Thermodynamic assessment of machine learning models for solid-state synthesis prediction

Area of Science:

• Materials Science
• Computational Chemistry
• Artificial Intelligence

Background:

• Machine learning models are increasingly used to predict the synthesizability of solid-state materials.
• These models learn from databases of successful syntheses, aiming to bypass computationally intensive first-principles calculations.
• However, the thermodynamic validity of these predictions remains largely unassessed.

Purpose of the Study:

• To evaluate the alignment of existing machine learning synthesis prediction models with fundamental material and reaction thermodynamics.
• To establish bounds for thermodynamic quantities (energy relative to convex hull, reaction selectivity) beyond which synthesis is unlikely.
• To introduce a novel method for assessing machine learning model quality using thermodynamic heuristics.

Main Methods:

• Computed thermodynamic quantities (energy relative to convex hull, reaction selectivity) for hypothetical materials using the CHGNet potential.
• Generated novel hypothetical materials using the Chemeleon generative model.
• Assessed four machine learning synthesizability prediction models against computed thermodynamics and established bounds.

Main Results:

• Machine learning models generally overpredict the likelihood of material synthesis.
• Some model scores correlate with thermodynamic heuristics, assigning lower scores to less stable materials or those lacking thermodynamically selective synthesis routes.
• Established bounds for thermodynamic quantities to identify potentially unsynthesizable materials.

Conclusions:

• Existing machine learning models for materials synthesis prediction have limitations and tend to overestimate synthesizability.
• Thermodynamic analysis provides a valuable framework for assessing and improving these predictive models.
• This work highlights gaps in current models and offers a new approach for quality assessment, particularly crucial in the absence of negative synthesis data.