Hybrid Machine Learning-Assisted Raman Spectroscopy with Generative Feature Augmentation for Pharmaceutical Identification

Quach Thi Thai Binh; Ton Nu Quynh Trang; Thang B. Phan; Vu Thi Hanh Thu; Nguyen Tuan Hung

cs.LGSep 18, 2026

Hybrid Machine Learning-Assisted Raman Spectroscopy with Generative Feature Augmentation for Pharmaceutical Identification

Authors: Quach Thi Thai Binh, Ton Nu Quynh Trang, Thang B. Phan, Vu Thi Hanh Thu, Nguyen Tuan Hung

Organizations: Faculty of Physics and Physics Engineering University of Science Ho Chi Minh City 700000 Viet Nam · Vietnam National University Ho Chi Minh City 700000 Viet Nam · Advanced Materials Technology Institute Vietnam National University Ho Chi Minh City (formerly Affiliated with Center for Innovative Materials and Architectures) Ho Chi Minh City 700000 Viet Nam · University of Health Sciences (UHS) Viet Nam National University Ho Chi Minh City Ho Chi Minh City 700000 Viet Nam ·

Abstract

Rapid and reliable identification of pharmaceutical residues is important for safeguarding public health, ensuring food safety, and enabling practical Raman-based screening. In this study, we propose HyMLRaman, a hybrid Raman spectroscopy framework that combines deep spectral feature extraction, generative models, and classical machine-learning classifiers to identify six pharmaceutical compounds, including amoxicillin, chloramphenicol, ciprofloxacin, tetracycline, ibuprofen, and paracetamol. Raman spectra are converted into spectral images and encoded with several deep neural-network backbones, among which EfficientNet-B3 yields the most effective representation. The resulting 1536-dimensional embeddings are then used to train downstream classifiers, including SVM, KNN, logistic regression, random forest, XGBoost, and ANN, using stratified 10-fold cross-validation. The hybrid EfficientNet-B3--SVM configuration achieves the strongest baseline performance, reaching 96.31% accuracy and a macro-F1 score of 96.36%, outperforming the standalone CNN baseline. To address limited-data conditions, a generative model, a DDPM-based feature augmentation, is introduced in a PCA-reduced EfficientNet-B3 latent space. The low-data ablation results show that DDPM augmentation provides selective benefits, particularly for KNN with reduced training fractions, and that its effect remains classifier-dependent. Finally, an application-level Raman Pharmaceutical Analyzer demonstrates the feasibility of embedding the trained model into an interactive Raman analysis workflow. These results suggest that HyMLRaman provides a practical and interpretable route for rapid Raman-based pharmaceutical screening.

Figures & tables

Figure 1: Schematic overview of the proposed HyMLRaman framework. (a) Cross-validation pipeline: (1) the data preprocessing, in which interpolated spectra are reshaped into $128\times 128$ spectral images, (2) feature embeddings (1536-D) are extracted using several ML architectures (EfficientNet-B0, EfficientNet-B3, RegNetY-400M, and ViT-B16), and multiple classifiers (SVM, KNN, LR, RF, ANN, XGBoost) are evaluated under stratified 10-fold cross-validation with visualization and confusion-matrix analysis, (3) an end-to-end application is built based on the trained EfficientNet-B3 classifier. (b) DDPM-based diffusion augmentation for generating feature embeddings via forward noise injection and learned reverse denoising.

Figure 2: Overview of the Raman spectral dataset. (a) Number of Raman spectral images in each pharmaceutical class: amoxicillin (AMX), chloramphenicol (CHL), ciprofloxacin (CIP), ibuprofen (IBU), paracetamol (PAR), and tetracycline (TET). (b) Representative normalized Raman spectra of the six pharmaceutical compounds, vertically offset for clarity. (c) t-SNE visualization of the EfficientNet-B3 feature embeddings for the six pharmaceutical classes. (d) Summary of the dataset, including the total number of samples, number of classes, data type, and classification task.

Figure 3: Top-1 accuracy comparison between standalone deep learning backbones and classical machine-learning classifiers trained on EfficientNet-B3 embeddings. The orange bar highlights the best-performing model, SVM, which achieves 96.31% accuracy. Deep learning backbones are shown in light gray, while classical machine-learning classifiers are shown in blue.

Hybrid ML model	Acc. (%)	Prec. (%)	Rec. (%)	F1 (%)	Acc. std.	F1 std.
EfficientNet-B3–SVM	96.31	96.57	96.35	96.36	2.14	2.12
EfficientNet-B3–ANN	95.91	96.19	95.91	95.93	1.61	1.61
EfficientNet-B3–LR	95.61	95.86	95.58	95.62	2.14	2.15
EfficientNet-B3–XGB	95.31	95.69	95.43	95.42	1.85	1.86
EfficientNet-B3–RF	94.91	95.26	94.99	94.97	1.60	1.55
EfficientNet-B3–KNN	94.62	95.03	94.64	94.66	1.75	1.77

Table 1: Baseline performance of the hybrid EfficientNet-B3–ML classifiers without DDPM augmentation, in which ML includes SVM, ANN, LR, XGB, RF, KNN, and CNN models.

Frac.	Model	Original Acc.	DDPM Acc.	$\boldsymbol{\Delta}$ Acc.	Original F1	DDPM F1	$\boldsymbol{\Delta}$ F1
25%	KNN	93.616	94.115	+0.499	93.628	94.154	+0.526
25%	SVM	94.510	94.811	+0.301	94.582	94.861	+0.279
25%	RF	93.715	93.714	-0.001	93.737	93.756	+0.019
25%	LR	94.311	94.212	-0.099	94.356	94.235	-0.121
25%	XGB	93.119	92.919	-0.200	93.179	92.902	-0.277
25%	ANN	94.810	94.210	-0.600	94.838	94.225	-0.613

Table 2: Low-data ablation of DDPM-based feature augmentation using PCA-reduced EfficientNet-B3 latent features. Accuracy and macro-F1 values are reported as mean percentages over stratified ten-fold cross-validation. Positive

\Delta

values indicate improvement after DDPM-based feature augmentation.

Figure 4: t-SNE visualization of original and filtered DDPM-generated latent features under the 25% training-data setting. The original samples are shown as circles, whereas DDPM-generated samples are shown as cross markers. The synthetic samples are generated in the 64-dimensional PCA-reduced EfficientNet-B3 feature space and retained after classifier-confidence and nearest-neighbor filtering.

Figure 5: Averaged confusion matrices of the embedding-level CNN baseline and hybrid CNN–ML classifiers under stratified ten-fold cross-validation, where CNN is the EfficientNet-B3 model and ML includes ANN, RF, XGBoost, SVM, KNN, and LR models. Values are reported as class-wise recall percentages. Class labels denote amoxicillin (AMX), chloramphenicol (CHL), ciprofloxacin (CIP), ibuprofen (IBU), paracetamol (PAR), and tetracycline (TET).

Figure 6: Class-level performance of the best-performing EfficientNet-B3–support vector machine (SVM) classifier under stratified ten-fold cross-validation. (a) Class-normalized confusion matrix, with values reported as class-wise recall percentages. (b) One-vs-rest ROC curves for the six pharmaceutical classes, together with the micro-averaged ROC curve and corresponding AUC values.

Figure 7: Application-level demonstration using the deployed Raman Pharmaceutical Analyzer for an independently measured paracetamol (PAR) sample. The interface displays the input Raman spectrum, the model-attention visualization, and the predicted class-probability distribution. The sample is correctly identified as PAR with a prediction confidence of 98.0%. Additional representative results for CIP, IBU, and TET are provided in Figure S4 of the Supporting Information.

Raman Spectroscopy

Raman spectroscopy uses inelastically scattered light to identify molecular vibrations and characterize materials, chemicals, and biological samples.

See all 16 papers

RamanPFN: learning from Raman spectral structure with a tabular foundation model
Aug 3, 2026Xingyu Pan, Huan Wang, Jinjia Guo +3Tabular Foundation Models Representation Learning
Beihang University · Cleer Science
Generative Augmentation of Raman Spectra for Glioma Classification
Jul 11, 2026Andrei Iuşan, Iulian Vasile, Daria Voiculescu +4Synthetic Data Augmentation Raman Spectroscopy
National Institute of Research and Development for Biological Sciences, Romania · Department of Mathematics and Statistics, University of Turku, Finland · Research Institute for Artificial Intelligence “Mihai Dragănescu”, Romanian Academy +2
Sharpness-Aware Minimization (SAM) Improves Classification Accuracy of Bacterial Raman Spectral Data Enabling Portable Diagnostics
Sep 16, 2026Kaitlin Zareno, Jarett Dewbury, Siamak K. Sorooshyari +2Raman Spectroscopy
Department of Mechanical Engineering MASSachusetts Institute of Technology Cambridge, MA 02139, USA · Department of Statistics Stanford University Palo Alto, CA 94305, USA · Google Research Mountain View, CA 94043, USA

Aug 3, 2026J/K move · Enter open · S save

RamanPFN: learning from Raman spectral structure with a tabular foundation model

Xingyu Pan, Huan Wang, Jinjia Guo, Zhenlin Zhao+2

Beihang University · Cleer Science

Tabular Foundation Models Representation Learning Physics-Informed ML Raman Spectroscopy Spectral Methods

Open paper AskPDF arXiv

Raman spectroscopy enables label-free molecular characterization across materials science, analytical chemistry, biomedicine, and industrial process monitoring. However, machine learning for high-dimensional spectroscopy remains constrained by limited labelled data and a mismatch between the physical organization of spectra and feature-agnostic models. Channel coverage alone does not ensure that related bands share a common inference context. Here we present RamanPFN, a general-purpose spectral foundation framework that enables unified in-context inference through physics-guided spectral learning. It captures full-spectrum compositional covariation via Global Compositional Unmixing (GCU), which decomposes distributed, multi-band mixture signatures into shared non-negative latent bases. Simultaneously, it resolves local vibrational structure through Local Vibrational Subspace Encoding (LVSE), which preserves fine-grained peak morphology, intensity fluctuations, and peak shifts within contiguous spectral neighborhoods. Extensive evaluation across 74 diverse public Raman datasets covered 129 regression targets and was further extended to 21 classification tasks. RamanPFN achieved state-of-the-art performance across all reported aggregate metrics against 28 independently reproduced methods spanning chemometrics, spectral neural networks, deep tabular learners and tabular foundation models. RamanPFN establishes a physics-guided paradigm for scientific spectroscopy, enabling data-efficient predictive learning across diverse chemical systems.

Abstract

Raman Spectroscopy

Explore similar work

RamanPFN: learning from Raman spectral structure with a tabular foundation model

Generative Augmentation of Raman Spectra for Glioma Classification

Sharpness-Aware Minimization (SAM) Improves Classification Accuracy of Bacterial Raman Spectral Data Enabling Portable Diagnostics