Predicting Mohs Hardness of Minerals Using Machine Learning on Atomic Descriptors

Predicting the hardness of a material is a challenging task due to the complex relationships between the atomic properties and the mohs hardness. To tackle this challenge, we leveraged machine learning models of linear regression, ridge regression, random forests, and gradient boosting to predict the hardness from the atomic descriptors. Among the models that were tested, random forests achieved the best results on average, reducing the error by around 30% compared to linear models and explained 60% of the variation in the hardness. Finally, an ablation study was conducted to figure out the importance of each atomic descriptor. The results showed that electron and bond related features, specifically the bond ZA Ratio and covalent radius, play a major role in predicting the hardness while more macroscopic properties such as density contribute very minimally. These findings show the effectiveness of machine learning in materials prediction and reveal the possibility for future works involving computer vision.Acknowledgement: This work was under under the guidance of Akshansh Mishra, current (April 2026) a Materials and Nanotechnology Engineer at Politecnico Di Milano. He works on applying artificial intelligence-based algorithms in the manufacturing and materials sectors. Grateful for his help and support in publishing my first research paper! Akshansh's Linkedin: https://www.linkedin.com/in/akshansh11/