Antisymmetrische neuronale netzwerke
Patent Information
- Application Number
- AT2020767536T
- Authority / Receiving Office
- AT · AT
- Patent Type
- Patents
- Current Assignee / Owner
- Priority Date
- 2019-09-03
- Filing Date
- 2020-09-03
- Publication Date
- 2026-07-15
- Estimated Expiration
- 2040-09-03
AI Technical Summary
Current methods for predicting the wavefunction of chemical systems, such as molecules, are either computationally impractical for larger systems or inaccurate on non-equilibrium geometries, and existing neural network approaches fail to effectively incorporate Fermi-Dirac statistics.
An antisymmetric neural network architecture that uses permutation-equivariant functions in intermediate layers to process electron and pair features, allowing for accurate wavefunction approximation while mitigating computational load by separating input and output streams for electrons and pairs, and applying linear transformations and activation functions.
The antisymmetric neural network achieves high accuracy in predicting wavefunctions for challenging systems, outperforming variational quantum Monte Carlo and other ab-initio methods, and enables direct optimization of wavefunctions for previously intractable molecules and solids, without requiring basis set selection or extrapolation.