Reading the whole corpus: a reproducible cross-sample Raman workflow for materials characterisation, with a spectral 3D Gaussian splatting representation for hyperspectral maps
- 1Department of Chemical Engineering & Applied Chemistry, University of Toronto, 200 College Street, Toronto, Canada, M5S 3E5
- 2Leslie Dan Faculty of Pharmacy, University of Toronto, 144 College Street, Toronto, Canada, M5S 3M2
- 3Department of Mechanical and Industrial Engineering, University of Toronto, 5 King’s College Road, Toronto, Canada, M5S 3G8
Abstract
Raman mapping produces a hyperspectral cube, and turning that cube into numbers depends strongly on a chain of preprocessing choices. Two further problems appear once a laboratory holds many maps rather than one: the corpus is analysed one file at a time, and the analysable corpus is only the subset somebody remembered to export by hand from the instrument’s proprietary format. We present RAMANSP, a compact open framework providing typed spectral containers, composable preprocessing protocols and an analysis layer, and extend it in three directions. First, a direct reader for the undocumented Horiba LabSpec 6 container, reverse engineered here and verified bit identical against the instrument’s own text export, which removes the manual export step and takes this corpus from 1 usable map to 23. Second, a cross-sample knowledge graph linking every acquisition by spectral similarity, shared unmixing endmembers and disorder-metric proximity, with each metric edge carrying the preprocessing protocol under which it was computed. Third, and as the main methodological contribution, spectral 3D Gaussian splatting: a map V (x,y,ν) is represented by a differentiable field of anisotropic ellipsoids, each localising a domain in the two spatial axes and a band along the wavenumber axis, fitted by first-order descent on closed-form gradients, entirely in NumPy on a CPU. On an anonymised corpus of 108 acquisitions the graph recovers 23 communities at modularity Q = 0.65 and flags 5 groups of byte-identical duplicates, and across 6 maps the field reconstructs held-out wavenumber channels to within 0.70 to 1.34 of each map’s own measurement noise at 8.1 to 37.7× compression, that is, essentially to the noise floor.
Related
- Scaling multimodal materials representation learning with synthetic narratives
- Scaling Density Functional Theory with Gaussian Splatting
- Angle-tunable air-gap Fabry–Pérot cavities for inline spectroscopy and hyperspectral imaging
- Hyperspectral Imaging for Agricultural and Food Sensing: Applications, Data Analytics, and Deployment Challenges.
- Hyperspectral reflectance imaging to establish non-invasive phenotyping markers in microalgae
BibTeX
@misc{pahari2026reading,
title = {{Reading the whole corpus: a reproducible cross-sample Raman workflow for materials characterisation, with a spectral 3D Gaussian splatting representation for hyperspectral maps}},
author = {Siddhartha Pahari and Jainish Shailesh Solanki},
howpublished = {ChemRxiv},
year = {2026},
month = sep,
doi = {10.26434/chemrxiv.15008805/v1},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15008805/v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - Reading the whole corpus: a reproducible cross-sample Raman workflow for materials characterisation, with a spectral 3D Gaussian splatting representation for hyperspectral maps
AU - Siddhartha Pahari
AU - Jainish Shailesh Solanki
PY - 2026
DA - 2026/09/14
PB - ChemRxiv
DO - 10.26434/chemrxiv.15008805/v1
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15008805/v1
AB - Raman mapping produces a hyperspectral cube, and turning that cube into numbers depends strongly on a chain of preprocessing choices. Two further problems appear once a laboratory holds many maps rather than one: the corpus is analysed one file at a time, and the analysable corpus is only the subset somebody remembered to export by hand from the instrument’s proprietary format. We present RAMANSP, a compact open framework providing typed spectral containers, composable preprocessing protocols and an analysis layer, and extend it in three directions. First, a direct reader for the undocumented Horiba LabSpec 6 container, reverse engineered here and verified bit identical against the instrument’s own text export, which removes the manual export step and takes this corpus from 1 usable map to 23. Second, a cross-sample knowledge graph linking every acquisition by spectral similarity, shared unmixing endmembers and disorder-metric proximity, with each metric edge carrying the preprocessing protocol under which it was computed. Third, and as the main methodological contribution, spectral 3D Gaussian splatting: a map V (x,y,ν) is represented by a differentiable field of anisotropic ellipsoids, each localising a domain in the two spatial axes and a band along the wavenumber axis, fitted by first-order descent on closed-form gradients, entirely in NumPy on a CPU. On an anonymised corpus of 108 acquisitions the graph recovers 23 communities at modularity Q = 0.65 and flags 5 groups of byte-identical duplicates, and across 6 maps the field reconstructs held-out wavenumber channels to within 0.70 to 1.34 of each map’s own measurement noise at 8.1 to 37.7× compression, that is, essentially to the noise floor.
ER -