SPHNet uses learned sparse gates to prune atom pairs and tensor-product combinations, cutting training cost up to 7.1x while matching or improving Hamiltonian prediction accuracy on QH9, PubChemQH, and MD17.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Efficient and Scalable Density Functional Theory Hamiltonian Prediction through Adaptive Sparsity
SPHNet uses learned sparse gates to prune atom pairs and tensor-product combinations, cutting training cost up to 7.1x while matching or improving Hamiltonian prediction accuracy on QH9, PubChemQH, and MD17.