🔬 materials science

Generative Inverse Estimation of 3D Atomic Coordination from Near-Edge Spectra via Equivariant Diffusion Models

This paper introduces an equivariant diffusion model that accurately reconstructs full 3D atomic coordination geometries, including bond angles and precise coordination numbers, directly from near-edge spectra (ELNES/XANES), outperforming traditional EXAFS analysis and template matching while demonstrating robust generalization from crystalline training data to amorphous systems.

Original authors: Ren Okubo, Yu Fujikata, Izumi Takahara, Teruyasu Mizoguchi

Published 2026-02-19

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Ren Okubo, Yu Fujikata, Izumi Takahara, Teruyasu Mizoguchi

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Picture: Solving a "Reverse Puzzle"

Imagine you are a detective trying to figure out what a house looks like inside, but you are only allowed to stand outside and listen to the sound of wind whistling through the windows. You can't see the walls, the furniture, or the layout. You only have the sound (the "spectrum").

For decades, scientists have struggled with this exact problem in materials science. They have powerful tools (like electron microscopes and X-rays) that give them a "sound" or a "shadow" of a material's atoms, but turning that shadow back into a clear 3D picture of where every single atom is located has been incredibly difficult.

This paper introduces a new AI detective that can look at that "sound" (a specific type of light spectrum) and instantly draw a perfect 3D map of the atoms inside.

The Problem: The "Blurry Photo" vs. The "3D Model"

Traditionally, scientists use a method called EXAFS (think of it as a high-tech ruler).

How it works: It measures the distance between atoms very precisely.
The flaw: It's like trying to describe a house by only measuring the distance from the front door to the back wall. You know the length, but you don't know the shape. You don't know if the rooms are square, if the roof is slanted, or how many chairs are in the living room. It's a 1D measurement of a 3D world.
The result: Scientists get good distance numbers but often guess wrong about how many neighbors an atom has (the "coordination number").

The Solution: The "Generative Diffusion Artist"

The authors created a new AI model based on Diffusion Models.

The Analogy: Imagine a sketch that starts as a chaotic cloud of static (like TV snow). The AI acts like a sculptor who slowly chips away the noise, refining the cloud until it becomes a clear, detailed statue.
The Magic: This AI is trained on thousands of known crystal structures. It learns the "language" of how atoms arrange themselves. When you feed it a new spectrum (the "sound"), it doesn't just look up a match in a database; it imagines and builds the 3D structure from scratch, atom by atom.

What Did They Achieve? (The "Wow" Factors)

The team tested this AI on Silicon-Oxygen materials (the stuff in glass, sand, and computer chips). Here is what they found:

It's as precise as a ruler, but sees more:
- The AI predicted the distance between atoms with an error of only 0.06 Angstroms (that's 0.000000006 meters!). This is almost as good as the traditional "ruler" method (EXAFS).
- But unlike the ruler, the AI also figured out the angles and the shape of the structure. It gave them a full 3D model, not just a line.
It counts neighbors perfectly:
- Traditional methods often get the "count" of neighboring atoms wrong by about 20%.
- This AI got it wrong less than 4% of the time. It's like a chef who can taste a soup and tell you exactly how many cloves of garlic were used, whereas the old method would just guess "maybe a few."
The "Crystal to Glass" Superpower:
- The AI was trained only on perfect, ordered crystals (like a neatly stacked brick wall).
- Then, the scientists asked it to predict the structure of amorphous glass (like a pile of sand where everything is jumbled).
- The Result: Even though it had never seen glass before, the AI figured out the local structure of the glass almost as well as it did the crystals. It learned the rules of the game, not just the specific moves.
Real-World Proof:
- They didn't just use computer simulations; they fed it real experimental data from a quartz crystal. The AI successfully reconstructed the real-world structure, proving it works outside the lab simulation.

Why Does This Matter?

Think of this as moving from 2D maps to Google Earth.

Before: Scientists had to guess the 3D shape of new materials based on 1D clues. It was slow, required a lot of human expertise, and was often inaccurate.
Now: This AI can take a quick spectrum scan and instantly generate the full 3D atomic map.

The Impact:

Faster Innovation: We can design better batteries, faster computer chips, and stronger glasses much faster because we can "see" the atomic structure instantly.
No More Guessing: It solves the "counting" problem that has plagued scientists for years.
Universal Application: Since it learned the underlying rules of how atoms bond, it can likely be applied to other materials (like metals or complex drugs) in the future.

Summary in One Sentence

The authors built a "3D imagination engine" for AI that can look at a flat, blurry shadow of a material's atoms and instantly reconstruct the full, detailed 3D house they live in, even if the house is made of messy, jumbled glass.

1. Problem Statement

The extraction of 3D atomic coordinates from spectroscopic data (specifically Electron Energy-Loss Near-Edge Structure [ELNES] and X-ray Absorption Near-Edge Structure [XANES]) is a longstanding inverse problem.

Limitations of Current Methods:
- Imaging (STEM/TEM): Provides 2D projections; 3D reconstruction via tomography is limited by beam damage, tilt range constraints, and artifacts.
- Extended Fine Structure (EXAFS/EXELFS): While it offers high precision in bond distances (~0.01–0.02 Å), it suffers from:
  - Poor coordination number accuracy (~20% uncertainty) due to correlations with Debye-Waller factors.
  - Lack of angular information (bond angles/orientations).
  - Requirement for wide energy ranges, making data acquisition slow and prone to damage.
- Near-Edge Spectra (ELNES/XANES): Contain rich 3D structural information (bond angles, coordination geometry) but are traditionally interpreted only qualitatively or via first-principles calculations, which is low-throughput and expertise-dependent.
Gap in Machine Learning: Existing ML approaches have predicted averaged structural descriptors (e.g., radial distribution functions) but have failed to generate site-specific, full 3D atomic coordinates directly from spectra.

2. Methodology

The authors propose a conditional equivariant diffusion model to generate 3D atomic structures directly from near-edge spectra.

Model Architecture:
- Equivariant Graph Neural Network (EGNN): The core architecture preserves rotational and translational symmetries inherent to atomic structures.
- Conditioning: The diffusion process is conditioned on the input spectrum ( $S$ ). The spectral vector is concatenated to the node features ( $h$ ) of the graph, ensuring the denoising process is guided by spectroscopic data.
- Generative Process: The model starts with Gaussian noise ( $z_T$ ) representing atomic coordinates ( $x$ ) and species ( $h$ ) and progressively refines them through reverse diffusion timesteps to produce a physically realistic 3D structure.
Dataset:
- Training: 2,067 computed O K-edge spectra from 274 Si-O crystal structures (Materials Project). The dataset is dominated by SiO₂ but includes various stoichiometries (SiOx, x=0.5–4.0).
- Target: Local coordination environments extending to the third-nearest neighbors from an excited oxygen site (within a 2.0 Å cutoff).
- Representation: Structures are represented as fully connected graphs where nodes encode Cartesian coordinates, elemental species (one-hot), and the spectral vector.
Evaluation Strategy:
- Alignment: Due to the equivariant nature of the model, the generated graph orientation is arbitrary. The authors use a modified Kabsch algorithm combined with the Hungarian algorithm to align generated structures with ground truth for RMSD calculation.
- Validation Tiers:
  1. Crystalline Test: Quantitative accuracy on held-out crystal structures.
  2. Out-of-Distribution (OOD): Testing on amorphous SiOx (generated via MD) despite training only on crystals.
  3. Experimental Application: Application to experimental O K-edge spectra of $\alpha$ -quartz.

3. Key Contributions

First Site-Specific 3D Generation: Demonstrated the ability to generate explicit 3D atomic coordinates for specific excitation sites directly from spectra, moving beyond averaged descriptors.
Superior Coordination Number Precision: Achieved coordination number errors of **<4.3%** (1.86% for the first shell), representing a >10x improvement over traditional EXAFS (~20% error).
Angular Information Recovery: Successfully reconstructed bond angles and full 3D geometries, a capability absent in 1D radial distribution analysis.
Robust Generalization: The model, trained exclusively on crystalline lattices, successfully generalized to amorphous systems (disordered materials), capturing local motifs without explicit training on disorder.
Outperformance of Template Matching: The generative approach outperformed conventional template matching (database lookup) in 52.1% of crystalline cases and >70% of amorphous cases.

4. Key Results

Radial Accuracy:
- Mean RMSD for the first coordination shell is 0.063 Å (6.3 pm).
- This is comparable to EXAFS fitting precision (0.01–0.02 Å) but achieved de novo without pre-defined structural models.
- Overall mean RMSD (1st–3rd shells) is 0.804 Å.
Coordination Number Accuracy:
- First shell error: 1.86%.
- Second shell error: 4.27%.
- Significantly outperforms EXAFS, which struggles with coordination number determination due to parameter correlation.
Elemental Species Prediction:
- Achieved 94.9% accuracy in predicting the correct elemental species (Si vs. O) for generated structures.
Amorphous Generalization:
- For amorphous SiO₂ (trained on crystals), the model achieved a first-shell RMSD of 0.137 Å and 84.4% species accuracy.
- For SiO₀.₂₅ (highly disordered, far from training distribution), performance dropped (RMSD ~0.778 Å, species accuracy 16.4%), but the model still outperformed template matching by >70%, indicating learned transferable spectral-structural relationships.
Experimental Validation:
- Applied to experimental $\alpha$ -quartz spectra, yielding an RMSD of 0.705 Å. The model correctly reproduced the first and second shell peaks in the Radial Distribution Function (RDF), validating practical applicability.

5. Significance and Impact

Paradigm Shift: Moves materials characterization from "fitting" spectra to pre-defined models or "looking up" spectra in databases, to generating novel 3D structures directly from data.
Complement to EXAFS: Offers a method that matches EXAFS in bond distance precision but vastly exceeds it in coordination number accuracy and provides unique access to bond angles.
Materials Discovery: Enables rapid, automated determination of local structures in complex systems (amorphous, interfaces, defects) where traditional methods fail or are too slow.
Future Directions: The framework is scalable to multi-edge data (e.g., combining O K-edge with Si L-edge), higher-order systems, and operando studies (tracking dynamic structural changes during battery cycling or catalysis).

In conclusion, this work establishes equivariant diffusion models as a powerful tool for solving the inverse problem of spectroscopy, enabling the direct, quantitative, and automated reconstruction of 3D atomic environments from near-edge spectra with accuracy rivaling or exceeding established physical characterization techniques.