Using Data-Derived Priors to Guide CNN Architecture Design for NIR Chemometrics

Abstract

Convolutional neural networks (CNN) for near-infrared (NIR) chemometrics are often designed using generic architectural rules, although spectral datasets differ in sampling, smoothness, redundancy, and sample size. We tested whether these properties can provide empirical priors for CNN design. Across 25 NIR regression tasks, we computed descriptors of dataset size, spectral length and spacing, entropy, intrinsic rank, autocorrelation, and wavelet-scale structure. Two interpretable 1D-CNN scaffolds (a minimal single-convolution model and an extended shallow model with optional branching, dilation, etc) were optimized using five-fold cross-validated Bayesian hyperparameter optimization (HPO). Relationships extracted from near-optimal trials were converted into warm-start heuristics and evaluated directly and through leave-one-dataset-out (LODO) validation. The clearest relationships involved convolutional receptive fields. In the minimal CNN, the preferred kernel fraction decreased with spectral entropy and intrinsic rank, increased with the wavelet energy-support fraction, and the learning rate tended to decrease with training-set size. Direct and LODO heuristics were competitive with HPO, with median test-RMSE ratios of 0.953 and 1.017, respectively. The extended CNN showed similar but less transferable structure across branch usage, dilation, dropout, filter counts, and receptive-field choices. Ten stochastic refits showed seed sensitivity comparable to that of HPO-selected configurations. In a separate experiment, joint preprocessing and CNN HPO outperformed standardized-spectra HPO in 19 of 25 tasks, although gains were dataset-dependent. These results show that spectral descriptors can provide practical CNN design priors, guiding shallow NIR models toward plausible hyperparameter regions before target-specific tuning

Constitutive Priors for Inverse Design
Jinkyo Han, Bahador Bahmani
arXiv21 Aug 2026PreprintCC BY 4.0Full textComputational physicsarXiv
Shielding design and construction for the XINGZHI cold neutron triple-axis spectrometer: integrating velocity selector, S-bender polarizer, and neutron guide
Hongxia Zhang, Daye Xu, Florian Grünauer +5
arXiv30 Oct 2026PreprintarXiv licenceFull textApplied physics & instrumentationarXiv
Project Setu: 3D Multi-Physics Design and Scaled Structural Analysis for a Relativistic Lightsail Architecture
Prashant Suresh Kamble
arXiv10 Sep 2026PreprintarXiv licenceFull textInstrumentation & methods Solar & stellar physicsarXiv
From AI-assisted to AI-native Industrial Control: Architecture, Agent Design, and Challenges
Lisi Liu, Bin Wang, Quan Tang +3
arXiv26 Oct 2026PreprintarXiv licenceFull textStatistical & nonlinear physicsarXiv
Design and Finite-Element Analysis of a New Inclined Blade-Electrode Architecture for Trapped-Ion Quantum Information Processing
Nahiyan Archa, Abhinand P, Ahammed Shabeeb +1
arXiv11 Sep 2026PreprintCC BY 4.0Full textCondensed matter Quantum physics & quantum informationarXiv

Cite

BibTeX RIS

BibTeX

@misc{passos2026using,
  title = {{Using Data-Derived Priors to Guide CNN Architecture Design for NIR Chemometrics}},
  author = {Dário Passos},
  howpublished = {arXiv},
  year = {2026},
  month = oct,
  doi = {10.48550/arxiv.2607.25636},
  eprint = {2607.25636},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2607.25636},
  note = {CC BY 4.0},
}

RIS

TY  - UNPB
TI  - Using Data-Derived Priors to Guide CNN Architecture Design for NIR Chemometrics
AU  - Dário Passos
PY  - 2026
DA  - 2026/10/20
PB  - arXiv
DO  - 10.48550/arxiv.2607.25636
UR  - https://arxiv.org/abs/2607.25636
AB  - Convolutional neural networks (CNN) for near-infrared (NIR) chemometrics are often designed using generic architectural rules, although spectral datasets differ in sampling, smoothness, redundancy, and sample size. We tested whether these properties can provide empirical priors for CNN design. Across 25 NIR regression tasks, we computed descriptors of dataset size, spectral length and spacing, entropy, intrinsic rank, autocorrelation, and wavelet-scale structure. Two interpretable 1D-CNN scaffolds (a minimal single-convolution model and an extended shallow model with optional branching, dilation, etc) were optimized using five-fold cross-validated Bayesian hyperparameter optimization (HPO). Relationships extracted from near-optimal trials were converted into warm-start heuristics and evaluated directly and through leave-one-dataset-out (LODO) validation. The clearest relationships involved convolutional receptive fields. In the minimal CNN, the preferred kernel fraction decreased with spectral entropy and intrinsic rank, increased with the wavelet energy-support fraction, and the learning rate tended to decrease with training-set size. Direct and LODO heuristics were competitive with HPO, with median test-RMSE ratios of 0.953 and 1.017, respectively. The extended CNN showed similar but less transferable structure across branch usage, dilation, dropout, filter counts, and receptive-field choices. Ten stochastic refits showed seed sensitivity comparable to that of HPO-selected configurations. In a separate experiment, joint preprocessing and CNN HPO outperformed standardized-spectra HPO in 19 of 25 tasks, although gains were dataset-dependent. These results show that spectral descriptors can provide practical CNN design priors, guiding shallow NIR models toward plausible hyperparameter regions before target-specific tuning
ER  -

Abstract

Related

Cite