Title: Leveraging Latent Diffusion Pretraining to Enhance Crystal Property Prediction

Abstract:

Developing new materials relies heavily on the ability to rapidly and precisely predict crystal properties, a persistent hurdle in the field. While graph neural networks and Transformer architectures have become prominent solutions—thanks to their capacity to represent the local atomic environments within crystals—they suffer from a significant dependency on large volumes of data. In reality, labeled datasets for crystal properties are often limited. To overcome this bottleneck, pretraining-finetuning methodologies, especially those utilizing diffusion models, have demonstrated considerable potential.

This study presents CrysLDNet, a new pretraining framework grounded in latent diffusion, specifically engineered to alleviate issues related to data scarcity. The methodology combines a Variational Autoencoder (VAE) with a diffusion model during the initial pretraining phase. Specifically, the VAE encoder transforms three-dimensional crystal structures into a continuous latent space, where the diffusion mechanism is subsequently executed. By employing latent diffusion for pretraining, the graph encoder can efficiently extract structural and chemical insights from extensive unlabeled datasets. These learned features are then fine-tuned to address specific property prediction objectives.

Extensive testing on widely used DFT datasets for property prediction demonstrates that CrysLDNet achieves superior performance compared to both models trained from scratch and existing pretrained baselines. The framework yielded accuracy improvements of 4.26% on the JARVIS dataset and 4.90% on the MP dataset. Furthermore, the representations acquired through this method prove resilient under sparse-data scenarios. They possess sufficient expressiveness to rectify DFT inaccuracies when fine-tuned using small sets of experimental data.

The source code for this project is accessible at: https://github.com/shrimonmuke0202/CrysLDNet.git.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC