EELS

Electron Energy Loss Spectroscopy

Join the Competition Contribute to Benchmark

Standard reconstruction benchmark — forward model perfectly known, no calibration needed. Score = 0.5 × clip((PSNR−15)/30, 0, 1) + 0.5 × SSIM

#	Method	Score	PSNR (dB)	SSIM	Source
🥇	DiffEELS DiffEELS Gao et al. 2024 39.3 dB SSIM 0.954 Checkpoint unavailable	0.882	39.3	0.954	✓ Certified	Gao et al. 2024
🥈	PhysEELS PhysEELS Chen et al. 2024 37.9 dB SSIM 0.942 Checkpoint unavailable	0.853	37.9	0.942	✓ Certified	Chen et al. 2024
🥉	SwinEELS SwinEELS Wang et al. 2023 36.7 dB SSIM 0.932 Checkpoint unavailable	0.828	36.7	0.932	✓ Certified	Wang et al. 2023
4	TransEELS TransEELS Li et al. 2022 35.1 dB SSIM 0.915 Checkpoint unavailable	0.792	35.1	0.915	✓ Certified	Li et al. 2022
5	N2V-EELS N2V-EELS Krull et al. 2019 32.6 dB SSIM 0.876 Checkpoint unavailable	0.731	32.6	0.876	✓ Certified	Krull et al. 2019
6	DnCNN-EELS DnCNN-EELS Kovarik et al. 2016 30.0 dB SSIM 0.838 Checkpoint unavailable	0.669	30.0	0.838	✓ Certified	Kovarik et al. 2016
7	ICA-EELS ICA-EELS Bosman et al. 2006 27.1 dB SSIM 0.786 Try in SpecLab →	0.595	27.1	0.786	✓ Certified	Bosman et al. 2006
8	MLS-EELS MLS-EELS Verbeeck & Van Aert 2004 24.5 dB SSIM 0.744 Try in SpecLab →	0.530	24.5	0.744	✓ Certified	Verbeeck & Van Aert 2004
9	PowerLaw-EELS PowerLaw-EELS Egerton 2011 21.8 dB SSIM 0.699 Try in SpecLab →	0.463	21.8	0.699	✓ Certified	Egerton 2011

Dataset: PWM Benchmark (9 algorithms)

Blind Reconstruction Challenge — forward model has unknown mismatch, must calibrate from data. Score = 0.4 × PSNR_norm + 0.4 × SSIM + 0.2 × (1 − ‖y − Ĥx̂‖/‖y‖)

#	Method	Overall Score	Public PSNR / SSIM	Dev PSNR / SSIM	Hidden PSNR / SSIM	Trust	Source
🥇	SwinEELS + gradient SwinEELS + gradient Wang et al., npj Comput. Mater. 2023 Score 0.773 Correct & Reconstruct →	0.773	0.806 34.4 dB / 0.964	0.767 31.84 dB / 0.941	0.745 31.06 dB / 0.932	✓ Certified	Wang et al., npj Comput. Mater. 2023
🥈	PhysEELS + gradient PhysEELS + gradient Chen et al., Microsc. Microanal. 2024 Score 0.768 Correct & Reconstruct →	0.768	0.842 36.71 dB / 0.977	0.739 30.78 dB / 0.928	0.722 28.85 dB / 0.898	✓ Certified	Chen et al., Microsc. Microanal. 2024
🥉	DiffEELS + gradient DiffEELS + gradient Gao et al., NeurIPS 2024 Score 0.751 Correct & Reconstruct →	0.751	0.859 37.82 dB / 0.981	0.739 30.29 dB / 0.921	0.654 26.34 dB / 0.841	✓ Certified	Gao et al., NeurIPS 2024
4	TransEELS + gradient TransEELS + gradient Li et al., Ultramicroscopy 2022 Score 0.743 Correct & Reconstruct →	0.743	0.783 32.29 dB / 0.946	0.752 31.47 dB / 0.937	0.694 28.56 dB / 0.892	✓ Certified	Li et al., Ultramicroscopy 2022
5	N2V-EELS + gradient N2V-EELS + gradient Krull et al., NeurIPS 2019 Score 0.640 Correct & Reconstruct →	0.640	0.744 29.72 dB / 0.912	0.612 23.43 dB / 0.748	0.563 22.1 dB / 0.694	✓ Certified	Krull et al., NeurIPS 2019
6	DnCNN-EELS + gradient DnCNN-EELS + gradient Kovarik et al., npj Comput. Mater. 2016 Score 0.566 Correct & Reconstruct →	0.566	0.733 28.93 dB / 0.899	0.517 20.17 dB / 0.607	0.448 18.4 dB / 0.520	✓ Certified	Kovarik et al., npj Comput. Mater. 2016
7	PowerLaw-EELS + gradient PowerLaw-EELS + gradient Egerton, EELS in the EM, Springer 2011 Score 0.482 Correct & Reconstruct →	0.482	0.488 18.87 dB / 0.543	0.485 19.02 dB / 0.551	0.474 18.73 dB / 0.536	✓ Certified	Egerton, EELS in the EM, Springer 2011
8	MLS-EELS + gradient MLS-EELS + gradient Verbeeck & Van Aert, Ultramicroscopy 2004 Score 0.471 Correct & Reconstruct →	0.471	0.612 23.06 dB / 0.733	0.429 17.62 dB / 0.481	0.372 16.0 dB / 0.401	✓ Certified	Verbeeck & Van Aert, Ultramicroscopy 2004
9	ICA-EELS + gradient ICA-EELS + gradient Bosman et al., Ultramicroscopy 2006 Score 0.456 Correct & Reconstruct →	0.456	0.645 24.94 dB / 0.800	0.412 16.92 dB / 0.446	0.310 13.77 dB / 0.300	✓ Certified	Bosman et al., Ultramicroscopy 2006

Complete score requires all 3 tiers (Public + Dev + Hidden).

Join the competition →

Scoring: 0.4 × PSNR_norm + 0.4 × SSIM + 0.2 × (1 − ‖y − Ĥx̂‖/‖y‖) PSNR 40% · SSIM 40% · Consistency 20%

Public 3 scenes

Full-access development tier with all data visible.

What you get & how to use

What you get: Measurements (y), ideal forward operator (H), spec ranges, ground truth (x_true), and true mismatch spec.

How to use: Load HDF5 → compare reconstruction vs x_true → check consistency → iterate.

What to submit: Reconstructed signals (x_hat) and corrected spec as HDF5.

Public Leaderboard

#	Method	Score	PSNR	SSIM
1	DiffEELS + gradient	0.859	37.82	0.981
2	PhysEELS + gradient	0.842	36.71	0.977
3	SwinEELS + gradient	0.806	34.4	0.964
4	TransEELS + gradient	0.783	32.29	0.946
5	N2V-EELS + gradient	0.744	29.72	0.912
6	DnCNN-EELS + gradient	0.733	28.93	0.899
7	ICA-EELS + gradient	0.645	24.94	0.8
8	MLS-EELS + gradient	0.612	23.06	0.733
9	PowerLaw-EELS + gradient	0.488	18.87	0.543

Spec Ranges (3 parameters)

Parameter	Min	Max	Unit
energy_dispersion	-0.002	0.004	eV/channel
zero_loss_shift	-0.3	0.6	eV
aberration	-2.0	4.0	%

View Details →

Dev 3 scenes

Blind evaluation tier — no ground truth available.

What you get & how to use

What you get: Measurements (y), ideal forward operator (H), and spec ranges only.

How to use: Apply your pipeline from the Public tier. Use consistency as self-check.

What to submit: Reconstructed signals and corrected spec. Scored server-side.

Dev Leaderboard

#	Method	Score	PSNR	SSIM
1	SwinEELS + gradient	0.767	31.84	0.941
2	TransEELS + gradient	0.752	31.47	0.937
3	PhysEELS + gradient	0.739	30.78	0.928
4	DiffEELS + gradient	0.739	30.29	0.921
5	N2V-EELS + gradient	0.612	23.43	0.748
6	DnCNN-EELS + gradient	0.517	20.17	0.607
7	PowerLaw-EELS + gradient	0.485	19.02	0.551
8	MLS-EELS + gradient	0.429	17.62	0.481
9	ICA-EELS + gradient	0.412	16.92	0.446

Spec Ranges (3 parameters)

Parameter	Min	Max	Unit
energy_dispersion	-0.0024	0.0036	eV/channel
zero_loss_shift	-0.36	0.54	eV
aberration	-2.4	3.6	%

Submit Reconstruction View Details →

Hidden 3 scenes

Fully blind server-side evaluation — no data download.

What you get & how to use

What you get: No data downloadable. Algorithm runs server-side on hidden measurements.

How to use: Package algorithm as Docker container / Python script. Submit via link.

What to submit: Containerized algorithm accepting y + H, outputting x_hat + corrected spec.

Hidden Leaderboard

#	Method	Score	PSNR	SSIM
1	SwinEELS + gradient	0.745	31.06	0.932
2	PhysEELS + gradient	0.722	28.85	0.898
3	TransEELS + gradient	0.694	28.56	0.892
4	DiffEELS + gradient	0.654	26.34	0.841
5	N2V-EELS + gradient	0.563	22.1	0.694
6	PowerLaw-EELS + gradient	0.474	18.73	0.536
7	DnCNN-EELS + gradient	0.448	18.4	0.52
8	MLS-EELS + gradient	0.372	16.0	0.401
9	ICA-EELS + gradient	0.310	13.77	0.3

Spec Ranges (3 parameters)

Parameter	Min	Max	Unit
energy_dispersion	-0.0014	0.0046	eV/channel
zero_loss_shift	-0.21	0.69	eV
aberration	-1.4	4.6	%

Submit Algorithm View Details →

Blind Reconstruction Challenge

Challenge

Given measurements with unknown mismatch and spec ranges (not exact params), reconstruct the original signal. A method must be evaluated on all three tiers for a complete score. Scored on a composite metric: 0.4 × PSNR_norm + 0.4 × SSIM + 0.2 × (1 − ‖y − Ĥx̂‖/‖y‖).

Input

Measurements y, ideal forward model H, spec ranges

Output

Reconstructed signal x̂

About the Imaging Modality

STEM-EELS measures the energy distribution of electrons transmitted through a thin specimen, where inelastic scattering events encode information about elemental composition, bonding, and electronic structure. The energy loss spectrum contains core-loss edges (characteristic of specific elements) and low-loss features (plasmons, band gaps). A magnetic prism spectrometer disperses the energy spectrum onto a position-sensitive detector. Spectrum imaging acquires a full spectrum at each scan position, enabling elemental mapping with atomic-scale spatial resolution.

Principle

Electron Energy Loss Spectroscopy measures the energy lost by transmitted electrons due to inelastic interactions with the specimen. The energy-loss spectrum contains characteristic edges corresponding to inner-shell ionization of specific elements, enabling elemental mapping with atomic spatial resolution. Near-edge fine structure (ELNES) reveals chemical bonding, and low-loss features probe band structure and optical properties.

How to Build the System

Attach a post-column energy filter (Gatan GIF Quantum/Continuum) to a TEM/STEM. For STEM-EELS spectrum imaging: scan the probe and record a full energy-loss spectrum (0-2000 eV range) at each pixel. Use a monochromated source (ΔE < 0.3 eV) for near-edge fine structure studies. Energy dispersion is typically 0.1-0.5 eV/channel. Acquire both core-loss edges (elemental maps) and low-loss region (thickness mapping, optical properties).

Common Reconstruction Algorithms

Background subtraction (power-law fitting before edge onset)
Multiple linear least-squares (MLLS) fitting for overlapping edges
Principal component analysis (PCA) for denoising spectrum images
Kramers-Kronig analysis for optical constants from low-loss EELS
Deep-learning EELS denoising and quantification

Common Mistakes

Specimen too thick causing plural scattering that distorts edge shapes
Incorrect background model for edge extraction (wrong fitting window)
Energy drift during long spectrum-image acquisitions
Not accounting for plural scattering when quantifying elemental ratios
Beam damage altering the specimen chemistry during EELS acquisition

How to Avoid Mistakes

Keep specimen thickness < 0.5 inelastic mean free path (t/λ < 0.5)
Fit background in a window just before the edge; use multiple-window methods if needed
Apply energy drift correction using the zero-loss peak or a known edge
Deconvolve plural scattering using Fourier-log method before quantification
Use low-dose protocols and fast spectrum imaging to minimize beam damage

Forward-Model Mismatch Cases

The widefield fallback produces a 2D spatial image, but EELS acquires energy-loss spectra at each probe position — the spectral dimension encoding elemental composition (core-loss edges) and electronic structure (near-edge fine structure) is entirely absent
Each EELS spectrum contains characteristic ionization edges (e.g., C-K at 284 eV, O-K at 532 eV) that identify elements with atomic spatial resolution — the widefield spatial blur cannot access spectroscopic chemical information

How to Correct the Mismatch

Use the EELS operator that models energy-loss spectrum formation: each probe position produces a spectrum with background (power-law), core-loss edges (proportional to elemental concentration), and near-edge fine structure (bonding information)
Quantify elemental maps using background subtraction and edge integration, or MLLS fitting for overlapping edges; apply PCA denoising to spectrum images before quantification

Experimental Setup — Signal Chain

Experimental setup diagram for Electron Energy Loss Spectroscopy

Reconstruction Gallery — 4 Scenes × 3 Scenarios

Method: CPU_baseline | Mismatch: nominal (nominal=True, perturbed=False)

I Ideal II Mismatch (uncorrected) III Oracle correction

I — Ideal

Ground Truth

Measurement

Reconstruction

23.51322858456625 dB SSIM 0.3680334261076263

II — Mismatch

Ground Truth

Measurement

Reconstruction

18.321872109157592 dB SSIM 0.10385815436617447

III — Oracle

Ground Truth

Measurement (perturbed)

Reconstruction

20.03444592103245 dB SSIM 0.1989493017944018

I — Ideal

Ground Truth

Measurement

Reconstruction

20.039820761387634 dB SSIM 0.29166542696823705

II — Mismatch

Ground Truth

Measurement

Reconstruction

18.85922982415404 dB SSIM 0.12798625129650684

III — Oracle

Ground Truth

Measurement (perturbed)

Reconstruction

20.031642094176974 dB SSIM 0.19752602173448866

I — Ideal

Ground Truth

Measurement

Reconstruction

23.283012767044198 dB SSIM 0.357495155445897

II — Mismatch

Ground Truth

Measurement

Reconstruction

18.4465452279848 dB SSIM 0.09369496983474383

III — Oracle

Ground Truth

Measurement (perturbed)

Reconstruction

20.2796683331393 dB SSIM 0.18597290071365627

I — Ideal

Ground Truth

Measurement

Reconstruction

23.31589368033676 dB SSIM 0.36687406178777526

II — Mismatch

Ground Truth

Measurement

Reconstruction

18.405990402819242 dB SSIM 0.10758808193384366

III — Oracle

Ground Truth

Measurement (perturbed)

Reconstruction

20.100674151892584 dB SSIM 0.20008535528637864

Mean PSNR Across All Scenes

22.53798894833371 dB

I — Ideal

18.50840939102892 dB

II — Mismatch

20.111607625060326 dB

III — Oracle

Degradation: -4.0 dB Recovery: +1.6 dB

Per-scene PSNR breakdown (4 scenes)

Scene	I (PSNR)	I (SSIM)	II (PSNR)	II (SSIM)	III (PSNR)	III (SSIM)
scene_00	23.51322858456625	0.3680334261076263	18.321872109157592	0.10385815436617447	20.03444592103245	0.1989493017944018
scene_01	20.039820761387634	0.29166542696823705	18.85922982415404	0.12798625129650684	20.031642094176974	0.19752602173448866
scene_02	23.283012767044198	0.357495155445897	18.4465452279848	0.09369496983474383	20.2796683331393	0.18597290071365627
scene_03	23.31589368033676	0.36687406178777526	18.405990402819242	0.10758808193384366	20.100674151892584	0.20008535528637864
Mean	22.53798894833371	0.3460170175773839	18.50840939102892	0.10828186435781721	20.111607625060326	0.19563339488223133

Experimental Setup

Instrument: Gatan Quantum GIF / Gatan Continuum / Nion HERMES

Accelerating Voltage Kv: 100

Energy Resolution Ev: 0.3

Dispersion Ev Per Ch: 0.1

Collection Angle Mrad: 30

Dwell Time Ms: 50

Spectrum Range: core-loss + low-loss

Analysis: elemental mapping, ELNES fine structure

Key References

Egerton, 'Electron Energy-Loss Spectroscopy in the Electron Microscope', Springer (2011)

Canonical Datasets

EELS Atlas (Ahn & Krivanek)

Spec DAG — Forward Model Pipeline

P(e⁻) → Λ(energy) → D(g, η₁)

P Electron Beam (e⁻)

Λ Energy Disperser (energy)

D CCD Spectrometer (g, η₁)

Mismatch Parameters

Symbol	Parameter	Description	Perturbed
ΔD_E	energy_dispersion	Energy dispersion error (eV/channel)	0.002
ΔE_0	zero_loss_shift	Zero-loss peak shift (eV)	0.3
ΔC_c	aberration	Chromatic aberration error (%)	2.0

Credits System

40%

Platform Profit Pool

Revenue allocated to benchmark rewards

30%

Winner Share

Top algorithm receives from pool

$100

Min Withdrawal

Minimum payout threshold

Spec Primitives Reference (11 primitives)

P Propagation

Free-space or medium propagation kernel (Fresnel, Rayleigh-Sommerfeld).

M Mask / Modulation

Spatial or spatio-temporal amplitude modulation (coded aperture, SLM pattern).

Π Projection

Geometric projection operator (Radon transform, fan-beam, cone-beam).

F Fourier Sampling

Sampling in the Fourier / k-space domain (MRI, ptychography).

C Convolution

Shift-invariant convolution with a point-spread function (PSF).

Σ Summation / Integration

Summation along a physical dimension (spectral, temporal, angular).

D Detector

Sensor readout with gain g and noise model η (Gaussian, Poisson, mixed).

S Structured Illumination

Patterned illumination (block, Hadamard, random) applied to the scene.

W Wavelength Dispersion

Spectral dispersion element (prism, grating) with shift α and aperture a.

R Rotation / Motion

Sample or gantry rotation (CT, electron tomography).

Λ Wavelength Selection

Spectral filter or monochromator selecting a wavelength band.