Spectral kernel machines with electrically tunable photodetectors
Data files
Sep 16, 2025 version files 7.14 MB
-
Fig_2D.txt
13.37 KB
-
Fig_2E.txt
24.81 KB
-
Fig_2F.txt
24.32 KB
-
Fig_2G.txt
59.73 KB
-
Fig_2H.txt
57.62 KB
-
Fig_2I.txt
58.27 KB
-
Fig_2J.txt
57.62 KB
-
Fig_3B.txt
275 B
-
Fig_3C.txt
193 B
-
Fig_3F.txt
349 B
-
Fig_4C.txt
3.51 KB
-
Fig_4D.txt
44.52 KB
-
Fig_4F.txt
35.08 KB
-
Fig_4H.txt
280 B
-
Fig_5A.txt
2.88 MB
-
Fig_5B.mat
2.10 MB
-
Fig_5C.mat
1.77 MB
-
Fig_5D.txt
126 B
-
Fig_5E.txt
92 B
-
Fig_5F.txt
74 B
-
Fig_5G.txt
76 B
-
README.md
5.81 KB
Abstract
Spectral machine vision collects spectral and spatial information as dense 3D hypercubes and digitally processes them into scene recognition, which causes a data bottleneck, limiting power efficiency, frame rate, and spectral-spatial resolution. This work introduces a device architecture called spectral kernel machines (SKM) to overcome these bottlenecks. SKM directly compresses spectral analysis through the output photocurrent and learns from example objects to identify and classify new samples in a 'sniff-and-seek' mode. We experimentally demonstrated SKM with electrically tunable bipolar black phosphorous (bP)-MoS2 photodiodes in the near/mid-infrared band and silicon photoconductors in the visible band, performing versatile intelligent tasks from chemometrics to semiconductor metrology. This architecture consumes significantly lower power and is more than an order of magnitude faster than existing solutions for hyperspectral image analysis, defining an intelligent imaging and sensing paradigm with intriguing possibilities.
Dataset DOI: 10.5061/dryad.jh9w0vtpv
Description of the data and file structure
Data for the main figures of the paper entitled "Spectral kernel machines with electrically tunable photodetectors". The photocurrent values were measured from a silicon photoconductor for the visible and near infrared applications (Fig. 2, 3), and from a black phosphorus-MoS2 photodiode for the mid-infrared applications (Fig. 4). We used both Arduino and Keysight B2901A SMU to measure the readout. The simulation results were produced with python codes (Fig. 5).
Files and variables
File: Fig_2D.txt
Description: The LED spectra for the spectral encoding.
Variables
- Wavelength(nm)
- normalized_power_band1
- normalized_power_band2
- normalized_power_band3
- normalized_power_band4
- normalized_power_band5
- normalized_power_band6
File: Fig_2E.txt
Description: The weighted spectrum Q(λ) for highlighting the leaf regions.
Variables
- wavelength (nm)
- weighted power
File: Fig_2F.txt
Description: Q(λ) for highlighting the bird regions.
Variables
- wavelength (nm)
- weighted power
File: Fig_2G.txt
Description: photocurrent highlighting the leaf regions.
Variables
- 2D photocurrent distribution in unit of nA
File: Fig_2H.txt
Description: sign of photocurrent in Fig. 2G.
Variables
- 2D distribution of boolean variables
File: Fig_2I.txt
Description: photocurrent highlighting the bird regions.
Variables
- 2D photocurrent distribution in unit of nA
File: Fig_2J.txt
Description: sign of photocurrent in Fig. 2I.
Variables
- 2D distribution of boolean variables.
File: Fig_3B.txt
Description: Photocurrent values for wafer identification. Each pillar corresponds to one testing point illustrated in (A) with the same color. Regions with 90.5 nm oxide are supposed to show positive current.
Variables
- Sample point label
- Photocurrent (nA)
- Sample point label
- Photocurrent (nA)
File: Fig_3C.txt
Description: Photocurrent-based inference of oxide thicknesses.
Variables
- Actual oxide thickness (nm)
- Expected total photocurrent for ideal prediction (nA)
- Actual readout photocurrent from test group (nA)
- Actual oxide thickness (nm)
- Actual readout photocurrent from training group (nA)
File: Fig_3F.txt
Description: SKM results for leaf hydration level classification. The photocurrent values are plotted against the relative water concentrations of the test leaves never seen before.
Variables
- Relative water concentration
- Photocurrent (nA)
- Relative water concentration
- Photocurrent (nA)
File: Fig_4C.txt
Description: The responsivity calibrated at λ = 1.32 μm.
Variables:
- Gate voltage (V)
- Responsivity (mA/W)
File: Fig_4D.txt
Description: The detectivity calibrated with FTIR.
Variables
- Wavelength (um)
- Detectivity (Jones)
File: Fig_4F.txt
Description: The total photocurrent for the C8F18 sample and C2H6OS sample.
Variables
- Time (s)
- Photocurrent (nA)
- Time (s)
- Photocurrent (nA)
File: Fig_4H.txt
Description: The total photocurrents, trained at 1 pA every percentage, versus the ground truth water concentration in acetone.
Variables
- Actual concentration (%)
- Measured concentration (%) after training
- Actual concentration (%)
- Measured concentration (%) before training
File: Fig_5A.txt
Description: Crop identification with the Salinas Valley hyperspectral dataset.
Variables
- 2D distribution of the output photocurrent (arbitrary units).
File: Fig_5B.mat
Description: Example SKM identification results from a hyperspectral city scene dataset LIB-HSI.
Variables
- 2D distribution of the output photocurrent (arbitrary units).
File: Fig_5C.mat
Description: Example tumor diagnosis results from a hyperspectral human brain dataset.
Variables
- 2D distribution of the output photocurrent (arbitrary units).
File: Fig_5D.txt
Description: The sensitivity of the classification results in (A)-(C), with a linear-kernel support vector machine run in a digital processor, a transformer run in a digital processor, and an in-sensor SKM.
Variables
- Position labels
- Sensitivity for linear SVM with digital processor.
- Position labels
- Results for transformer with digital processor.
- Position labels
- Results with in-sensor SKM.
File: Fig_5E.txt
Description: The specificity comparisons.
Variables
- Position labels
- Specificity for linear SVM with digital processor.
- Position labels
- Results for transformer with digital processor.
- Position labels
- Results with in-sensor SKM.
File: Fig_5F.txt
Description: Power consumption for capturing and processing one hypercube of each dataset.
Variables
- Position labels
- Power consumption for linear SVM with digital processor (J/hypercube)
- Position labels
- Results for transformer with digital processor (J/hypercube)
- Position labels
- Results with in-sensor SKM (J/hypercube)
File: Fig_5G.txt
Description: The latency for capturing and processing one hypercube.
Variables
- Position labels
- Latency for linear SVM with digital processor (seconds/hypercube)
- Position labels
- Results for transformer with digital processor (seconds/hypercube)
- Position labels
- Results with in-sensor SKM (seconds/hypercube)
Code/software
A MATLAB or related Python package is necessary to read the data in Fig. 5B and 5C in .mat format. All other data are stored as .txt files.