Reviewing the Current State of ML-Enabled Raman Spectroscopy Across Applied Fields

This video was made with the help of NotebookLM.

Over the past five years, there have been several significant strides made in machine learning-assisted Raman spectral classification. A recent review article published in Sensors highlighted these advancements.¹ This article, written by lead author Jiantao Qi of China University of Petroleum, with colleagues from this institution and Ocean Chemical Industry Research Institute Co., proposes a workflow-oriented framework to improve reproducibility, model selection, and deployment across applications from biomedicine to materials analysis. Their review addresses persistent barriers, such as weak signals, complex spectra, and high-dimensional data, that have limited Raman spectroscopy’s routine use outside specialized laboratories.¹

Machine learning (ML) and artificial intelligence (AI) are becoming more widely used in spectral classification, and the amount of literature that showcases the use of ML and AI reinforces this point.² As part of their study, the researchers compiled scientific literature from Web of Science and Google Scholar between 2020 and 2025. They searched for studies that integrated ML or deep learning models with Raman spectroscopy for classification tasks and transparent performance metrics.¹ The selected studies span biomedical diagnostics, food authentication, mineralogical classification, and plastic and microplastic identification, which are areas where the nondestructive nature of Raman spectroscopy is helpful, but data complexity has hindered automation.¹

The review article also delves into upstream choices being made that helps deep learning architectures outperform chemometrics. The paper proposes a workflow that begins with standardized acquisition and preprocessing, followed by model selection matched to data volume and variability.¹ For limited data sets, classical ML or shallow neural networks with engineered features may remain competitive and more interpretable. For larger, heterogeneous data sets, convolutional or transformer-based DL models can capture complex spectral structure, provided augmentation and cross-validation are carefully designed.¹ The authors recommended reporting minimum dataset descriptors to enable cross-study comparison.¹ They also highlight the need for external test sets to assess generalization.¹

However, in their review article, the researchers explain that there is still a concern with data scarcity and reproducibility. To solve this, the authors argue that open spectral repositories and standard operating procedures (SOPs) should be accessible to everyone that covers the basics of sample preparation.¹ Multimodal integration is another priority: coupling Raman with microfluidics or mass spectrometry could enhance phenotyping of bacteria, elucidate antibiotic resistance, and monitor single-cell dynamics. The authors also point to generative AI for spectral augmentation and biomarker discovery, while stressing that interpretability methods must advance in parallel to build user trust.¹

There are several key takeaways from this study. For one, the workflow highlighted here gives spectroscopy vendors a way to improve ML outcomes. The workflow highlights hardware stability and metadata capture are two key features that could improve ML outcomes.¹ Meanwhile, for applied laboratories, standardized reporting and validation pathways may accelerate qualification of ML-enabled Raman methods in quality control, food authentication, and environmental monitoring.¹

References

1. Liu, Y.; Wu, Y.; Wang, J. et al. Recent Advances in Raman Spectral Classification with Machine Learning. Sensors 2026, 26 (1), 341. DOI: 10.3390/s26010341
2. Acevedo, A. Raman and AI for Pathological Classification. Spectroscopy. Available at: https://www.spectroscopyonline.com/view/raman-and-ai-for-pathological-classification (accessed 2026-02-20).