Are diffusion models ready for materials discovery in unexplored chemical space?

Area of Science:

• Materials Science
• Computational Chemistry
• Artificial Intelligence

Background:

• Diffusion models are emerging tools for materials discovery.
• Their performance in generating novel, low-energy structures across diverse chemical spaces requires systematic evaluation.

Purpose of the Study:

• To assess the efficacy of diffusion models (MatterGen, DiffCSP) in materials discovery.
• To evaluate their performance on well-sampled (oxides, nitrides) and uncommon (GNoME) chemical spaces.
• To investigate their size-extrapolation capabilities.

Main Methods:

• Evaluation of MatterGen and DiffCSP against three distinct material databases.
• Analysis of model performance across different chemical spaces and varying system sizes.
• Identification of limitations related to periodic boundary conditions.

Main Results:

• Diffusion models perform reliably in well-sampled chemical spaces like oxides and nitrides.
• Performance degrades significantly in uncommon chemical spaces, such as the GNoME database with rare-earth elements.
• A notable decrease in performance occurs when extrapolating to larger system sizes beyond the training data range.

Conclusions:

• Diffusion models are effective for materials discovery in familiar chemical environments.
• Challenges remain in generating stable structures in unexplored chemical spaces and for large systems.
• The 'curse of periodicity' limits the size-extrapolation ability of current diffusion models.