Xinyuan Bi、Li Lin、Zhou Chen、Jian Ye

Small Methods 2024, 8, 2301243

Abstract

Surface-enhanced Raman spectroscopy (SERS), well acknowledged as afingerprinting and sensitive analytical technique, has exerted highapplicational value in a broad range of fields including biomedicine,environmental protection, food safety among the others. In the endlesspursuit of ever-sensitive, robust, and comprehensive sensing and imaging,advancements keep emerging in the whole pipeline of SERS, from the designof SERS substrates and reporter molecules, synthetic route planning,instrument refinement, to data preprocessing and analysis methods. Artificialintelligence (AI), which is created to imitate and eventually exceed humanbehaviors, has exhibited its power in learning high-level representations andrecognizing complicated patterns with exceptional automaticity. Therefore,facing up with the intertwining influential factors and explosive data size, AIhas been increasingly leveraged in all the above-mentioned aspects in SERS,presenting elite efficiency in accelerating systematic optimization anddeepening understanding about the fundamental physics and spectral data,which far transcends human labors and conventional computations. In thisreview, the recent progresses in SERS are summarized through the integrationof AI, and new insights of the challenges and perspectives are provided in aimto better gear SERS toward the fast track.

Introduction and Methods

As for the way we perceive the world, techniques for sensing atthe molecular level are rapidly innovating, driven by the endless pursuit of truth. In general, two technical routes are taken to meet different needs, namely labeling and label-free strategy. The former helps to locate the specific regions and sense the molecular species of interest only by focusing on the specific signals predesigned on the sensor. While in the latter strategy, the signals are directly generated by the detected molecules, which is relevant with the chemical and/or physical states of the molecules in the sensing environment.In all areas of sensing, robustness, sensitivity, specificity, and generality are the common criteria for both strategies. Raman spectroscopy is an inelastic scattering generated by the molecules inresponse to the incident light. Since the molecules can either take up or give off the energy of the incident photons as the induced polarization, the frequency of Raman scattering can be further divided into the Stokes shift when the scattering frequency is red-shifted from the incident one and the anti-Stokes shift when the scattering is blue-shifted. Given that the vibrational energy is determined by the chemical bonds of a molecule, Raman spectroscopy is well acknowledged as afinger printing technique, by which one can infer the molecular structure and chemical/physical behaviors as well as the mixture composition in the molecular level, enabling specific sensing of a wide range of molecules even if in a complex matrix.

Substrate is the foundation for all SERS sensing, primarily determining the overall performance. From the first observation of SERS spectra, floods of papers worked on diverse compositions and structures based on a range of criteria, namely: i) Excellent plasmonic effect: generally, noble metals such as silver (Ag) and gold (Au) and their alloys have good plasmonic effect and structures of sharp tips and narrow gaps can cause high EM fields nearby, namely, hotspots, especially under the matched incident frequency. Excellent substrates such as Auconcave nanocubes, Au or Ag nanostars, Au or Ag nanourchins, Ag nanoflowers, etc. have been proven with high sensitivity owing to the rich sharp ends for strong hotspots and large surface areas for molecules to adsorb. ii) Efficient charge transfer: by delicately designing the substrates, such as Ag/Au alloys, and metal–organic framework hybrids, platinum nanocrystals, etc., the charge transfer between the substrate surface and the molecules can be aroused under the specific incident frequency to further generate CM enhancement apart from EM enhancement. iii) Other criteria include temporal stability and spatial uniformity so as to maintain robust detectability. However, though with a few manual experimental efforts for optimization, there is still no reason to believe that the proposed ones are the optimal under those certain application scenarios in case of the complex geometric and compositional configurations.

Key Results and Conclusions

In recent years, as the advancements of computer science, we have witnessed that AI dramatically revolutionized and expedited the development of the whole SERS pipeline, from the very beginning of the design and selection of proper SERS substrates and reporter molecules, the strategic planning of on-demand and efficient synthetic routes, the optimization of the instrumentation for data acquisition, and the development of data preprocessing methods, toward the applicational end with various purposes. To facilitate the widespread adoption of AI techniques in the SERS community and to encourage ongoing innovation, the most direct way involves the creation of user interfaces for facile public use, thus to collect a wider range of user demands and stimulate extensive evaluation and feedback.

To date, AI is primarily playing an assisting role in all fields including SERS, where human judgment is undoubtedly the central processor considering, for instance, the ethical concerns associated with AI deployment. In this case, processes and tasks with explicitly determined objectives by human are automated and accelerated by leveraging big data and learning-based methodologies to capture high-level features and patterns that often elude human comprehension.[340] The current progress of SERS assisted by all abovementioned AI models remains largely incremental, which is somehow affected by human’s cognitive bias away from rational objectivity and the lack of causality and explainability inherently. Consequently, there is a need to evolve from AI-assisted to AI-driven mode, which would revolutionize the identification, optimization, discovery, and evaluation across all aspects of SERS. For instance, given certain functions the generative AI models or large Raman models can generate hypothesis about SERS substrate and reporter design. Based on the hypothesis, AI models help establish computer simulation to accurately deduce observables and guide laboratory evaluations. The corresponding discoveries promote and unlock novel SERS research and further refine the AI models with uncertainty indicators. Eventually, AI predictors analyze spectral data to extract the fingerprint for specific applications. This more advanced form releases AI systems from to a greater extent human intervention, where the main purpose is to stimulate the full potential of AI holding strict accountability and responsibility rather than completely taking over human control. We anticipate that step-change improvements and unprecedented capabilities will be achieved and orient the future direction of SERS research.

Fig.1：Scheme of AI for the whole SERS pipeline.

Along the whole SERS pipeline, AI can be used in the design of SERS substrates, the reporter molecules and the synthetic routes, the optimization of instrumentations and data preprocessing methods, as well as SERS related applications. In the meanwhile, there is an urgent need for SERS database in aim to achieve further development.

Fig. 2：I for SERS substrate design. 

a) AI is capable of performing both forward and inverse tasks that links the substrate structures and their properties. The structures of the substrates can be described by explicit parameters, 2D or 3D images and material compositions. To date, multiple properties have been successfully predicted or used as the targets for on-demand substrate design, including far-field extinction, near field enhancement distribution and charge mobility. b) Forward prediction of the substrate properties based on the 2D image of the substrate structures. Inverse design of the substrate structures by c) tandem neural networks and d) cGAN. e) DeepAdjoint, proposed as a standard framework for substrate design.

Fig.3：AI for SERS reporters.

a) In case of AI-assisted molecular design, the molecular structures are required to be first translated into notations readable by computers. Generally, for the molecular structures, the 3D conformation and the 2D structure are taken into consideration, and the image of the molecules can also be directly used. Thereafter, various types of notations with different generation methods have been developed for different downstream purposes. Multiple algorithms have been utilized for the generation of new on-demand molecules and the prediction of molecular properties, that is, cross section, absorption, and emission wavelength as well as quantum yield. b) A GCN for the prediction of various molecular properties based on the molecular structures. c) A VAE and d) a latentGAN to generate new molecules with the desired properties.

Fig.4：AI for SERS synthesis.

a) The training data can be collected from literatures and from experimental trials as well. Algorithms for synthetic route optimization include GA and BO followed by the visualization of key parameter conditions. b) Text mining to collect the experimental data. Reproduced with permission. c) Microfluidic systems to generate experimental data equipped with online characterization, including i) a T-junction microfluidic device and ii) a pulse mixing microfluidic system. d) GA-assisted microfluidic synthetic route optimization. Reproduced with permission. e) Parameter visualization realized by i) classification and ii) SHAP values.

Fig. 5: AI to improve the instrumentation for spectral collection and the methods for data preprocessing

a) Main aspects of instrumentation improvement and preprocessing methods. Herein, the instrumentation has been advanced in b) transmission matrices, c) light source shape, and d) optical information storage capacity.  For spike removal, background correction and denoising, different network architectures have been explored such as e) cascade U-net, f) joint fully connected network model and U-net, g) DAE, and h) denoising cycle-consistent GAN.

Fig. 6: The promotion of SERS-related applications by AI.

a) Aspects of AI assisted SERS applications. b) Qualification ascertains the existence of certain molecular species in an unknown sample. c) Quantification is to determine the detailed relative or absolute abundance of certain target molecules. d) Molecular phenotypes can also be profiled and discriminated with the assistance of AI. e) SERS imaging with AI can realize ultra-high multiplexity in both labeling and label-free strategy with demultiplexed endmembers corresponding to certain biological components.

https://doi.org/10.1002/smtd.202301243