Learning the Geometry of Materials Space with Trained Neural Tangent Kernels
Abstract
Neural networks trained to predict materials properties learn representations whose physical content can be difficult to extract. We use the neural tangent kernel (NTK), evaluated at trained parameters, to recover the geometry learned by a network that predicts superconducting transition temperatures from elemental composition alone. Kernel ridge regression with the trained NTK closely reproduces the network's predictive performance and substantially outperforms regression with the initial kernel. Spectral analysis and multidimensional scaling reveal a geometry dominated by a small number of directions. Clustering and local neighborhood analysis organize compounds into recognizable superconducting families and connect materials with related chemical and correlated-electron properties despite substantial differences in Tc. These results establish trained tangent kernels as practical probes of learned materials representations, connecting property prediction with the exploration of materials space.
Related
- Poisson Pencils, Lie Symmetries and Hamiltonian Reductions of a Coupled Nonlinear Wave System:Tangent KdV Geometry and Elliptic Moduli
- Chern bounds and tangent geometry of polarized Calabi-Yau threefolds
- Tangent-Plane Evidential Uncertainty in Active Learning for Magnetic Interatomic Potentials
- Learning Interaction Kernels from Collective Steady States
- An all-transistor physical-learning transformer can be trained to perform in-context learning
BibTeX
@misc{kaplan2026learning,
title = {{Learning the Geometry of Materials Space with Trained Neural Tangent Kernels}},
author = {Daniel Kaplan and Gabriel Kotliar and Anirvan M Sengupta},
howpublished = {arXiv},
year = {2026},
month = oct,
doi = {10.48550/arxiv.2610.27412},
eprint = {2610.27412},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2610.27412},
}
RIS
TY - UNPB
TI - Learning the Geometry of Materials Space with Trained Neural Tangent Kernels
AU - Daniel Kaplan
AU - Gabriel Kotliar
AU - Anirvan M Sengupta
PY - 2026
DA - 2026/10/20
PB - arXiv
DO - 10.48550/arxiv.2610.27412
UR - https://arxiv.org/abs/2610.27412
AB - Neural networks trained to predict materials properties learn representations whose physical content can be difficult to extract. We use the neural tangent kernel (NTK), evaluated at trained parameters, to recover the geometry learned by a network that predicts superconducting transition temperatures from elemental composition alone. Kernel ridge regression with the trained NTK closely reproduces the network's predictive performance and substantially outperforms regression with the initial kernel. Spectral analysis and multidimensional scaling reveal a geometry dominated by a small number of directions. Clustering and local neighborhood analysis organize compounds into recognizable superconducting families and connect materials with related chemical and correlated-electron properties despite substantial differences in T_c. These results establish trained tangent kernels as practical probes of learned materials representations, connecting property prediction with the exploration of materials space.
ER -