DiffSyn: a generative diffusion approach to materials synthesis planning

Area of Science:

• Materials Science
• Chemical Engineering
• Computational Chemistry

Background:

• Synthesizing crystalline materials like zeolites is challenging due to complex structure-synthesis relationships and vast experimental spaces.
• Existing methods struggle to navigate the high-dimensional synthesis landscape and predict optimal conditions efficiently.

Purpose of the Study:

• To develop an advanced computational model, DiffSyn, for predicting and optimizing zeolite synthesis pathways.
• To leverage a large dataset of synthesis recipes to train a generative diffusion model.
• To demonstrate the model's capability in discovering and synthesizing novel crystalline materials.

Main Methods:

• Developed DiffSyn, a generative diffusion model trained on over 23,000 zeolite synthesis recipes from 50 years of literature.
• Conditioned the model on desired zeolite structures and organic templates to generate probable synthesis routes.
• Utilized density functional theory (DFT) for rationalizing predicted synthesis routes via binding energy calculations.

Main Results:

• DiffSyn achieved state-of-the-art performance by effectively modeling the multi-modal structure-synthesis relationships.
• The model successfully differentiated between competing phases and proposed optimal synthesis routes.
• A UFI material was synthesized using DiffSyn-generated routes, yielding a high Si/Al ratio of 19.0, indicating enhanced thermal stability.

Conclusions:

• DiffSyn offers a powerful computational approach to accelerate the discovery and synthesis of crystalline materials.
• The model's ability to predict synthesis routes based on desired structures and templates significantly reduces experimental trial and error.
• Successful synthesis of the UFI material validates DiffSyn's predictive accuracy and potential for advancing materials science.