SE(3) Equivariant Neural Networks for Regression on Molecular Properties: The QM9 Benchmark
By Benjamin Kurt Miller. 2020
e3nn is an artificial neural network which operates on atomic coordinates and achieves equivariance to the special euclidean group in three dimensions by using spherical harmonics as features. The main experiment is to benchmark the model against a standard chemical data set called QM9, on which e3nn achieves state of the art performance on three of twelve regression targets. Along with empirical results, this thesis presents theoretical argumentation for why e3nn outperforms its closest relatives, SchNet and Cormorant, on some regression targets. Significant background regarding machine learning, quantum chemistry, and the special euclidean group is also presented. test Without significant hyperparameter optimization, the dipole moment result is already state of the art. a 0 is defined to be the bohr radius. The results are reported after fifty-five epochs of training.
How researchers see this paper
The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.
- Mario Geiger
- Tess E. Smidt
- Benjamin Kurt Miller
- Frank No'e
- Hao Wu
- Andreas Mardt
- Maurice Weiler
- Jonas Köhler
Citations in the corpus
- Relevance of Rotationally Equivariant Convolutions for Predicting Molecular Properties (2020, background): “A more general framing in terms of spherical harmonics can be found in the e3nn library [12, 23] .”