Crystal Generation using the Fully Differentiable Pipeline and Latent Space Optimization

2026-01-08

AI Summary

This work proposes a fully differentiable generative framework for efficiently producing crystal structures that satisfy prescribed local chemical environments and crystallographic constraints. By integrating a symmetry-conditioned variational autoencoder (CVAE) with an SO(3) power spectrum–based objective function, the method alternates optimization between real space and latent space representations of crystals. To mitigate gradient conflicts arising from multiple objectives and overcome local energy barriers, a two-level relaxation optimization strategy is introduced. The approach naturally extends to multi-component systems with diverse local environments, achieving approximately fivefold acceleration over conventional CPU-based workflows while maintaining high generation fidelity. This significantly enhances the success rate in generating complex crystal structures.

Technology Category

• Machine Learning: Deep Generative Models & Autoencoders
• Search and Optimization: Sampling/Simulation-based Search
• Computer Vision: Generative Adversarial Networks (GANs) for Vision

Application Category

• Economics, Online Markets and Human Computation: Economic ramifications for generative AI infrastructure and applications
• Web Mining and Content Analysis: Web data generation and simulation
• Graph Algorithms and Modeling for the Web: Representation, reconstruction, and subgraph or motif discovery in Web-related graphs

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.

Problem

• crystal generation
• local environment
• crystallographic constraints
• materials design
• structure optimization

Innovation

• differentiable pipeline
• latent space optimization
• symmetry-conditioned VAE
• SO(3) power spectrum
• dual-level relaxation