Neural Tangent Kernels Probe Learned Materials Representations

Researchers have developed a method using the neural tangent kernel (NTK) to understand the internal representations learned by neural networks that predict material properties. By evaluating the NTK at trained parameters, they can reveal the geometry learned by a network predicting superconducting transition temperatures. This approach, when combined with kernel ridge regression, closely matches the network's predictive accuracy and significantly outperforms using the initial kernel. Analysis of the learned geometry shows it is dominated by a few key directions, allowing for the organization of compounds into recognizable superconducting families and the identification of materials with similar chemical and correlated-electron properties. AI

IMPACT Provides a new method for understanding and interpreting the internal workings of AI models used in materials science research.

RANK_REASON The item is an academic paper published on arXiv detailing a new method for analyzing neural network representations in materials science. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• arXiv
• Chemical properties and optical properties of carbonaceous particles
• clustering algorithm
• correlated-electron properties
• Kernel Ridge Regression
• Materials space of solid-state electrolytes: unraveling chemical composition-structure-ionic conductivity relationships in garnet-type metal oxides using cheminformatics virtual screening approaches.
• multidimensional scaling
• Neural tangent kernel
• superconducting families
• Superconducting Transition Temperatures of Semiconducting SrTi O 3
• Turks and Caicos Islands

• paper
• other

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →