US 12,248,861 B2
Antisymmetric neural networks
David Benjamin Pfau, London (GB); James Spencer, London (GB); and Alexander Graeme de Garis Matthews, London (GB)
Assigned to DeepMind Technologies Limited, London (GB)
Filed by DeepMind Technologies Limited, London (GB)
Filed on Sep. 3, 2020, as Appl. No. 17/011,569.
Claims priority of provisional application 62/895,491, filed on Sep. 3, 2019.
Prior Publication US 2021/0064961 A1, Mar. 4, 2021
Int. Cl. G06N 3/04 (2023.01); G06F 17/18 (2006.01)
CPC G06N 3/04 (2013.01) [G06F 17/18 (2013.01)] 23 Claims
OG exemplary drawing
 
1. A method comprising:
obtaining input data characterizing a chemical system having a plurality of nuclei and a plurality of electrons, the input data comprising electron features for each of the plurality of electrons and pair features for each of a plurality of pairs of the plurality of electrons; and
processing the input data using an antisymmetric neural network having a plurality of intermediate layers, wherein:
the antisymmetric neural network is configured to process the input data to generate as output a predicted value of one or more properties of the chemical system, the one or more properties comprising one or more of a wavefunction of the chemical system or a ground state energy of the chemical system,
each of the plurality of intermediate layers of the antisymmetric neural network is configured to generate a respective layer output for the intermediate layer from a respective layer input to the intermediate layer by applying a respective permutation-equivariant function to the respective layer input,
the antisymmetric neural network is configured to generate the predicted value from the respective layer output of the last intermediate layer of the plurality of intermediate layers,
a respective layer output of a last intermediate layer of the plurality of layers comprises a respective output stream for each of the plurality of electrons,
generating the predicted value comprises:
generating, from the respective layer output of the last intermediate layer, a respective input for each of a plurality of pairs of determinants, each pair of determinants including (i) one determinant that operates on a respective spin up matrix that is generated from respective output streams for spin up electrons and has a respective row and a respective column for each of the spin up electrons and (ii) another determinant that operates on a respective spin down matrix generated from respective output streams for spin down electrons and has a respective row and a respective column for each of the spin down electrons, and wherein each pair of determinants of the plurality of determinants has respective spin down and spin up matrices that are different from each other pair of determinants of the plurality of determinants;
determining a respective output of each determinant in the plurality of pairs of determinants from the respective input for the determinant; and
determining the predicted value from the respective outputs of each of the determinants in the plurality of pairs of determinants,
and
the antisymmetric neural network has been trained on training data to optimize at least a variational energy objective.