At SciX 2026, Thursday morning’s “New Challenges in Chemometrics” session questioned long-held assumptions, from reliance on autonomous model selection to the restriction of spectral data to real numbers.
Five talks in session CHEM-Thu1 addressed emerging problems in the multivariate analysis of spectroscopic data. Presenters described virtual reality–assisted local regression for near-infrared (NIR) spectra, a comparison of target detection algorithms for spectral imaging, complex-valued regression for infrared and Raman spectra, outlier diagnostics for calibration-free models in continuous pharmaceutical manufacturing, and a unified covariance-based framework for feature selection. Together, the talks showed a field adapting to larger, more complex data.
Chemometrics supplies the mathematics that turns spectra into chemical answers. Larger libraries, hyperspectral images, and multiblock data now raise new questions about how models are built, selected, and monitored. Session CHEM-Thu1, “New Challenges in Chemometrics,” took up these questions on Thursday, October 8, 2026, in Southern Pacific C. Five 20-minute presentations ran from 10:30 AM to 12:10 PM PDT, spanning visualization, imaging, vibrational spectroscopy, process analytical technology (PAT), and variable selection.
At 10:30 AM, John H. Kalivas, professor emeritus at Idaho State University, opened the session with work coauthored with Hyrum Redd of Idaho State University. Kalivas argued that virtual reality (VR) lets analysts base data decisions on how the data look, rather than on algorithmic score values alone. He briefly reviewed earlier VR work on outlier detection, classification, and tuning parameter selection.
The main topic was local regression, in which a separate model is built for each prediction sample from calibration samples drawn out of a large spectral library. Those samples must be matrix-matched to the unknown and should bracket its analyte amount. The group’s autonomous local adaptive fusion regression (LAFR) algorithm performs this selection but depends on many operational parameters; it had been tested on a NIR soil library of nearly 100,000 spectra. In the VR version, users viewed the matrix effect patterns of library and prediction samples and picked visually similar calibration samples. Kalivas reported that for several NIR data sets, the VR approach could outperform LAFR and ran much faster. The work matters because selecting matrix-matched calibration samples from large libraries remains a bottleneck for local modeling, and a visual route could open such methods to analysts with limited chemometrics training.
Talk summary: Visual comparison of matrix effect patterns in VR offered a faster, parameter-light alternative to autonomous local regression for NIR data.
At 10:50 AM, Neal B. Gallagher, vice-president of Eigenvector Research Inc., presented a comparison coauthored with Nathanial Watson, also of Eigenvector Research. Gallagher explained that spectral imaging is well suited to detecting minor signal sources: a target can dominate individual pixels, many pixels are available to characterize interference (clutter), and clutter suppression can be tailored to each image.1
Four algorithms were evaluated: spectral angle mapper (SAM), generalized SAM (GSAM), adaptive cosine estimator (ACE), and iterative target detection (ITD).2 The test case was a laboratory NIR image of plastic particles on sand, with ground truth established by multivariate curve resolution. SAM, while simple and interpretable, performed worst. GSAM, ACE, and ITD performed similarly on pure or nearly pure target pixels. For partially filled pixels, GSAM gave the best results, ITD was close behind, and ACE was less accurate. Because the comparison rested on rigorously established ground truth, it gives practitioners evidence-based guidance for choosing a detection algorithm, particularly when targets only partly fill a pixel.
Talk summary: Tested against ground truth, GSAM performed best for partially filled pixels, while SAM performed worst.
At 11:10 AM, Thomas G. Mayerhöfer, senior researcher at the Leibniz Institute of Photonic Technology in Jena, Germany, presented work with Jürgen Popp, the institute’s scientific director. Mayerhöfer questioned the convention that chemometric analysis uses only real-valued spectra.3 Building on earlier findings that using both parts of the complex refractive index reduces systematic error in infrared spectroscopy,4 the team described a framework based on causal response functions. Through Kramers–Kronig relations, both absorbance and Raman spectra were converted into complex-valued representations.5
The group presented complex-valued versions of classical least squares (CLS), inverse least squares (ILS), principal component regression (PCR), and partial least squares regression (PLSR). These were tested on ideal mixtures (benzene–toluene and benzene–cyclohexane), a non-ideal mixture (acetone–chloroform), and blood plasma. The complex-valued methods consistently lowered prediction errors compared with conventional approaches.
Complex-valued ILS matched or exceeded complex-valued PLSR, and imaginary concentration components revealed deviations from linear mixing. The transforms take seconds with FFT-based routines. If broadly adopted, the approach could reduce systematic errors that arise when linear models are applied to real mixtures, while adding physical interpretability to familiar regression tools.
Talk summary: Complex-valued forms of absorbance and Raman spectra lowered prediction errors across four standard regression methods.
At 11:30 AM, Md Nahid Hasan, a PhD candidate at Duquesne University, presented work coauthored with James K. Drennen, dean of the School of Pharmacy, and Carl A. Anderson, associate dean for research and graduate programs, both at Duquesne. The talk addressed PAT in continuous pharmaceutical manufacturing, where pure component models such as Iterative Optimization Technology (IOT) predict active pharmaceutical ingredient (API) concentration without calibration.
Unlike PLS models, which rely on Q-residuals and Hotelling’s T2, pure component models lack a defined calibration space, which complicates in-model diagnostics. Hasan proposed two complementary metrics: scaled net analyte signal distance (SND), which measures in-model variance relative to a theoretical calibration center, and relative squared spectral residual (rSSR), which measures out-of-model variance. Used together, the two metrics reliably flagged outlier samples. The diagnostics address a practical gap in regulated manufacturing, where calibration-free models need defensible ways to signal when a prediction should not be trusted.
Talk summary: Combining SND and rSSR enabled outlier detection for calibration-free pure component models.
The session closed at 11:50 AM with Federico Marini, full professor at the University of Rome “La Sapienza,” Italy. Marini presented the Covariance Selection (CovSel) family as a unified framework for feature selection. The original CovSel iteratively selects variables with maximum covariance with the response. Its extensions redefine the candidate feature: window CovSel (wCovSel) selects contiguous spectral intervals; group CovSel (g-CovSel) selects correlated but not necessarily adjacent variables; sequential orthogonalized CovSel (SO-CovSel) handles multiblock data while removing redundancy among blocks; and N-CovSel works directly on multiway arrays. Applications to spectroscopic, hyperspectral, and multimodal data illustrated more parsimonious, interpretable models. Marini concluded that the structure of the data, not the algorithm, should define what a selected feature is. A single optimization principle that adapts to intervals, blocks, and tensors could simplify method choice as hyperspectral and multimodal data sets become routine.
Talk summary: The CovSel family applies one covariance-based principle to variables, intervals, groups, blocks, and multiway arrays.
The session showed chemometrics revisiting its foundations. Whether by restoring human judgment, testing algorithms against ground truth, or rethinking the numbers themselves, the speakers shared a goal of models that are more accurate, more interpretable, and better suited to their data.
Gallagher, N. B.; Goyetche, R.; Amigo, J. M.; Kucheryavskiy, S. Extended Least Squares (ELS) and Generalized Least Squares (GLS) for Clutter Suppression in Hyperspectral Images: A Theoretical Description. Chemom. Intell. Lab. Syst. 2024, 244, 105032. DOI:
Forland, B. M.; Gallagher, N. B.; Johnson, T. J. Application of Automated Iterative Target Detection for Standoff Hyperspectral Imaging. J. Appl. Remote Sens. 2025, 19 (1), 016509. DOI:
Mayerhöfer, T. G. Wave Optics in Infrared Spectroscopy: Theory, Simulation, and Modeling; Elsevier: Philadelphia, 2024. ISBN 978-0-443-22031-9.
Mayerhöfer, T. G.; Ilchenko, O.; Kutsyk, A.; Popp, J. Complex-Valued Chemometrics in Spectroscopy: Partial Least Squares Regression. Appl. Spectrosc. 2026, 80 (4), 416–427. DOI:
Mayerhöfer, T. G.; Ilchenko, O.; Kutsyk, A.; Piehler, S.; Silge, A.; Ramoji, A.; Winterfeld, A.; Ryabchykov, O.; Kiehntopf, M.; Bocklitz, T.; Popp, J. Complex-Valued Chemometrics for Analyzing Absorbance or Raman Spectra. Anal. Chem. 2026, 98 (15), 10953–10965. DOI:
