µg/mL BSA blocking solution and a 10% ethanolamine solution at room temperature for 30 min. Then, clean the chip with PBS. Incubate the chip with different concentrations of CRP (in PBS) for 3 min. After incubation, the OD spectrum was detected using a Biotek Epoch Microplate spectrophotometer after incubating in 50 µg/mL anti-CRP antibody for 5 min. The LOD calculation of the IDMM-SPR platform LOD is the lowest analyte concentration that can be detected but not necessarily quantitated as an exact value. For direct assays, the LOD is primarily a function of the signal-to-noise ratio in the measurement itself and is set in relation to statistical variations in response values for blank samples. A commonly used xLOD is calculated as:6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:x\mathrm{L}\mathrm{O}\mathrm{D}\:=\:{f}^{-1}({y}_{\mathrm{b}\mathrm{l}\mathrm{a}\mathrm{n}\mathrm{k}}\:+\:3{{\upsigma\:}}_{\mathrm{b}\mathrm{l}\mathrm{a}\mathrm{n}\mathrm{k}})$$\end{document} where the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:3{{\upsigma\:}}_{\mathrm{b}\mathrm{l}\mathrm{a}\mathrm{n}\mathrm{k}}$$\end{document} is the standard deviation of the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{y}_{\mathrm{b}\mathrm{l}\mathrm{a}\mathrm{n}\mathrm{k}}$$\end{document} (if the deviation of the control group is zero, use the first group) and the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{y}_{\mathrm{b}\mathrm{l}\mathrm{a}\mathrm{n}\mathrm{k}}$$\end{document} is the mean value of control signals. And \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{f}^{-1}$$\end{document} is the inverse of the fitting function. Optical settings Using an Olympus IX73 vertical fluorescence microscope to acquire transmitted bright-field microscopic color images. A 100 W halogen lamp was used as the light source, a condenser with an NA of 0.3, and an objective with an NA of 0.3 at 10× magnification. The exposure time was set to 25 ms, and the digital gain was 1. Images were acquired using a charge-coupled device (CCD) camera controlled by cellSens software. Results Principle and materials selection for metasurface plasma excitation Surface plasmon (SP) is a charge density wave that exists at the interface between a metal and a dielectric. SPP is collective oscillation of SP, which is induced by an electromagnetic wave. SPR are the collective oscillations of free electrons at the interface between a metal and dielectric material [29]. In our system, the influence of SPP can written as:7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\lambda\:}_{sp}=\frac{P}{\sqrt{{i}^{2}+{j}^{2}}}\sqrt{\frac{{\varepsilon}_{m}{\varepsilon}_{d}}{{\varepsilon}_{m}+{\varepsilon}_{d}}}$$\end{document} Where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\lambda\:}_{sp}$$\end{document} is the resonance wavelength of SPP, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:P$$\end{document} is the periodicity of nanostructures, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\varepsilon}_{m}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\varepsilon}_{d}$$\end{document} are the dielectric constants of the metal film and the medium (at the top or bottom surface of the nanostructures), And \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:i$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:j$$\end{document} are integers. Unlike conventional SPR sensors, the periodic gold nanostructure exhibits complex resonant behaviors due to the combined effects of LSPR within the holes and Wood’s anomalies on the outer surface [30]. Wood’s anomalies can expresse as:8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\lambda\:}_{w}=\frac{P}{\sqrt{{i}^{2}+{j}^{2}}}\sqrt{{\varepsilon}_{d}}$$\end{document} Both SPP and Wood’s anomalies can exist in the same problem and, depending on system details, lie close to one another. We have further analyzed and introduced the content of SPP and Wood’s anomalies in Supplementary. From Eqs. (7) and (8), it is evident that the sensor exhibits high sensitivity to the dielectric constants of the surrounding medium. When an analyte binds to the sensor surface, the resultant change in the surface refractive index induces a shift in the resonance spectrum, thereby enabling the monitoring of molecular interactions. However, the detailed detection of such spectral shifts typically necessitates high-resolution spectrometers, which can limit the practicality and accessibility of the sensing platform. To reduce reliance on complex instruments and simplify spectral shift detection, we used the differential OD spectral response (see Supplementary Fig. S2). Specifically, we utilize the endpoint method to quantify these shifts, which simplifies the detection process and enhances the robustness of our sensor design. This approach allows for the effective monitoring of molecular interactions without compromising the sensitivity or resolution of the measurements. The efficacy of SPR is largely contingent upon the properties of the metal film employed. To optimally harness the plasmonic effects on metasurfaces, the selection of appropriate coating materials is imperative. Silver and gold are the preferred plasmonic materials due to their low optical losses in the visible and near-infrared (NIR) spectra, with silver exhibiting superior performance in the visible range [31]. However, both silver and gold suffer from poor adhesion, necessitating the use of a chromium adhesive layer to ensure stable biosensor fabrication. As shown in Supplementary Fig. S3, we conducted simulation calculations on several common metals, which corroborated the aforementioned considerations. In summary, we have elected to utilize chromium as the adhesive layer, silver for its superior surface plasmon polariton (SPP) performance, and gold as the protective layer to ensure stability and biocompatibility. Inverse design of multilayer metasurface SPR sensors Our initial intention was to develop a metasurface that possesses both extremely high OD response and resonance peak shift response, enabling it to be highly compatible with commercial microplate absorbance readers or cameras, thereby achieving superior sensing performance. This approach addresses significant challenges in current molecular interaction sensors, such as insufficient sensitivity, poor stability, and high cost. As previously discussed, the combination of multilayer materials and periodic nanocup arrays offers distinct advantages. This geometric design allows the transformation of the SPR displacement response into an OD response. This method effectively reduces device dependency and enhances the resolution of RI measurements. We designed the nanocup metasurface sensor based on multilayer nanocup arrays, combining FDTD simulations with an inverse design optimization algorithm. To simultaneously optimize the both metrics, we introduced a double-objective function to serve as the optimization criterion. The optimization function can be written as:9\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:FoM=w*\varDelta\:difOD\:+\:\:v*\:\varDelta\:\:Peak,\:({OD}_{max}<3.8)$$\end{document}10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned}FoM & =w*\varDelta\:difOD\:+\:\:v*\:\varDelta\:\:Peak\\ & \quad -{(20*10}^{{10*(OD}_{max}-3.8)}-1),(3.8<{OD}_{max}) \end{aligned}$$\end{document} Here, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\varDelta\:difOD\:$$\end{document}represents the OD response, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\:\varDelta\:\:Peak\:$$\end{document} represents the resonance peak shift response. The weights \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:w$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:v$$\end{document} were set to 3 and 1, respectively, based on preliminary calculations (Supplementary Fig. S1) that showed this weighting brings the two objectives to comparable magnitudes, ensuring the final design excels in both ΔdifOD and ΔPeak. The segmented function with OD = 3.8 as the threshold is based on our observation that the signal-to-noise ratio (SNR) in commercial spectrometers significantly declines when OD exceeds 3.8. This decline directly impacts the reliability and accuracy of measurements. As discussed in Sect. "Definition of sensitivity and FoM of inverse design metasurface", balancing the thickness and SNR of metasurface sensors is crucial for optimizing sensing performance. When OD exceeds 3.8, the SNR deteriorates, necessitating the introduction of a penalty term to counteract this adverse effect. This penalty term, derived from empirical testing, indicates that exceeding a certain OD threshold compromises sensor performance. More precise definition of this penalty term can be explored in future work. Notably, this threshold should be instrument-specific—more advanced detection systems can extend the upper OD limit, enabling sensors with even higher performance. Designing metasurface sensors with multilayer plasmonic materials requires consideration of both the geometric parameters of the metasurface and the thicknesses of the multilayer materials, significantly increasing the dimensionality of the parameter space. For instance, evaluating ten values for each parameter—such as period, radius, height, and thicknesses of three plasmonic layers—results in \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{10}^{6}$$\end{document} possible combinations. This complexity demands extensive parameter sweeps and specialized optimization techniques, which are both time-consuming and reliant on expertise in metasurface engineering. To efficiently identify the optimal parameter combination that maximizes the FoM within this high-dimensional space, we employed a Particle Swarm Optimization (PSO) algorithm coupled with FDTD simulations (Fig. 3a). This approach utilizes a population-based optimization strategy, where a swarm of particles collectively explores the parameter space and iteratively improves the optimization objective through collaborative movement, ultimately maximizing the FoM. Fig. 2Inverse design of multilayer MetaSPR Sensor Chip. (a) Schematic of the working principle and the associated FoM of our multilayer MetaSPR sensor. (b) Schematic illustrations of the five parameters that define the architecture of the multilayered nano-cup arrays and their respective ranges utilized in the PSO algorithm. Within this five-dimensional parameter space, 10 initial populations are randomly generated and subsequently evolved iteratively via the PSO algorithm to identify the sensor configuration that achieves the highest FoM (Here, “Best” refers to the historical best value, and “average” refers to the mean value excluding zero values.). (c) The optimal results of each parameter for the multilayer MetaSPR Sensor during each generation. (d) Comparison of experimental and simulation results under different thicknesses and periods (red line represents experimental results, blue line represents simulation results) Inverse design of multilayer MetaSPR Sensor Chip. (a) Schematic of the working principle and the associated FoM of our multilayer MetaSPR sensor. (b) Schematic illustrations of the five parameters that define the architecture of the multilayered nano-cup arrays and their respective ranges utilized in the PSO algorithm. Within this five-dimensional parameter space, 10 initial populations are randomly generated and subsequently evolved iteratively via the PSO algorithm to identify the sensor configuration that achieves the highest FoM (Here, “Best” refers to the historical best value, and “average” refers to the mean value excluding zero values.). (c) The optimal results of each parameter for the multilayer MetaSPR Sensor during each generation. (d) Comparison of experimental and simulation results under different thicknesses and periods (red line represents experimental results, blue line represents simulation results) We initialize the optimization process with 12 particles, each assigned a random set of initial parameter values. The corresponding FoM for each particle is evaluated using FDTD calculations. In subsequent generations, each particle updates its position based on its own experience and the collective experience of the swarm, moving towards parameter values that yield higher FoM values (Supplementary Fig. S4). As shown in Fig. 2b, after iterating through 15 generations, the average FoM of the swarm increases from 10.3 to 12.5, with the best-performing particle achieving an FoM of 14.1. During the optimization process, some particles exhibited an FoM of 0. This occurred when the penalty was so large that it rendered the FoM negative, in which case we set the FoM to 0. Additionally, we assigned an FoM of 0 to particles that produced abnormal results during the calculations, in order to prevent erroneous results from affecting the overall optimization process. The optimized sensor architecture obtained through this process is characterized by the following parameters: p = 475 nm, h = 400 nm, r = 140 nm, Cr = 5 nm, Ag = 90 nm, Au = 10 nm. The optimal results of each parameter in each generation are shown in Fig. 2c. For detailed results, please refer to Supplementary Figs. S5-S9. Additionally, we further investigate and validate the results of the PSO inverse design through a parametric sweep study. As shown in Supplementary Fig. S10 and S11, our proposed double-objective function inverse design approach effectively yields optimal design results that possess both high OD response and peak shift response. Moreover, this approach also enables us to assess whether the desired design outcomes can still be achieved within a certain range of manufacturing tolerances. At this point, the reverse engineered sensor can excite the SPR effect using only vertical light incidence (Supplementary Fig. S12), eliminating the need for precision equipment such as prism coupling and angle measurement. To further verify the conclusions of the above numerical calculation, we manufactured chips with different periods and thicknesses at a diameter-to-period ratio of 0.60. Figure 2d shows the variation of OD response spectra of IDMM-SPR sensors under different period and thickness parameters at 4% sucrose concentration. The simulation and experimental results are highly consistent, confirming the accuracy of our design. Notably, the OD response increases with the period, and the resonance peak redshifts, indicating a strong correlation between structural parameters and optical performance. Under a fixed period, the OD response also increases with the increase of Ag layer thickness. These findings provide a basis for tuning the OD response and position in metasurface research. Mechanism analysis based on LSPR, wood anomalies, and SPR resonance enhancement Although inverse design has provided us with a nanocup array metasurface that exhibits optimal performance, the underlying enhancement mechanisms remain unclear. To elucidate the physical significance behind these parameters, further spectral and electric field analyses are required. We calculated and plotted the transmission spectra of the optimal parameters in different refractive index media, as shown in Fig. 3a. We observed that the spectral shifts in different refractive index media primarily occur in the 500–750 nm wavelength range. Within this range, two distinct transmission maxima and minima are evident, which we labeled in wavelength order as λ1, λ2, λ3, and λ4. To further identify the physical mechanisms of these resonance peaks, we plotted the electric field distributions of the four peaks, as shown in Fig. 3b. We observed complex field distributions at the shorter wavelength λ1, with enhanced electric fields around the nanocups and maxima located between the nanocups. This phenomenon is similar to the findings of F. A. A. Nugroho et al. [22], suggesting a significant contribution from Rayleigh anomalies (RA) to the resonance. At λ2, the electric field distribution is similar to that at λ1, but the field enhancement is directed inward toward the nanocups, indicative of a clear LSPR phenomenon. At λ3, we observed magnetic field stripes similar to those reported by Yanik et al. [32] in the context of Wood’s anomalies, and this position is also close to that described in Equation S6. At λ4, the electric field exhibits a pronounced enhancement, corresponding to the extraordinary optical transmission (EOT) phenomenon induced by the nanocup array metasurface BW-SPP, which aligns with the description in Equation S4. Through electric field analysis, we can confirm the presence of multiple complex optical effects within the nanocup array metasurface. As shown in Fig. 3c, the period of the nanocup array significantly influences the sensing response. We extracted the response values using the method described in S2 and plotted them against the period, as shown in Fig. 3d, revealing a clear positive correlation between the sensing response and the period. To further understand the physical mechanisms behind this phenomenon, we plotted the cross-sectional electric field distributions for LSPR, Wood’s anomalies, and BW-SPP, which are the three dominant contributors to the metasurface sensing response, and calculated the proportion of the electric field strength greater than 2.5 in the enhanced region relative to the total area. As shown in Figs. 3e-j, the electric field responses of LSPR, Wood’s anomalies, and BW-SPP are all highly positively correlated with the period. This result is also supported by the derivations in Equations S4 and S6, where both Wood’s anomalies and BW-SPP are positively correlated with the period. As the period increases, the resonance wavelength is coupled to longer wavelengths, and the resonance at longer wavelengths causes the electric field to couple to regions farther from the interface, thereby increasing the sensing area and enhancing the sensing response. This result is consistent with Fig. 3d. Notably, as shown in Figs. 3e and i, LSPR and BW-SPP not only have larger sensing areas with increasing period but also exhibit stronger local electric field enhancements, with enhanced electric fields inside the nanocups. In contrast, as shown in Fig. 3g, the intensity of Wood’s anomalies remains relatively unchanged, and this change primarily occurs above the metasurface. Among these three effects, the electric field of BW-SPP dominates. It is these three effects that extend the electric field of light interacting with the metasurface beyond the interface, making it highly sensitive to the refractive index near the interface. Furthermore, we analyzed the electric field at the difOD maximum, as shown in Fig. 3k, and found distinct features of LSPR and Wood’s anomalies at this wavelength, with significantly enhanced electric fields. Therefore, this wavelength is likely a region where LSPR, Wood’s anomalies, and BW-SPP are coupled, resulting in the optimal sensing performance. Fig. 3Mechanism analysis of resonance enhancement in nanocup array metasurfaces. (a) Transmission spectra of the nanocup array metasurface (P = 475 nm, R = 135 nm, H400, Cr = 5 nm, Ag = 80 nm, Au = 10 nm). (b) Top-view electric field distributions for RA, LSPR, Wood’s anomalies, and BW-SPP. (c) difOD response curves for different periods (P = 400 nm, 425 nm, 450 nm, 475 nm). (d) Bar chart of the maximum difOD response against period. (e) Cross-sectional electric field distributions for LSPR at different periods (P = 400 nm, 425 nm, 450 nm, 475 nm). (f) Bar chart of the proportion of the electric field area greater than 2.5 relative to the total area for LSPR at different periods (P = 400 nm, 425 nm, 450 nm, 475 nm). (g) Cross-sectional electric field distributions for Wood’s anomalies at different periods (P = 400 nm, 425 nm, 450 nm, 475 nm). (h) Bar chart of the proportion of the electric field area greater than 2.5 relative to the total area for Wood’s anomalies at different periods (P = 400 nm, 425 nm, 450 nm, 475 nm). (i) Cross-sectional electric field distributions for BW-SPP at different periods (P = 400 nm, 425 nm, 450 nm, 475 nm). (j) Bar chart of the proportion of the electric field area greater than 2.5 relative to the total area for BW-SPP at different periods (P = 400 nm, 425 nm, 450 nm, 475 nm). (k) Cross-sectional electric field distributions at the difOD maximum for different periods (P = 400 nm, 425 nm, 450 nm, 475 nm) Mechanism analysis of resonance enhancement in nanocup array metasurfaces. (a) Transmission spectra of the nanocup array metasurface (P = 475 nm, R = 135 nm, H400, Cr = 5 nm, Ag = 80 nm, Au = 10 nm). (b) Top-view electric field distributions for RA, LSPR, Wood’s anomalies, and BW-SPP. (c) difOD response curves for different periods (P = 400 nm, 425 nm, 450 nm, 475 nm). (d) Bar chart of the maximum difOD response against period. (e) Cross-sectional electric field distributions for LSPR at different periods (P = 400 nm, 425 nm, 450 nm, 475 nm). (f) Bar chart of the proportion of the electric field area greater than 2.5 relative to the total area for LSPR at different periods (P = 400 nm, 425 nm, 450 nm, 475 nm). (g) Cross-sectional electric field distributions for Wood’s anomalies at different periods (P = 400 nm, 425 nm, 450 nm, 475 nm). (h) Bar chart of the proportion of the electric field area greater than 2.5 relative to the total area for Wood’s anomalies at different periods (P = 400 nm, 425 nm, 450 nm, 475 nm). (i) Cross-sectional electric field distributions for BW-SPP at different periods (P = 400 nm, 425 nm, 450 nm, 475 nm). (j) Bar chart of the proportion of the electric field area greater than 2.5 relative to the total area for BW-SPP at different periods (P = 400 nm, 425 nm, 450 nm, 475 nm). (k) Cross-sectional electric field distributions at the difOD maximum for different periods (P = 400 nm, 425 nm, 450 nm, 475 nm) Characterization and validation of IDMM-SPR sensors Following the completion of chip fabrication (the main process is shown in Supplementary Fig. S13), we employed scanning electron microscopy (SEM) to examine the morphology of the IDMM-SPR chip and verify that its parameters align with our design specifications. Figure 4a presents SEM images of the chip surface at different periods, revealing a uniform arrangement of 3D nanocup structures on the metasurface. The images are taken at the same magnification, and it is evident that the nanocup period increases progressively from left to right. Figure 4b illustrates the cross-sectional view of the nanocup metasurface arrays, confirming that nanocups with a height of 450 nm are uniformly arranged. The SEM results indicate that the actual radius of the nanocup metasurface is slightly smaller than the designed radius. This discrepancy is primarily attributed to manufacturing and measurement errors, as well as the deposition of plasmonic materials on the surface. The depth discrepancy may arise from the nanoimprint lithography process. However, as discussed in Sect. "Software", our multilayer nanocup design allows for robust performance even within a certain range of fabrication errors, ensuring the high performance of the metasurface sensor.These results demonstrate that nano-etching and nanoimprinting processes can effectively produce metasurfaces that meet our design requirements. We then assembled the fabricated IDMM-SPR sensor with a 3D-printed 96-well plate to create a portable detection device. By placing sucrose solutions of varying concentrations into the wells and measuring their absorbance, we obtained differential OD spectra by subtracting the absorbance curves from that of the control group (water). Figure 4c shows typical differential OD responses of the four chips at different refractive indices (RIs), while Fig. 4d presents a strong correlation between OD response and RI (R² >0.99) for the four types of IDMM-SPR chips. These findings demonstrate that the designed IDMM-SPR biosensors exhibit exceptional capabilities for RI sensing and hold significant potential for applications in biological detection. Additionally, the results indicate that larger periods correspond to higher sensor responses and improved detection resolution. This conclusion is further supported by numerical simulations. This conclusion is further supported by numerical simulations. Supplementary Fig. S14a and b show that significant local electric field enhancement only occurs at the resonant wavelength. Moreover, the amplitude of this enhancement increases significantly with the period. As shown in Supplementary Fig. S14c, magnetic field stripes similar to those reported by Yanik et al. [32] in the context of Wood’s anomalies are observed. According to Supplementary Fig. S14c and d, under our optimal structure with a period of 475 nm, both Wood’s anomalies and bound-wave surface plasmon polaritons (BW-SPP) contribute to the sensing capability, with the SPP mode being predominant. As shown in Supplementary Fig. S15, we further evaluated the quantitative detection sensitivity of the above parameter chips using the differential OD response of Protein A against immunoglobulin G (IgG) antibodies, and the results show a similar trend with the sucrose experiment. Meanwhile, as shown in Supplementary Fig. S16, we further calculated the noise and signal-to-noise ratio of different parameter metasurfaces to evaluate our parameter selection. Finally, we confirmed that the optimal thickness combination is 5 nm for Cr, 80 nm for Ag, and 10 nm for Au. According to the results of the simulation and experiment, when the thickness of Ag is greater than 100 nm, the differential OD response no longer increases significantly. On the contrary, the signal-to-noise ratio decreases due to the performance limitations of the instrument. And increasing the thickness of the metal will also increase manufacturing costs. Figure 4e displays transmission microscopy images of four IDMM-SPR chips in the air, exhibiting distinct color changes that correspond to variations in resonance wavelengths induced by periodic structural differences. These color variations demonstrate the chips’ potential for wavelength-selective filtering. Given the differing periods of the four fabricated parameters, their resonance wavelengths are expected to vary, as predicted by Ebbesen’s formula [30]. To more intuitively illustrate the influence of period and thickness on transmitted light intensity within the designed metasurface, we present transmission microscopy images for all metasurfaces with periods ranging from 400 to 475 nm and Ag thicknesses from 60 to 140 nm in Supplementary Fig. S17. As shown in the Supplementary Fig. S17, the transmission images progressively darken with increasing thickness while the colors shift across different periods. The unprocessed OD spectrum of the sensors in Supplementary Fig. S18 also confirms this. Owing to our double-objective function design, which takes into account the resonance peak shift response, and given that peak shifts induce color changes, the inversely designed sensor is also highly suitable for colorimetric sensing in SPR imaging. We demonstrated the feasibility of using it for optical imaging colorimetric sensing in Supplementary Fig. S19. As shown in Supplementary Fig. S19a, the sensor surface exhibits distinct color changes under different sucrose concentrations. We further separated the RGB channel image data and found that the R channel plays a dominant role in this change, as the strong resonance and peak shift mainly occur in the red band of 620–720 nm (Supplementary Fig. S19b). Moreover, we extracted the mean values of the RGB channels and fitted them against sucrose concentration, revealing a high correlation between RGB channel data and sucrose concentration, with an R2 value exceeding 0.9. These findings demonstrate that the sensor developed using the double-objective inverse design method is not only compatible with commercial microplate readers but also suitable for camera-based sensors, enabling its application in a wider range of scenarios. However, considering that microplate absorbance readers have well-established commercial solutions that are more conducive to controlled conditions and rapid dissemination of our practical application value, this study focuses on showcasing the potential research scenarios and value based the perspective of OD response. To evaluate and demonstrate the performance of the developed IDMM-SPR sensor, we also compared it with relevant reported works in Supplementary Table S1. Although FoM and sensitivity were not our optimization goals, the inverse design optimal sensor (p = 475 nm, h = 480 nm, r = 140 nm, Cr = 5 nm, Ag = 80 nm, Au = 10 nm.) we ultimately determined still achieved 812 nm ⁄ RIU (Refractive Index Unit) and 26.3 RIU− 1, respectively. Fig. 4Characterization and verification of IDMM-SPR chips. (a) SEM images of structures P1 - P4 with varying periods. (b) Cross-sectional SEM image of the IDMM-SPR chip. (c) Differential OD response spectra of structures P1 - P4. (d) Differential OD response spectra of structures P1 - P4 at different sucrose concentrations. (e) Response curves of varying sucrose concentrations for structures P1 - P4 (differential OD response peak minus valley). (f) Transmission microscopy images of structures P1 - P4 in air Characterization and verification of IDMM-SPR chips. (a) SEM images of structures P1 - P4 with varying periods. (b) Cross-sectional SEM image of the IDMM-SPR chip. (c) Differential OD response spectra of structures P1 - P4. (d) Differential OD response spectra of structures P1 - P4 at different sucrose concentrations. (e) Response curves of varying sucrose concentrations for structures P1 - P4 (differential OD response peak minus valley). (f) Transmission microscopy images of structures P1 - P4 in air Application of IDMM-SPR sensor in ultra-high affinity antibody drugs development Therapeutic antibodies have grown rapidly in recent years, becoming a dominant force in the therapeutics market [33]. The IDMM-SPR biosensing platform developed by our research institute provides real-time affinity data to evaluate the binding and dissociation rates between antibody components and target molecules, which can help understand drug mechanisms, screen lead compounds, optimize drug design, target fishing and validation, and accelerate the discovery process of new drugs. The measurement of ultra-high affinity interactions typically involves rapid binding and slow dissociation processes, necessitating real-time detection capabilities and high sensor stability. Insufficient sensor stability and sensitivity can lead to baseline drift, thereby compromising data accuracy. Therefore, strategies to control this issue are required [34]. To address this, we further evaluated the anti-nonspecific adsorption capability of the IDMM-SPR biosensor (Supplementary Fig. S20). The results demonstrated that the sensor exhibits excellent anti-nonspecific adsorption performance when configured with a 10 nm Au overlayer. As shown in Supplementary Figs. S21 – S23, we optimized the blocking and detection systems for the sensing platform. A 1% BSA solution was selected as the blocking agent, while CBS and HBS-ET were used as the incubation and detection buffer systems, respectively. Additionally, Supplementary Fig. S24 presents the raw OD spectra and calculated differential OD data for each critical step in biological sample detection, confirming that the data accurately reflect the adsorption and desorption of biomolecules on the sensor surface. Finally, as shown in Supplementary Fig. S25, we detail the establishment of an affinity analysis platform based on IDMM-SPR sensors. Furthermore, in Supplementary Fig. S26, we validate the anti-non-specific adsorption properties, regeneration ability, and stability of the IDMM-SPR sensors. Notably, over nearly 60 regeneration experiments, the sensors demonstrated no significant signal attenuation, thereby confirming their robust regeneration capability and stability. These experiments collectively demonstrate the feasibility, accuracy, and stability of the affinity detection platform developed based on the IDMM-SPR sensor. To further validate the applicability of our IDMM-SPR biosensor for ultra-high affinity molecular interaction analysis, we selected Programmed Death-1 (PD-1) and its ligand Sintilimab as a model system. Sintilimab is a safe and effective human IgG4 monoclonal antibody that binds to PD-1 with a \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\boldsymbol{K}}_{\boldsymbol{D}}$$\end{document} value of 74 pM [35]. Given its ultra-high affinity, a prolonged dissociation phase is required to achieve sufficient dissociation for accurate data fitting and determination of the dissociation constant (typically observing a 5% decrease). Insufficient sensor sensitivity and stability may lead to drift in the dissociation curve during prolonged dissociation periods. This drift can cause the curve, which should ideally decline, to appear as a horizontal line or even show a slow increase, mimicking non-dissociation or spurious re-binding phenomena. These issues complicate data fitting and prevent accurate calculation of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\boldsymbol{K}}_{\boldsymbol{D}}$$\end{document}. Figure 5a illustrates the rapid binding kinetics between PD-1 and Sintilimab, with a swift increase in the binding curve (data normalized to a zero-concentration reference). Upon introducing buffer solution, the dissociation phase is notably slow, prompting a 30-minute dissociation period for adequate data collection. Figure 5b demonstrates the IDMM-SPR platform’s ability to stably monitor this slow dissociation process, characteristic of ultra-high-affinity interactions. Using the Langmuir model with 1:1 binding stoichiometry, we fitted the kinetic data (Fig. 5c) and obtained a \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\boldsymbol{K}}_{\boldsymbol{D}}$$\end{document} value of 70–127 pM, closely aligning with previously reported values [35, 36]. This result underscores the exceptional performance of our SPR sensor in ultra-high-affinity kinetic analysis and target validation. Due to cost control rather than the limitations of the testing equipment we developed. Here, a monoclonal antibody drug (Adalimumab), a potential Antibody–Drug Conjugate (ADC) drug candidate (Polatuzumab) and an ADC drug (Sacituzumab govitecan) are used to demonstrate our simple, rapid, and high-throughput molecular fishing and affinity detection scheme based on the IDMM-SPR biosensor. As shown in Fig. 4d, we developed a high-throughput affinity detection protocol based on the IDMM-SPR biosensor. This protocol includes sensor surface cleaning, functionalization, binding detection, and dissociation detection. Figure 4e illustrates a schematic of sample deposition on the sensor surface during the detection process. The entire measurement process can be completed in 20 min or less, enabling high-throughput detection. To further validate the performance of the inverse-designed sensor, we compared it with an unoptimized sensor (parameters: p = 400 nm, h = 550 nm, r = 90 nm, Cr = 5 nm, Ag = 60 nm, Au = 5 nm). It is worth noting that the non-optimized sensor was fabricated near the optimal parameters to confirm our inverse-designed results, rather than being deliberately selected as the worst-case scenario. Despite this, the performance of the non-optimized sensor remains highly advanced compared to many previously reported works. Figures 5f–h show the detection kinetic curves obtained using the unoptimized sensor, while Figs. 5i–k present the corresponding curves for the optimized sensor. Although the unoptimized sensor performed well, the optimized sensor demonstrated significantly enhanced detection capabilities, particularly in distinguishing low-concentration samples. In Supplementary Fig. S27, we quantitatively fit the binding signals, further highlighting the difference between the two sensors. Furthermore, we calculated the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\boldsymbol{K}}_{\boldsymbol{D}}$$\end{document} values of the three antibody drugs using both sensors and summarized the results in Table 1. The \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{K}_{D}$$\end{document} values obtained from the unoptimized sensor were nearly one order of magnitude lower than those from the optimized sensor. This discrepancy is primarily attributed to the insufficient precision of the unoptimized sensor, which fails to accurately capture the kinetic binding and dissociation processes, especially at low analyte concentrations. A secondary contributing factor is the relatively thin surface Au layer of the unoptimized sensor, which reduces anti-nonspecific adsorption capability and compromises sensor stability; both factors introduce bias into the calculated \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\boldsymbol{K}}_{\boldsymbol{D}}$$\end{document} values. In contrast, the optimized sensor’s results are consistent with previously reported values, confirming its superior real-time affinity detection capabilities. The results from Supplementary Fig. S28 and Fig.S29 further demonstrate the platform’s potential as a molecular-fishing tool for small molecule affinity evaluation and ligand screening, underscoring the IDMM-SPR sensor’s capability in ultra-high affinity kinetic analysis and its promise for advancing the development of next-generation antibody drugs. Fig. 5High-throughput affinity sensing platform for ADC drug development. (a) The association process of incubating PD-1 on the chip to detect Sintilimab. (b) The long-time dissociation process of PD-1 with Sintilimab. (c) Dynamic real-time and fitting curves of the association and dissociation phases of the interaction between PD-1 and Sintilimab. (d) Schematic diagram of the high-throughput antibody affinity detection protocol, with each group, repeated three times in the three samples tested, and the concentration of A-H in each column decreasing incrementally. (e) Schematic diagram of the main events in the antibody drug fishing and affinity detection process. (f-h) Dynamic real-time and fitting curves of the association and dissociation phases for Polatuzumab with CD79B, Adalimumab with TNF-α, and Sacituzumab govitecan with TROP2, detected using the unoptimized sensor. (i-k) Dynamic real-time and fitting curves of the association and dissociation phases for Polatuzumab with CD79B, Adalimumab with TNF-α, and Sacituzumab govitecan with TROP2, detected using the optimized IDMM-SPR sensor High-throughput affinity sensing platform for ADC drug development. (a) The association process of incubating PD-1 on the chip to detect Sintilimab. (b) The long-time dissociation process of PD-1 with Sintilimab. (c) Dynamic real-time and fitting curves of the association and dissociation phases of the interaction between PD-1 and Sintilimab. (d) Schematic diagram of the high-throughput antibody affinity detection protocol, with each group, repeated three times in the three samples tested, and the concentration of A-H in each column decreasing incrementally. (e) Schematic diagram of the main events in the antibody drug fishing and affinity detection process. (f-h) Dynamic real-time and fitting curves of the association and dissociation phases for Polatuzumab with CD79B, Adalimumab with TNF-α, and Sacituzumab govitecan with TROP2, detected using the unoptimized sensor. (i-k) Dynamic real-time and fitting curves of the association and dissociation phases for Polatuzumab with CD79B, Adalimumab with TNF-α, and Sacituzumab govitecan with TROP2, detected using the optimized IDMM-SPR sensor Table 1Results and comparison of ADCs affinity measurementNO.Name\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\boldsymbol{K}}_{\boldsymbol{D}}$$\end{document}-unopt (10−9 M)\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\boldsymbol{K}}_{\boldsymbol{D}}$$\end{document}-opt (10−9 M)Ref_\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\boldsymbol{K}}_{\boldsymbol{D}}$$\end{document} (10−9 M)Ref.1Polatuzumab0.321.181.33[37]2Adalimumab0.040.190.127[38]3Sacituzumab govitecan0.070.480.26[39] Table 1Results and comparison of ADCs affinity measurementNO.Name\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\boldsymbol{K}}_{\boldsymbol{D}}$$\end{document}-unopt (10−9 M)\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\boldsymbol{K}}_{\boldsymbol{D}}$$\end{document}-opt (10−9 M)Ref_\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\boldsymbol{K}}_{\boldsymbol{D}}$$\end{document} (10−9 M)Ref.1Polatuzumab0.321.181.33[37]2Adalimumab0.040.190.127[38]3Sacituzumab govitecan0.070.480.26[39] Results and comparison of ADCs affinity measurement Development of the quantification platform with C-reactive protein (CRP) Quantitative detection of target analytes (biomarkers) holds significant importance in numerous fields, particularly in disease screening and diagnosis [40, 41]. Traditional methods for biomarker quantification, including immunoassays, enzyme-linked immunosorbent assay (ELISA), Western blot (WB), and RT-PCR, are often limited by their sensitivity and accuracy. Moreover, these approaches are typically costly, intricate, and time-consuming, requiring sizable sample volumes [41]. CRP is an inflammatory biomarker whose levels are usually low in healthy individuals but can increase by 100 to 200 times or more in acute systemic inflammation [42]. CRP and other inflammatory markers have been measured in various clinical settings, and it has been shown that CRP levels are higher in patients with acute coronary syndrome, which can be used to predict the prognosis of these patients [42]. In this study, we demonstrate a CRP detection scheme based on the designed IDMM-SPR platform, which is ultra-sensitive, has a wide response range, and is fast and low-cost. This platform can be rapidly and efficiently deployed for CRP detection applications. As shown in Fig. 6a, to maximize the performance of the sensor, we self-assembled a monolayer of MUA on the chip surface to reduce interference and improve fixation efficiency. Subsequently, the signal was further amplified using the double-antibody sandwich method reported previously [21]. To test the effectiveness of the developed quantitative analysis platform, we also conducted experiments on specific recognition and non-specific adsorption. As shown in Fig. 6b, all test groups showed no significant signal changes like CRP at 0 concentration, which reflects the excellent ability of our developed quantitative analysis platform to resist non-specific adsorption and specifically recognize CRP (Refer to Supplementary Fig. S26d for the experiment of the 8 µg/mL CRP group). The 8 µg/mL CRP group, which was the experimental group without CRP antibody incubation, showed similar results to the 0 µg/mL CRP group, both lacking significant signal response, further validating the platform’s anti-nonspecific adsorption capability. Figure 6c shows the differential OD curve after the second addition of antibodies, and the results show good discrimination at different concentrations. To verify the effectiveness of the inverse design approach, we conducted identical experiments using both an unoptimized sensor and a sensor without any metasurface design. The results are plotted together in Fig. 6d, while Fig. 6e presents the fitting results of the response values relative to concentration. Figure 6f illustrates the three sensor configurations. The LOD for the optimized and unoptimized sensors was calculated to be 0.84 ng/mL and 124.02 ng/mL, respectively. This indicates that the detection performance of the sensor was enhanced by at least 147 times after inverse design optimization, and which is 60-fold lower than before [21]. In contrast, the sensor that was merely coated without any design could hardly distinguish the concentration of any samples in this application scenario. This clearly demonstrates the necessity of rational inverse design in the complex, multidimensional design of metasurface biosensors, which can significantly improve sensor performance. Furthermore, we compared our quantitative analysis platform based on the IDMM-SPR sensor with several previously reported works. As shown in Table 2, our CRP detection scheme using the IDMM-SPR platform demonstrates significant advantages in terms of the limit of detection (LOD) compared to traditional methods. For instance, compared to the fiber optic SPR, our approach achieves a LOD that is nearly 260 times lower. Additionally, when compared to the colorimetric immunoassay based on gold nanoparticles, our method exhibits a LOD that is nearly 120 times lower. Moreover, the inverse-designed IDMM-SPR sensor shows a nearly 150-fold improvement in LOD compared to the unoptimized MetaSPR sensor. These comparisons highlight the substantial potential of inverse design in enhancing sensor performance. The multilayer IDMM-SPR biosensor exhibits considerable potential for applications in clinical detection and diverse areas of bio-detection. Characterized by high sensitivity, rapid response, and cost-effectiveness, this detection platform is poised for broad application in clinical diagnostics and bio-detection, thereby providing robust support for early disease screening and precision medicine. Its advanced capabilities are expected to enhance the accuracy and efficiency of diagnostic processes, ultimately contributing to improved healthcare outcomes and advancements in biomedical research. Fig. 6Application of IDMM-SPR Sensing Platform for CRP Quantitative Detection. (a) Schematic diagram of the CRP detection process using the double-antibody sandwich method with the IDMM-SPR sensor. (b) Schematic diagram of the CRP detection process using the double-antibody sandwich method with the IDMM-SPR sensor. (c) Differential OD spectral curve of CRP detected by the double-antibody sandwich method, with a concentration range from 0 to 8 µg/mL. (d) Bar chart of end-point method OD responses for CRP detection by the double-antibody sandwich method at different concentrations, using three types of sensors: non-metasurface (Cr = 5 nm, Ag = 60 nm, Au = 5 nm), unoptimized (parameters: p = 400 nm, h = 480 nm, r = 90 nm, Cr = 5 nm, Ag = 60 nm, Au = 5 nm), and optimized (p = 475 nm, h = 480 nm, r = 140 nm, Cr = 5 nm, Ag = 80 nm, Au = 10 nm). (e) The standard response curve for CRP detection using the three types of sensors on the quantitative sensing platform. (f) Comparison of the performance of the three different structural sensors Application of IDMM-SPR Sensing Platform for CRP Quantitative Detection. (a) Schematic diagram of the CRP detection process using the double-antibody sandwich method with the IDMM-SPR sensor. (b) Schematic diagram of the CRP detection process using the double-antibody sandwich method with the IDMM-SPR sensor. (c) Differential OD spectral curve of CRP detected by the double-antibody sandwich method, with a concentration range from 0 to 8 µg/mL. (d) Bar chart of end-point method OD responses for CRP detection by the double-antibody sandwich method at different concentrations, using three types of sensors: non-metasurface (Cr = 5 nm, Ag = 60 nm, Au = 5 nm), unoptimized (parameters: p = 400 nm, h = 480 nm, r = 90 nm, Cr = 5 nm, Ag = 60 nm, Au = 5 nm), and optimized (p = 475 nm, h = 480 nm, r = 140 nm, Cr = 5 nm, Ag = 80 nm, Au = 10 nm). (e) The standard response curve for CRP detection using the three types of sensors on the quantitative sensing platform. (f) Comparison of the performance of the three different structural sensors Table 2Comparison of different strategies of CRP testingNO.YearMethodStructureWavelength (nm)Time (min)LODRef.12019SPRnanocup400–80018050 ng/mL[21]22020colorimetric immunoassayAuNP400–8005100 ng/mL[43]32020capillary-based ELISAcapillary400–8003031 ng/mL[44]42021surface relief gratingsgratings400–80018 h0.3 mg/mL[45]52023SPRfiber optic SPR500–750150.22 µg/mL[46]62023LSPRgold nanorod250–8503025 ng/mL[47]72024IDMM-SPRnanocup400–800100.84 ng/mLthis work82024MetaSPRnanocup400–80010124.02 ng/mLthis work Table 2Comparison of different strategies of CRP testingNO.YearMethodStructureWavelength (nm)Time (min)LODRef.12019SPRnanocup400–80018050 ng/mL[21]22020colorimetric immunoassayAuNP400–8005100 ng/mL[43]32020capillary-based ELISAcapillary400–8003031 ng/mL[44]42021surface relief gratingsgratings400–80018 h0.3 mg/mL[45]52023SPRfiber optic SPR500–750150.22 µg/mL[46]62023LSPRgold nanorod250–8503025 ng/mL[47]72024IDMM-SPRnanocup400–800100.84 ng/mLthis work82024MetaSPRnanocup400–80010124.02 ng/mLthis work Comparison of different strategies of CRP testing Conclusion In this study, we have successfully developed an ultrasensitive plasmonic biosensor through inverse design, leveraging the unique optical properties of multilayer nanocup arrays. Our work not only demonstrates the effectiveness of inverse design in optimizing sensor performance but also unveils novel physical mechanisms underpinned by the synergistic effects of LSPR, Wood’s anomalies, and BW-SPP. These collective resonances significantly enhance the sensor’s sensitivity and stability, enabling it to address critical limitations in current molecular interaction analysis technologies, such as insufficient sensitivity for small molecules and ultra-high affinity interactions. A important contribution of this work lies in the mechanistic analysis of the nanocup metasurface, where we identified and characterized the contributions of LSPR, Wood’s anomalies, and BW-SPP to the sensor’s performance. The electric field analysis revealed that larger periods of the nanocup array extend the sensing area and intensify local field enhancements, further amplifying the sensor’s response. This mechanistic insight not only deepens our understanding of plasmonic metasurfaces but also provides a foundation for future designs targeting specific resonance enhancements. Another standout feature of our work is the successful application of inverse design, which combines numerical simulations with machine learning to optimize the sensor’s structural parameters. Traditional sensor development often involves iterative trial-and-error processes, which are time-consuming and costly. By employing a double-objective optimization function and the PSO algorithm, we efficiently identified the optimal configuration (e.g., p = 475 nm, Ag = 80 nm) that balances high OD response and resonance peak shift. This approach not only accelerates the design process but also ensures robust performance even within manufacturing tolerances, making the sensor highly practical for real-world applications. From a practical standpoint, our IDMM-SPR sensor offers significant advantages over existing technologies. This metasurface requires only a vertically incident light source to achieve surface plasmon excitation, eliminating the need for complex and expensive detection equipment. This scalable production method, combined with the sensor’s compatibility with commercial microplate absorbance readers, positions it as a viable solution for high-throughput molecular interaction analysis. The sensor’s ability to perform real-time, label-free detection of ultra-high affinity interactions (e.g., PD-1/Sintilimab), high sensitive (0.84 ng/mL of CRP) and small molecules (e.g., Quercetin) further highlights its versatility in drug development and disease diagnosis. Notably, this work also identifies several avenues for future exploration. First, expanding the IDMM-SPR platform to portable detection devices (e.g., smartphone-based systems) could enable point-of-care applications. Second, integrating emerging materials (e.g., graphene, van der Waals materials) or exploring advanced surface functionalization strategies may further enhance sensitivity and specificity. Third, systematic investigations into how metasurface nanostructures influence metal deposition conformality (e.g., step coverage on high-aspect-ratio sidewalls) and subsequent plasmonic mode excitation will help refine fabrication-performance relationships. In summary, this study addresses critical limitations of conventional SPR and metasurface-based sensors by combining inverse design with deep mechanistic insights. The IDMM-SPR sensor not only achieves superior performance but also exemplifies how metal-based metasurfaces can overcome the limitations of dielectric counterparts in scalability and cost. We anticipate that this work will inspire further advancements in metasurface biosensors, paving the way for their widespread use in life sciences, drug discovery, molecular fishing, and clinical diagnostics. Supplementary Information Supplementary Material 1 Supplementary Material 1