Continued challenges in high-throughput materials predictions: MatterGen predicts compounds from the training...

Area of Science:

• Materials Science
• Computational Chemistry
• Crystallography

Background:

• Generative AI and high-throughput computational tools promise to accelerate inorganic compound discovery.
• A key challenge is accurately modeling crystallographic sites occupied by multiple elements, particularly disordered phases.

Purpose of the Study:

• To evaluate the accuracy of AI-driven materials prediction tools, specifically Microsoft's MatterGen.
• To investigate the misclassification of known disordered materials as novel compounds by AI models.

Main Methods:

• Crystallographic analysis of a synthesized material predicted by MatterGen.
• Comparison of the synthesized material with existing crystallographic databases and literature.
• Evaluation of the AI model's training dataset for completeness and accuracy regarding disordered phases.

Main Results:

• MatterGen predicted TaCr2O6, which was synthesized as a disordered Ta1/3Cr2/3O2.
• Crystallographic analysis revealed this synthesized material is identical to the known Ta1/2Cr1/2O2, reported in 1971.
• The known compound Ta1/2Cr1/2O2 was present in MatterGen's training dataset, indicating a failure in distinguishing known disordered phases.

Conclusions:

• AI tools like MatterGen require rigorous human oversight and validation, especially for complex disordered materials.
• Current generative models are limited in predicting disorder and validating datasets, hindering rapid, large-scale materials discovery.
• Integrating crystallographic expertise is essential to enhance the reliability and potential of AI in materials science.