Discovering new photovoltaics using optimal transport theory

Area of Science:

• Materials Science
• Computational Chemistry
• Data Science

Background:

• Materials discovery often relies on chemical and structural similarity, but algorithmic implementation faces challenges in defining and balancing these measures.
• Existing methods for materials similarity often focus on either composition or structure, leaving the integration of both aspects open.

Purpose of the Study:

• To develop a novel metric for crystalline materials similarity that effectively balances both compositional and structural information.
• To apply this metric to accelerate the discovery of new photovoltaic materials.

Main Methods:

• Adaptation of graph-based distance calculation methods, specifically the Fused Gromov-Wasserstein (FGW) metric, to assess crystalline materials similarity.
• Utilizing optimal transport theory to map between graph representations of materials, considering both graph topology and node (atomic) information.
• Application of the FGW metric to the Materials Project database for a photovoltaic materials discovery campaign.

Main Results:

• The FGW metric demonstrated competitive performance compared to sophisticated equivariant graph neural networks, even with minimal training data.
• FGW successfully identified seven previously unexplored photovoltaic absorber candidates from the Materials Project database.
• One identified material, Cs5Sb8, showed a predicted solar light material efficiency (SLME) exceeding 30% and good thermodynamic stability.

Conclusions:

• The Fused Gromov-Wasserstein (FGW) metric offers a powerful approach for materials exploration, leveraging strong inductive bias for efficient discovery with limited training data.
• This method effectively integrates chemical and structural information, paving the way for accelerated discovery of advanced materials like novel photovoltaics.