🔬 materials science

Fourier Neural Operators for Composition-Driven Crystal Structure Discovery

This paper introduces a scalable framework for crystal structure discovery that combines a Fourier Neural Operator-based solver, capable of capturing long-range periodic correlations, with a conditional variational autoencoder to generate diverse and valid crystal structures from prescribed chemical compositions.

Original authors: Zhijie Yu, Jingyu Li, Yang Huang, Jingrun Chen

Published 2026-09-02

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Zhijie Yu, Jingyu Li, Yang Huang, Jingrun Chen

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 search for new materials is a quest as old as civilization itself, driven by the need to store energy more efficiently, build faster electronics, and create catalysts that can transform our industrial world. At the heart of this search lies the crystal, a solid where atoms are arranged in a repeating, three-dimensional pattern. The specific way these atoms stack determines whether a material is a hard diamond, a soft metal, or a superconductor. However, the universe of possible crystal structures is so vast that trying to find a new one by testing every combination of elements is like searching for a specific grain of sand on every beach on Earth. Traditional computer methods attempt to solve this by simulating the energy of millions of candidate structures, but this process is so slow and expensive that it can only explore a tiny fraction of the possibilities.

In recent years, scientists have turned to artificial intelligence to navigate this chemical wilderness. These new models learn from existing databases of known crystals, learning the hidden rules that govern how atoms arrange themselves. They can then propose entirely new structures that have never been seen before. Yet, even these advanced AI systems face a significant hurdle: they often struggle to capture the long-range connections that define a crystal's repeating nature, and when they try to generate complex 3D images of atoms, they frequently produce blurry or physically impossible results. A team of researchers has now developed a new approach that bypasses these limitations, offering a faster and more reliable way to discover stable crystals from a simple list of ingredients.

The researchers, working at the University of Science and Technology of China, created a system called CrystalFNO. Instead of trying to guess the exact position of every atom in a single, massive leap, they broke the problem into two distinct steps. First, the system takes a chemical formula—the recipe of elements, such as calcium, palladium, and oxygen—and generates a set of candidate shapes for the crystal's container, known as the lattice parameters. Think of this as determining the size and angles of the box that will hold the atoms. To do this, they used a type of neural network trained to understand the relationship between a chemical recipe and the geometric box it usually fits into.

Once a candidate box is proposed, the system moves to its second, more innovative step: predicting how the atoms will fill that space. Here, the researchers replaced the standard tools used in previous AI models with a new type of solver based on Fourier neural operators. While traditional methods look at a crystal by examining small, local neighborhoods of atoms, this new solver looks at the entire structure at once, analyzing the repeating patterns across the whole grid. This allows it to capture the long-range connections that are essential for a stable crystal. The solver predicts two invisible fields: one that shows where the atoms are likely to be, and another that indicates what kind of atoms they are. These fields are smooth, continuous maps rather than a list of coordinates, which helps the AI avoid the confusion that often plagues other models when dealing with complex, repeating patterns.

With these predicted fields in hand, the system then reconstructs the actual crystal structure. It identifies the peaks in the density maps to place the atoms, optimizes their positions to ensure they fit perfectly within the predicted shape, and assigns the correct element types based on the second field. This process turns the smooth, abstract predictions back into a concrete 3D arrangement of atoms. To ensure these new structures are not just mathematical curiosities but physically real possibilities, the team subjected them to a rigorous screening process. They first filtered out obviously flawed structures, then used a machine-learning model to relax the atoms, allowing them to settle into their most stable, low-energy positions. Finally, they performed high-level calculations to check if the resulting crystals would hold together or fall apart.

The results of this new approach were tested against a vast database of known materials. The system successfully generated novel structures for over one hundred different chemical formulas. In a significant test, the researchers found that the system could produce stable, geometrically reasonable crystals that did not exist in any known database. Out of thousands of candidates generated, hundreds passed the strict stability checks, including some with complex combinations of elements like silver, lanthanum, and silicon. The study showed that the new method is particularly effective at handling high-symmetry crystals, where the repeating patterns are strong and clear, though it still faces challenges with more irregular, low-symmetry structures where the atomic arrangements are less predictable.

Crucially, the researchers identified a fundamental difficulty in their task: predicting a crystal structure from a chemical formula is not a simple one-to-one puzzle. The same set of ingredients and box size can sometimes correspond to multiple different atomic arrangements, a phenomenon known as polymorphism. Because their system is deterministic, it learns to predict the most typical arrangement it has seen during training. When faced with a rare or unusual arrangement, it may produce a result that is close but not perfect. The team acknowledged this limitation, noting that the system's success depends heavily on the symmetry of the crystal; the more symmetrical the structure, the more accurately the system can predict the density fields and reconstruct the final material.

Despite these challenges, the work represents a significant step forward in the field of materials discovery. By separating the generation of the crystal's container from the prediction of its contents, and by using a solver that understands global patterns rather than just local details, the researchers have created a scalable path to finding new materials. The system does not just mimic existing crystals; it explores the vast space of possibilities to find stable, novel structures that could one day power new technologies. The study confirms that while the task is complex and the mapping from recipe to structure is not always unique, coupling advanced neural operators with careful physical screening offers a powerful new tool for scientists to discover the materials of the future.

Technical Summary: Fourier Neural Operators for Composition-Driven Crystal Structure Discovery

1. Problem Statement

The discovery of novel crystalline materials is critical for advancements in energy, electronics, and catalysis. However, the vast combinatorial space of chemical elements, stoichiometries, and crystal structures renders exhaustive experimental or computational screening infeasible. Traditional Crystal Structure Prediction (CSP) methods, such as genetic algorithms and particle swarm optimization, rely on iterative Density Functional Theory (DFT) evaluations, which are computationally expensive and scale poorly with system size.

While deep generative models offer a promising alternative by learning the underlying distribution of crystal structures, existing voxel-based approaches face two fundamental limitations:

Local Receptive Fields: Standard 3D Convolutional Neural Networks (CNNs) struggle to capture long-range periodic dependencies inherent in crystal density fields, as they primarily aggregate information from local neighborhoods.
Posterior Collapse: The high dimensionality of voxel representations (e.g., $N \times N \times N$ grids) causes the reconstruction term to dominate Variational Autoencoder (VAE) objectives, leading to posterior collapse where the latent space fails to align with the prior, resulting in low-quality and insufficiently diverse generations.

2. Methodology: CrystalFNO

The authors propose CrystalFNO, a coupled generation-solving framework that decouples lattice parameter generation from crystal density-field prediction. The workflow consists of four sequential stages:

A. Voxel Representation

Crystal structures are represented as continuous density fields on a $32 \times 32 \times 32$ grid under periodic boundary conditions. Two fields are defined:

Number Density Field ( $\rho_{num}$ ): Encodes the spatial distribution of atoms independent of element type, used for atom localization.
Electron Density Field ( $\rho_{Z}$ ): Encodes both positional and elemental information (weighted by atomic number), used for element assignment.

B. Lattice Parameter Generation (CVAE)

A Conditional Variational Autoencoder (CVAE) generates candidate lattice parameters ( $L = \{a, b, c, \alpha, \beta, \gamma\}$ ) conditioned on a chemical formula ( $F$ ).

Symmetry Constraints: To ensure positive definiteness and consistency with crystallographic symmetries, the lattice parameters are encoded into a six-dimensional coefficient vector using a symmetric basis decomposition strategy.
Generation: The CVAE samples from a low-dimensional latent space to produce diverse lattice parameters without directly modeling the high-dimensional voxel space.

C. Density-Field Prediction (FNO Solver)

A Fourier Neural Operator (FNO) serves as a crystal-field solver, mapping the chemical formula and generated lattice parameters to the corresponding $\rho_{num}$ and $\rho_{Z}$ fields.

Architecture: The solver employs a dual-branch structure within Fourier layers:
- Frequency Branch: Uses 3D Fast Fourier Transform (FFT), learnable spectral convolutions, and Inverse FFT (IFFT) to capture global periodic correlations and long-range dependencies.
- Spatial Branch: Uses 3D convolutions to extract local atomic-scale features.
Input Encoding: Chemical formulas are encoded via weighted one-hot vectors, while lattice parameters undergo Fourier feature encoding to capture high-frequency variations.
Training: Two independent FNO models are trained (one for each density field) using L1 loss.

D. Structure Reconstruction and Screening

Predicted density fields are converted back to atomic structures ( $M = \{A, X, L\}$ ) via a three-stage optimization:

Peak Detection: Identifies candidate atomic positions as local maxima of $\rho_{num}$ using non-maximum suppression.
Position Optimization: Refines atomic coordinates via gradient descent to minimize the discrepancy between reconstructed and predicted $\rho_{num}$ .
Weight Optimization: Assigns element types by optimizing weights to fit $\rho_{Z}$ .

Screening Pipeline:

Voxel-Level Filtering: Candidates are ranked by the L1 loss between predicted and reconstructed density fields.
DPA Relaxation: Top candidates undergo relaxation using the Deep Potential Atomistic (DPA) machine-learning interatomic potential to optimize geometry and forces.
Validation: Structures are screened based on unit cell variation, RMSD, and element match rate. Final candidates are evaluated for formation energy.

3. Key Results

A. Density Field Prediction Accuracy

The FNO solver achieved stable convergence with low validation losses ( $\approx 1.5 \times 10^{-5}$ ).
Prediction accuracy correlates with crystal symmetry: high-symmetry systems (cubic) showed the lowest errors ( $\approx 6.15 \times 10^{-6}$ ), while low-symmetry systems (triclinic) exhibited higher errors ( $\approx 2.36 \times 10^{-5}$ ).
Limitation Analysis: The authors identify the density field prediction task as an ill-posed inverse problem. Identical inputs (formula + lattice) can correspond to multiple distinct density fields due to global translation, discrete symmetry transformations, and polymorphism. As a deterministic model, FNO learns a "typical" mapping, leading to systematic deviations in cases that deviate from the training distribution.

B. Structure Reconstruction

Reconstruction accuracy varies by crystal system. Cubic systems achieved the best performance (RMSE = 0.247 Å, match rate = 90.5%), while monoclinic systems showed the largest positional errors (RMSE = 1.195 Å).
Under combined criteria (RMSE $\le$ 1.0 Å and element match rate $\ge$ 60%), only 53.3% of structures were successfully recovered, highlighting the difficulty of reconstructing complete structures from density fields alone.

C. End-to-End Generation

The framework was tested on 104 chemical formulas.
From 43,686 generated candidates, 3,663 passed geometric and energetic screening.
After deduplication against the Materials Project database, 320 novel structures were identified, covering 104 distinct chemical formulas.
55.9% of these novel candidates exhibited negative formation energies relative to elemental reference states, indicating energetic favorability.

4. Significance and Claims

The paper claims to present the first application of Fourier Neural Operators as crystal-field solvers within a voxel-based crystal structure generation framework.

Scalability: By separating lattice generation from density prediction and utilizing global spectral convolutions, the approach addresses the limitations of local CNNs and high-dimensional VAEs, providing a scalable route to crystal discovery from prescribed compositions.
Generative Diversity: The coupled generation-solving strategy successfully generates diverse, geometrically reasonable, and energetically stable structures across a wide range of chemical formulas and crystal systems.
Methodological Contribution: The work demonstrates that extending Fourier neural operators to periodic crystal fields, when coupled with composition-conditioned lattice generation, effectively captures long-range correlations and enables the discovery of novel materials without the computational cost of iterative DFT searches for every candidate.

The authors remain modest regarding the ill-posed nature of the inverse problem, noting that the deterministic FNO solver produces a single "typical" mapping, which may limit performance on rare polymorphs or non-standard symmetry operations. They suggest that future improvements could involve enriching input conditions (e.g., center-of-mass position or symmetry information) to better distinguish between multiple valid configurations.