Crystal Generation using the Fully Differentiable Pipeline and Latent Space Optimization

Osman Goni Ridwan, Gilles Frapper, Hongfei Xue, Qiang Zhu · 2026

Read Crystal Generation using the Fully Differentiable Pipeline and Latent Space Optimization by Osman Goni Ridwan, Gilles Frapper, Hongfei Xue, Qiang Zhu (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2601-04606.txt.

Introduction

Osman Goni Ridwan and three co-authors generate new crystals through a fully differentiable pipeline, in a 2026 arXiv paper. Designing a material means finding an atomic arrangement whose local setting gives the wanted traits, subject to crystallographic symmetry. Their system couples a symmetry-conditioned variational autoencoder, which proposes trial crystals, with an objective built from the SO(3) power spectrum that pulls trial crystals toward a specified local setting.

Because the whole chain is differentiable, trial crystals can be optimized in batches on graphics chips, running about five times faster than the team's earlier workflow with similar results. A dual-level strategy alternates optimization between the raw crystal representation and the autoencoder's latent space, which escapes the local minima that trap either view alone. The authors note the system extends to multi-component systems and multiple target settings.

Catalog record

CREATOROsman Goni Ridwan, Gilles Frapper, Hongfei Xue, Qiang Zhu
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE35 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2601-04606

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Crystal Generation using the Fully Differentiable Pipeline and Latent Space Optimization Osman Goni Ridwan, Gilles Frapper, Hongfei Xue, Qiang Zhu Abstract We present a materials generation framework that couples a symmetry-conditioned variational autoencoder (CVAE) with a differentiable SO(3) power spectrum objective to steer candidates toward a specified local environment under the crystallographic constraints. In particular, we implement a fully differentiable pipeline to enable batch-wise optimization on both direct and latent crystallographic representations. Using the GPU acceleration, this implementation achieves about fivefold speed compared to our previous CPU workflow, while yielding comparable outcomes. In addition, we introduce the optimization strategy that alternatively performs optimization on the direct and latent crystal representations. This dual-level relaxation approach can effectively escape local minima defined by different objective gradients, thus increasing the success rate of generating complex structures satisfying the target local environments. This framework can be extended to systems consisting of multi-components and multi-environments, providing a scalable route to generate material structures with the target local environment. I Introduction To enable rapid materials discovery prior to synthesis, it is pivotal to have a reliable and efficient crystal structure prediction (CSP) method. In the past, global optimization strategies have achieved major successes 22 . However, they are mostly limited to small unit cells. For large unit cells and high-throughput exploration of a large compositional space, the computational cost becomes prohibitive. Since 2018, machine-learning (ML) and artificial intelligence (AI) approaches have been increasingly applied to the CSP field. These models offer a fundamentally new approach to exploring structure space by learning data-driven representations, allowing for the efficient generation of low-energy crystal structures at a much faster rate than traditional global optimization methods. Despite the ongoing criticisms 7 , the AI-based CSP approaches have been widely used to accelerate the discovery of new materials such as magnets, ferroelectrics, and thermoelectric materials 20 ; 34 . Recently, we introduced a symmetry-informed approach called Local Environment Geometry-Oriented Crystal Generator ( LEGO-xtal ) 27 to rapidly generate low-energy crystals with a target motif, using sp 2 carbon allotropes as the benchmark example. Building on this foundation, we report the further implementation of a GPU-enabled differentiable framework to explore alternative optimization strategies at both direct and latent structural representation space. In the practical application, we demonstrate that the dual optimization strategy can notably improve the success rate of generating low-energy new sp 2 carbon allotropes. This implementation is expected to pave the way for accelerated discovery of novel materials with targeted local coordination environments. II Related Work II.1 Global optimization based CSP methods Most CSP algorithms cast structure discovery as a global optimization problem on a rugged potential-energy surface. Representative methods include evolutionary algorithms 19 ; 17 ; 26 , particle-swarm and swarm-intelligence searches 28 , random structure searching 25 , basin-hopping 2 , and related metadynamics and genetic search variants 38 ; 37 . While these strategies have enabled many high-impact discoveries 22 , their practical throughput is commonly dominated by repeated local relaxations and extensive samplings. This becomes increasingly challenging as the number of degrees of freedom grows (e.g. unit cells with tens to hundreds of atoms, multiple Wyckoff sites, and broad symmetry/composition search spaces), motivating complementary data-driven generation and accelerated screening workflows. II.2 AI generative models for crystals Deep generative modeling has become popular for crystal generation, spanning Variational Auto Encoder (VAE)-based latent-variable models 15 ; 30 , Generative Adversarial Network (GAN)-based formulations 14 , diffusion models 12 , and autoregressive generators 1 . These methods offer different trade-offs between sampling cost, controllability, and output fidelity, and they form the basis for modern ML-driven CSP pipelines. More recently, incorporating crystallographic priors is widely recognized as essential for validity and control, including space-group–conditioned diffusion 13 , Wyckoff-based generators 39 , and symmetry-informed transformer architectures 6 . II.3 Machine learning for crystal screening and relaxation After the generation of candidates, high-throughput workflows typically rely on fast screening models and machine-learned interatomic potentials (MLIPs) to reduce the number of structures that require expensive first-principles evaluation. For reference, the final validation of stability is commonly carried out using electronic-structure codes such as VASP 16 and Quantum Espresso 8 . In addition, MLIPs based on message passing and E(3)-equivariant architectures provide accelerated geometry relaxation and energy ranking across large candidate pools, improving practical throughput for screening and database construction 5 ; 32 ; 29 . Other than energy-based optimization, we recently proposed the concept of “pre-relaxation" for a rapid pre-process of massive AI-generated structures before the use of MLIPs or DFT energy-based optimizations 27 . III Method III.1 Dataset Preparation and Augmentation Following our previous work on LEGO-xtal 27 , we focus on developing a framework to harness sp 2 carbon allotropes. To begin, we construct the training set from the SACADA database 11 , which contains 154 experimentally known or hypothetical sp 2 -bonded carbon allotropes. Here we adopt a compact tabular representation that exploits space-group symmetry to reduce dimensionality while remaining fully reconstructible. Each structure is encoded as a fixed-length vector comprising: (i) the space-group number ( spg ); (ii) the six lattice parameters (a,b,c,\alpha,\beta,\gamma) ; and (iii) Wyckoff-site positions (WPs), each specified by a Wyckoff position index and the fractional coordinates of the corresponding representative atom in the asymmetric unit. By default, crystal structures are presented in the highest possible symmetry group. To increase diversity while preserving crystallographic validity, we applied subgroup augmentation 36 : for each parent structure, we enumerate possible subgroup combinations and the symmetry operations of the Wyckoff sites. This procedure yields 63 115 structures spanning a broad range of symmetry combinations. III.2 Conditional VAE In the LEGO-xtal approach 27 , we originally employed the standard VAE model to handle both continuous variables (lattice parameters and fractional coordinates) and discrete variables (space group and Wyckoff indices). In this work, we instead adopt a two-stage approach: in the first stage, we have trained the VAE or GAN 9 model on the discrete part of the dataset (space group and Wyckoff indices) to generate valid symmetry combinations. The target of the first stage model is to learn the distribution of valid (spg, wps) combinations from the training set and generate new combinations that will be used as conditions for the second stage model. In the second stage, we train a Conditional Variational Autoencoder (CVAE) 10 architecture, in which the generative process is explicitly conditioned on discrete crystallographic symmetry information. This design enables targeted exploration and optimization of the latent space while keeping the discrete symmetry fixed, allowing the continuous degrees of freedom (cell parameters and Wyckoff fractional coordinates) to adapt toward a desired local environment. The continuous crystallographic features X are transformed using a Gaussian Mixture Model (GMM) 31 cluster-based normalization scheme. Each scalar variable is encoded as a K -component categorical assignment together with a standardized continuous feature, yielding a transformed representation X^{\prime} . This GMM-based encoding helps to handle the strong multi-modality in crystallographic parameters and makes the representation smoother. The discrete symmetry information, consisting of the space-group index and Wyckoff site labels, is converted into a one-hot condition vector C . Here, X denotes ground-truth input features, while \tilde{X} denotes the corresponding features generated by the model. The encoder network takes X^{\prime} as input and outputs

Read the complete text