Abstract
Uranyl peroxide clusters, formed from uranyl ions, peroxide and other ligands, exhibit complex structural diversity and have significant potential in nuclear chemistry applications. Theoretical prediction is essential for atomic-level understanding of their formation and stability, offering insights where experiments face challenges from uranium's radioactivity and complex coordination. However, their large size and varied topologies pose challenges for theoretical study. In this work, we introduce Uni-Gen, a deep learning model based on the Uni-Mol framework, which generates uranyl cluster structures by leveraging topological similarities with carbon clusters, particularly fullerene. This newly-developed Uni-Gen not only validates the stability of the U<sub>28</sub> uranyl cluster consistent with previous experimental data, but also discovers a more stable U<sub>44</sub> uranyl cluster than the previously reported structure, and most notably, predicts the previously unreported isomers of the U<sub>38</sub> uranyl cluster. Our results demonstrate that Uni-Gen is a powerful tool for predicting diverse potentially stable uranyl cluster structures, thereby enriching our understanding of the stability landscape of uranyl clusters and offering potential for further exploration of novel uranyl materials and chemical phenomena.
Introduction
Uranyl peroxide clusters are "molecular cages" formed from uranium ions and peroxide ligands. These structures are essential for nuclear chemistry due to their ability to encapsulate ions, making them ideal for purifying contaminated water or optimizing nuclear fuel cycles. However, studying these clusters is exceptionally difficult. Experimental work is hampered by the radioactivity and chemical complexity of uranium, while traditional computational methods struggle with the astronomical number of possible atomic arrangements in these massive clusters. Brute-force search algorithms are often prohibitively slow and frequently miss the most stable structures.
To address these challenges, researchers from Tsinghua University have developed Uni-Gen, a generative deep learning framework. Uni-Gen bridges the gap between complex molecular design and computational efficiency by leveraging the structural "blueprints" of carbon fullerenes to predict and discover new uranyl peroxide clusters.
Methodology
Uni-Gen is built upon the Uni-Mol framework, utilizing an SE(3)-equivariant transformer. This neural network architecture is designed to understand 3D space, ensuring that fundamental molecular properties remain consistent regardless of rotation or translation. The model's training followed a two-stage process:
- Pre-training: The model learned the underlying physics of uranyl clusters through a 3D position denoising task. It was trained on a dataset of 64,000 structure coordination points generated via ab initio molecular dynamics (AIMD) across 16 types of stable clusters.
- Structural Fine-tuning: Recognizing that many uranyl clusters share topologies with carbon fullerenes (such as Buckyballs), the researchers used fullerene skeletons as "prompts." The model learned to map these carbon templates onto uranium skeletons, using the well-studied world of carbon chemistry to navigate the actinide landscape.
Key Insights and Results
The core innovation of Uni-Gen is Topological Mimicry. By treating carbon fullerenes as a structural cheat sheet, the model bypasses the need for exhaustive random searches. This approach allowed the model to effectively "hallucinate" stable new isomers that had previously eluded researchers.
Breakthrough Discoveries
- The $U_{44}$ Isomer: Uni-Gen discovered a new isomer of the $U_{44}$ cluster with $D_2$ symmetry that is 0.83 eV more stable than the bispherical structure currently known to science.
- Predicting the $U_{38}$ Cluster: The model predicted 17 stable isomers for the $U_{38}$ cluster, a molecule that has not yet been synthesized or reported in existing literature.
- Superior Efficiency: In benchmarks against traditional tools like RDKit and ABCluster, Uni-Gen achieved higher Coverage (COV) and Matching (MAT) accuracy while significantly reducing the average optimization time.
Conclusion and Future Outlook
Uni-Gen represents a significant leap in computational actinide chemistry. By incorporating chemical intuition—such as coordination environments and spherical symmetry—into a deep learning framework, the researchers have created a powerful tool for discovering novel uranyl materials.
While the newly predicted $U_{38}$ and $U_{44}$ isomers still require experimental synthesis for final validation, the success of Uni-Gen opens new doors. Future iterations may explore non-spherical arrangements, such as open-ring or cup-shaped clusters, and extend the framework to other heavy-metal systems across the bottom row of the periodic table.
Summary updated