Method and apparatus for converter-based operator learning
By employing the General Neural Operator Transformer (GNOT) architecture and utilizing heterogeneous normalized attention (HNA) and hybrid expert (MoE) soft gating mechanisms, irregular mesh and multi-scale problems are solved, improving the efficiency and accuracy of partial differential equation (PDE) solving.
CN121014046APending Publication Date: 2025-11-25ROBERT BOSCH GMBH +1
0 Cites 0 Cited by
Patent Information
- Application Number
- CN202380097368.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2023-04-21
- Publication Date
- 2025-11-25
AI Technical Summary
Technical Problem
Existing technologies struggle to effectively handle irregular meshes, multi-input, and multi-scale problems, resulting in low efficiency in solving partial differential equations (PDEs).
Method used
It employs a general neural operator transformer (GNOT) architecture, combining heterogeneous normalized attention (HNA) blocks, normalized multi-head cross-attention, and expert hybrid (MoE) soft gating mechanisms to handle different input functions and multi-scale problems.
Benefits of technology
Efficient operator learning was achieved in irregular grids and multi-scale scenarios, improving the efficiency and accuracy of PDE solutions.
✦ Generated by Eureka AI based on patent content.
Abstract
A computer-implemented method for learning operator mappings from a plurality of input functions to a solution function of a partial differential equation (PDE) based on a transformer model is disclosed. The method includes inputting coordinates of a query point and a plurality of input functions to a converter model; encoding, via respective encoders, the coordinates of the query point and the plurality of input functions as feature embedding; and learning an operator by minimizing a mean square error (MSE) loss between an output feature processed by the transformer model and a truth value of the solution function.
Need to check novelty before this filing date? Find Prior Art