Light That Thinks: How Near-Infrared Spectroscopy Learned to See Inside Bodies, Bioreactors, and the Biosphere

Abstract

Near-infrared (NIR) spectroscopy has spent nearly 58 years developing as a routine quantitative analysis technique, becoming the analytical workhorse of agricultural and pharmaceutical applications. Over the years it has been valued for its speed rather than for its glamour. That reputation is now being rewritten. Chip-scale sensors have shrunk benchtop instruments into pocket devices, while transformer networks, generative adversarial models, and uncertainty-aware machine learning are turning static calibration curves into adaptive, self-correcting systems. Clinicians are using functional NIR to detect hidden consciousness in unresponsive brain-injury patients and to distinguish mild cognitive impairment from Alzheimer’s disease. Biomanufacturers are letting NIR, fused with Raman data, run closed-loop feeding strategies without operator intervention. Environmental scientists are pointing handheld NIR units at rivers and soils to chase microplastics and carbon at a scale field chemistry never allowed. This article surveys the developments most likely to reshape how readers think about NIR over the next several years, drawn from primary literature and reporting published over the past five years.

Introduction

Ask a spectroscopist what near-infrared (NIR) spectroscopy is good for, and the answer used to be predictable: moisture in wheat, fat in milk, active ingredient in a tablet. The technique’s overtone and combination bands are broad, weak, and famously difficult to interpret by eye, which is precisely why NIR spent decades as a chemometrics-dependent specialty rather than a headline-grabbing method.¹ That dependency turned out to be an advantage once artificial intelligence (AI) caught up with the mathematics. Because NIR has always required a model to translate spectra into meaning, it was primed to absorb deep learning faster than techniques that could be interpreted directly. The result, visible across the literature from 2021 through 2026, is a technique that has quietly become one of the most AI-native tools in analytical science.² The developments below are not incremental refinements of existing calibration models. They represent NIR functioning in roles it was never designed for: reading brain activity through an intact skull, running a fermentation without supervision, and screening for plastic particles too small and too transparent to see.

From Benchtop to Pocket: The Miniaturization Leap

The most consequential hardware change in NIR over the past five years has been the collapse of the instrument itself. Fully integrated sensor-chip modules now cover roughly 850–1700 nm in a footprint that fits in a hand, eliminating the moving gratings and cooled detectors that once confined serious NIR work to a bench.¹ These chip-based modules retain enough spectral resolution for real quantitative work, not just crude sorting, which is why they have spread into agricultural grading, pharmaceutical raw-material screening, food authentication, and wood-species identification in the field rather than the lab.^1,3 Portable NIR has been used to distinguish visually similar Cinnamomum wood species on-site, a task that previously required destructive sampling and laboratory referral.³ The practical effect is that NIR measurements are migrating to wherever the sample already is, rather than requiring the sample to travel to the instrument.

Teaching Machines to Read Light: AI Rewrites NIR Calibration

Classical NIR calibration relied on partial least squares regression and principal component analysis, techniques that assume roughly linear relationships between absorbance and concentration.² That assumption breaks down in complex biological or industrial matrices, and 2025 supplied a wave of replacements. Convolutional neural networks are now used as checkpoint tools to characterize cell-culture media directly from raw spectra, linking media variability to downstream glycosylation outcomes without the intermediate step of hand-built calibration models.⁴ A separate framework, external variable augmented iterative optimization technology (EVA-IOT), cut the calibration burden for continuous pharmaceutical powder streams by as much as 97%, addressing the single biggest practical barrier to industrial NIR adoption: the expense of building and maintaining models in the first place.⁵ Model drift, long a quiet failure mode of process NIR, is also being addressed directly, with new maintenance and monitoring frameworks keeping in situ calibrations reliable across months of continuous fermentation.⁶

Autonomous Bioprocessing: NIR Learns to Drive the Factory Floor

Perhaps the clearest sign that NIR has changed roles is its move from monitoring to control. Researchers fusing NIR and Raman spectral streams with machine learning have demonstrated real-time, closed-loop control of gentamicin fermentation, using the combined spectral signal to automatically adjust feeding and stabilize substrate levels without a human adjusting a valve.⁷ That is a meaningfully different claim from “NIR can measure glucose concentration”; it is a claim that NIR can be trusted to make a process decision. Inline NIR combined with chemometrics is likewise now monitoring fast, exothermic reactions inside microreactors in real time, expanding process analytical technology into reaction environments once considered too fast or too hazardous for spectroscopic supervision.⁸ Doctoral-level work evaluating NIR alongside Raman as process analytical tools for monoclonal antibody manufacturing has further shown that NIR can track residual moisture during freeze-drying closely enough to influence release decisions, not just flag anomalies after the fact.⁹

Small Data, Big Answers: Synthetic Spectra Solve the Sample Problem

NIR calibration has always been data-hungry, and biological processes rarely produce enough labeled spectra to satisfy a deep neural network. The most inventive response has been to generate the missing data outright. A transformer-based architecture enhanced with a Wasserstein generative adversarial network created synthetic spectra to monitor hyaluronic acid fermentation, using the augmented dataset to overcome calibration limits imposed by sparse, noisy bioprocess sampling.¹⁰ The implication extends well beyond one fermentation: if a model can be trained on spectra that were never physically measured, the sample-scarcity problem that has limited NIR calibration in biopharmaceuticals, rare-disease diagnostics, and novel materials may no longer be a hard ceiling.

Seeing Consciousness, Diagnosing Disease: NIR Enters the Clinic

The clinical applications of NIR published since 2024 are striking for a technique historically associated with grain elevators. Functional NIR spectroscopy has been used to detect resting-state neural networks and task-driven brain activity in behaviorally unresponsive, severely brain-injured patients, revealing preserved awareness that standard bedside examination missed entirely.¹¹ That single finding challenges a diagnostic paradigm built on behavioral observation and raises real questions about how “unresponsive” is defined in intensive care. Elsewhere, NIR spectra of serum, combined with machine learning, have detected the presence of hepatitis C virus (HCV) in serum microsamples by identifying subtle shifts in the water and lipid matrix, and combining those spectral features with routine clinical variables raised diagnostic accuracy above 72%, with an area under the curve (AUC) of 0.850.^12,13 In neurology, functional NIR is now distinguishing mild cognitive impairment (MCI) from Alzheimer’s disease (AD) through consistent reductions in tissue oxygenation and functional connectivity, with machine-learning classifiers reportedly reaching accuracies near 90%.¹⁴ Liver disease has not been left out either: NIR combined with neural network and support vector machine classification is being used to stage fibrosis noninvasively, a task that traditionally required biopsy.¹⁵ Each of these results depends on NIR’s ability to probe hemoglobin, water, and lipid chemistry noninvasively through tissue, but the diagnostic reach is new.

Watching the Planet: NIR’s Environmental Frontier

Environmental monitoring has become one of NIR’s fastest-growing application areas, largely because the technique’s speed and minimal sample preparation suit field surveys better than laboratory-bound alternatives.¹⁶ NIR spectroscopy paired with machine learning has been used to detect colorless microplastic particles in environmental samples, a genuinely difficult target because such particles lack the strong, distinctive absorption bands that make mid-infrared identification straightforward.¹⁷ Soil and vegetation properties are being screened the same way, giving ecologists a rapid, nondestructive tool for large-scale conservation surveys where laboratory analysis of every sample would be impractical.¹⁸ None of these environmental use cases would be credible without the parallel rise of machine-learning classifiers capable of pulling a usable signal out of NIR’s comparatively weak and overlapping overtone bands.

Summary and Conclusions

Across biomedicine, bioprocessing, and environmental science, the pattern repeats: NIR spectroscopy’s underlying physics has not changed, but its interpretive layer has been transformed by machine learning, and that transformation has unlocked applications the technique was never built for. Chip-scale hardware put the instrument in the field.¹ Deep learning and synthetic data solved the calibration and sample-scarcity problems that once confined NIR to well-characterized industrial processes.^4,5,10 And the resulting systems are now trusted with tasks that once required a biopsy, a behavioral neurology exam, or a wet-chemistry laboratory.^11,14,15 The throughline is not any single instrument or algorithm; it is that NIR has shifted from a technique that reports a number to one that increasingly makes a judgment.

Future Outlook

The near-term trajectory points toward NIR systems that explain their own reasoning rather than simply outputting a prediction, an important shift as regulators and clinicians are asked to trust spectroscopic judgments in place of biopsies or wet chemistry.² Expect continued fusion of NIR with Raman, mass spectrometry, and imaging data streams, producing multimodal sensors that offer more complete chemical fingerprints than any single technique can provide.⁷ NIR is also beginning to support personalized medicine directly: NIR fused with Raman imaging is now being used to quantify drug content in porous, patient-specific formulations manufactured in small batches, a quality-control problem that destructive assay methods were never designed to solve at that scale.¹⁹ And as synthetic spectral generation matures, NIR calibration may increasingly be built before the relevant samples exist at all, a genuinely unfamiliar way to develop an analytical method.¹⁰ Whether these systems perform as reliably outside curated studies as they do within them remains the open question the field must answer next.²⁰

References

(1) Ou, F.; van Klinken, A.; Hakkel, K. D.; et al. Spectral Sensing Using a Handheld NIR Module Based on a Fully Integrated Sensor Chip. Spectroscopy 2022, 37 (S11), 34–38. DOI: 10.56530/spectroscopy.yd5989g6.

(2) Workman, J., Jr. AI Developments That Changed Vibrational Spectroscopy in 2025. Spectroscopy 2025, published online December 27, 2025. DOI: 10.56530/spectroscopy.no4686u3.

(3) Pan, X.; Qiu, J.; Yang, Z. Identification of Five Similar Cinnamomum Wood Species Using Portable Near-Infrared Spectroscopy. Spectroscopy 2022, 37 (6), 16–23, 49. DOI: 10.56530/spectroscopy.zg7089n4.

(4) Gangwar, N.; Balraj, K.; Rathore, A. S. Near-Infrared Spectroscopy Coupled with Convolutional Neural Network as a Checkpoint Tool for Cell Culture Bioprocess Media Characterization. Biotechnol. Prog. 2025, 41 (6), e70056. DOI: 10.1002/btpr.70056.

(5) Velez-Silva, N. L.; Rish, A. J.; Drennen, J. K., III; Anderson, C. A. Robust Near-Infrared Modeling for Pharmaceutical Powder Streams: External Variable Augmented Iterative Optimization Technology (EVA-IOT). Eur. J. Pharm. Biopharm. 2025, 207, 114626. DOI: 10.1016/j.ejpb.2025.114626.

(6) Chandrasekaran, K.; Chandrasekaran, K.; Jayaraman, G.; Bhatt, N. Model Maintenance, Monitoring, and Control Framework Using In-Situ NIR Spectroscopy and Offline Process Data for Lactococcus lactis Fermentation. Chem. Eng. J. 2025, 519, 165116. DOI: 10.1016/j.cej.2025.165116.

(7) Xu, F.; Su, L.; Gao, H.; Wang, Y.; Ben, R.; Hu, K.; Mohsin, A.; Li, C.; Chu, J.; Tian, X. Harnessing Near-Infrared and Raman Spectral Sensing and Artificial Intelligence for Real-Time Monitoring and Precision Control of Bioprocess. Bioresour. Technol. 2025, 421, 132204. DOI: 10.1016/j.biortech.2025.132204.

(8) Mahler, L.; Desel, P.; Sladkov, M.; Roppertz, A.; Mayer, C.; Jaeger, M. A Process Monitoring Microreactor Assembly for Real-Time Reaction Analysis Using Inline Near-Infrared Spectroscopy and Chemometrics. Anal. Bioanal. Chem. 2025, 417 (8), 1431–1439. DOI: 10.1007/s00216-025-05779-2.

(9) Massei, A. Spectroscopic Techniques as Process Analytical Technology Monitoring Tools for the Development of Biopharmaceutical Product Processes. Ph.D. Thesis, University of Bologna, Bologna, Italy, 2025; pp 26–52. https://tesidottorato.depositolegale.it/bitstream/20.500.14242/202045/1/conv_tesi_final_reviewed.pdf (accessed 2026-07-31).

(10) Tian, W.; Zang, L.; Li, Y.; Peng, C.; Zhang, J.; Shao, H.; Xian, R.; Sun, R.; Liu, F.; Tan, H.; Ling, P. A Transformer-Based Framework with Generative Spectral Augmentation for Online Monitoring of Hyaluronic Acid Fermentation. Carbohydr. Polym. 2025, 369, 124278. DOI: 10.1016/j.carbpol.2025.124278.

(11) Kazazian, K.; Abdalmalak, A.; Novi, S. L.; Norton, L.; et al. Functional Near-Infrared Spectroscopy: A Novel Tool for Detecting Consciousness after Acute Severe Brain Injury. Proc. Natl. Acad. Sci. U. S. A. 2024, 121 (36), e2402723121. DOI: 10.1073/pnas.2402723121.

(12) Gómez, J.; Barquero-Pérez, O.; Gonzalo, J.; Salgüero, S.; et al. Near Infrared Spectroscopy (NIRS) and Machine Learning as a Promising Tandem for Fast Viral Detection in Serum Microsamples: A Preclinical Proof of Concept. Spectrochim. Acta, Part A 2024, 322, 124819. DOI: 10.1016/j.saa.2024.124819.

(13) Pérez-Gómez, E.; Gómez, J.; Gonzalo, J.; Salguero, S.; Riado, D.; Casas, M. L.; Gutiérrez, M. L.; Jaime, E.; Pérez-Martínez, E.; García-Carretero, R.; Ramos, J.; Fernández-Rodríguez, C.; Català, M.; Martino, L.; Barquero-Pérez, Ó. Exploratory Integration of Near-Infrared Spectroscopy with Clinical Data: A Machine Learning Approach for HCV Detection in Serum Samples. Front. Med. 2025, 12, 1596476. DOI: 10.3389/fmed.2025.1596476.

(14) Li, H.; Yang, X.; Gong, L. Functional Near-Infrared Spectroscopy for Identifying Mild Cognitive Impairment and Alzheimer’s Disease: A Systematic Review. Front. Neurol. 2025, 16, 1578375. DOI: 10.3389/fneur.2025.1578375.

(15) Addissouky, T. A.; El Tantawy El Sayed, I.; Ali, M. M. A.; Alubiady, M. H. S. Optical Insights into Fibrotic Livers: Applications of Near-Infrared Spectroscopy and Machine Learning. Arch. Gastroenterol. Res. 2024, 5 (1), 1–10. DOI: 10.33696/Gastroenterology.5.048.

(16) Workman, J., Jr.; Wetzel, W. Infrared and Near-Infrared Spectroscopy: Recent Advances and Environmental Applications (2022–2026). Spectroscopy Online, June 18, 2026. https://www.spectroscopyonline.com/view/infrared-and-near-infrared-spectroscopy-recent-advances-and-environmental-applications-2022-2026 (accessed 2026-07-31).

(17) Wetzel, W. Detection of Colorless Microplastics in the Environment Using NIR Spectroscopy and Machine Learning. Spectroscopy Online, February 5, 2025. https://www.spectroscopyonline.com/view/detection-of-colorless-microplastics-in-the-environment-using-nir-spectroscopy-and-machine-learning (accessed 2026-07-31).

(18) Wetzel, W. Examining Soil and Leaf Properties Using NIR Spectroscopy. Spectroscopy Online, December 2, 2025. https://www.spectroscopyonline.com/view/examining-soil-and-leaf-properties-using-nir-spectroscopy (accessed 2026-07-31).

(19) Workman, J., Jr. Advancing Near-Infrared Spectroscopy and Machine Learning for Personalized Medicine. Spectroscopy Online, February 12, 2025. https://www.spectroscopyonline.com/view/advancing-near-infrared-spectroscopy-and-machine-learning-for-personalized-medicine (accessed 2026-07-31).

(20) Kos, J.; Pavelek, D.; Kaykhaii, M.; Olsen, M.; Jampilek, J.; Halko, R. Unveiling the Transformative Power of Near-Infrared Spectroscopy in Biomedical and Pharmaceutical Analysis: Trends, Advancements, and Applications. Eur. J. Pharm. Sci. 2025, 212, 107175. DOI: 10.1016/j.ejps.2025.107175.