US-12621069-B2 - Equivariant spatially-consistent wireless channel models

Abstract

Systems and techniques are described herein for operating an apparatus having a geometric algebra transformer. An apparatus to predict link properties between a transmitter and a receiver in a three-dimensional space can include one or more processor; and a computer-readable medium storing instructions which, when executed by the one or more processor, cause the one or more processor to be configured to: receive a three-dimensional geometry, a transmitter position and a receiver position; and predict, based on a neural network wireless channel model, network link properties related to one or more channel between the transmitter position and the receiver position. Other tasks can be performed as well such as inferring wall position and/or orientation in the three-dimensional geometry based on a map of reference signal received power in a portion of the three-dimensional geometry.

Inventors

• Thomas Markus HEHN
• Markus PESCHL
• Tribhuvanesh OREKONDY
• Arash BEHBOODI
• Johann Hinrich BREHMER

Assignees

• QUALCOMM INCORPORATED

Dates

Publication Date

20260505

Application Date

20240514

Claims (20)

1. 1 . A processor-implemented method of processing data using a geometric algebra transformer, the processor-implemented method comprising: encoding, at a wireless geometric algebra tokenizer, one or more of three-dimensional scene data, a transmitter position, a receiver position and link information as a sequence of tokens which are geometric algebra representations; and processing, at a neural network, the sequence of tokens to generate output comprising N geometric algebraic multivectors and M scalars to jointly model a relationship between the three-dimensional scene data, the transmitter position, the receiver position and wireless channels.
2. 2 . The processor-implemented method of claim 1 , wherein the neural network comprises a geometric algebra transformer.
3. 3 . The processor-implemented method of claim 1 , wherein the encoding further comprises encoding, at the wireless geometric algebra tokenizer, the three-dimensional scene data, the transmitter position, the receiver position and the link information as the sequence of tokens.
4. 4 . The processor-implemented method of claim 1 , wherein the encoding further comprises encoding, at the wireless geometric algebra tokenizer, any combination of two or more of the three-dimensional scene data, the transmitter position, the receiver position and the link information as the sequence of tokens.
5. 5 . The processor-implemented method of claim 1 , wherein the N geometric algebraic multivectors comprise 16-dimensional vectors that represent geometric data.
6. 6 . The processor-implemented method of claim 1 , wherein the neural network is trained to process the sequence of tokens and to be equivariant with respect to rotations, translations and mirrorings.
7. 7 . The processor-implemented method of claim 1 , wherein the sequence of tokens comprises multi-vector inputs to the neural network.
8. 8 . The processor-implemented method of claim 7 , wherein the multi-vector inputs comprise embedded geometric objects.
9. 9 . The processor-implemented method of claim 8 , wherein the embedded geometric objects are embedded into the multi-vector inputs using a geometric algebra embedding component.
10. 10 . The processor-implemented method of claim 8 , wherein the embedded geometric objects comprise at least one of a scalar, a vector, a bivector, a trivector, or a pseudoscalar.
11. 11 . The processor-implemented method of claim 2 , wherein the geometric algebra transformer further comprises: an input equilinear layer; a transformer block; and an output equilinear layer.
12. 12 . The processor-implemented method of claim 11 , wherein the geometric algebra transformer further comprises a plurality of transformer blocks.
13. 13 . The processor-implemented method of claim 11 , wherein the transformer block further comprises: a first normalization layer; a first equilinear layer; a geometric attention layer; a first geometric product engine; a second equilinear layer; a first addition engine; a second normalization layer; a third equilinear layer; a geometric bilinear layer; a scalar-gated nonlinearity layer; a fourth equilinear layer; and a second addition engine.
14. 14 . The processor-implemented method of claim 13 , wherein the scalar-gated nonlinearity layer comprises a scalar-gated Gaussian Error Linear Units nonlinearity layer.
15. 15 . The processor-implemented method of claim 1 , wherein the three-dimensional scene data is encoded into a respective mesh face token per mesh face, wherein the respective mesh face token comprises one or more of a mesh face center position, vertex positions, relative vertex position from a center of the mesh face, mesh face plane vector, mesh face normal vector and material properties.
16. 16 . The processor-implemented method of claim 1 , wherein the transmitter position is encoded into a transmitter position token which comprises one or more of a transmitter position and a transmitter antenna orientation.
17. 17 . The processor-implemented method of claim 1 , wherein the receiver position is encoded into a receiver position token which comprises one or more of a receiver position and a receiver antenna orientation.
18. 18 . The processor-implemented method of claim 1 , wherein the link information is encoded into a link token which comprises one or more of a signal strength, a phase and a delay.
19. 19 . An apparatus to predict link properties between a transmitter and a receiver in a three-dimensional space, the apparatus comprising: one or more processor; and a computer-readable medium storing instructions which, when executed by the one or more processor, cause the one or more processor to be configured to: receive a three-dimensional geometry, a transmitter position and a receiver position; and predict, based on a neural network wireless channel model, network link properties related to one or more channel between the transmitter position and the receiver position.
20. 20 . An apparatus to infer scene properties in a three-dimensional space, the apparatus comprising: one or more processor; and a computer-readable medium storing instructions which, when executed by the one or more processor, cause the one or more processor to be configured to: receive a three-dimensional geometry and a signal strength data associated with the three-dimensional geometry; and infer, based on the signal strength data and a neural network wireless channel model, a scene property related to the three-dimensional geometry.

Description

TECHNICAL FIELD The present disclosure generally relates to processing data using machine learning systems. For example, aspects of the present disclosure include systems and techniques for tokenizing data associated with three-dimensional room geometry, a transmitter position and a receiver position, and processing the data via a geometric algebra transformer (e.g., tailored for a three-dimensional space and equivariant with respect to symmetries of three-dimensional space). The output of the geometric algebra transformer can be used to predict link properties or signal strength at different locations within the three-dimensional room geometry and for other purposes. BACKGROUND Wireless signal propagation in a given environment is most accurately described by Maxwell's equations through the electro-magnetic (EM) fields. Solving these equations analytically is intractable for most non-trivial scenarios, and while numerical solutions are feasible at small scales, it is not practical for large scale environments. In the so-called far-field regime, macroscopic effects dominate the signal behavior. The shift has led to the emergence of geometric optics and the uniform theory of diffraction as essential tools for modeling wireless propagation. These foundational concepts underpin state-of-the-art ray tracing simulations, enabling accurate predictions of signal behavior in complex environments. Ray tracing has been a substantial tool in modeling the channel given an environment, yet for future applications, such as sensing, some model properties and speed of computation are lacking. SUMMARY Systems and techniques are described for providing an equivariant and spatially-consistent wireless channel model that uses a geometric algebra transformer. The geometric algebra transformer can be used as a general model that can be used to solve any problem with inputs and outputs that are geometric in nature. The geometric algebra transformer is agnostic as to the problem to be solved. The geometric algebra transformer may be trained to perform a single task or to perform more than one task. The geometric algebra transformer may act either on specific instructions regarding the task to be achieved or operate on another type of input according to its training. The present disclosure introduces a new neural network or neural data-driven differentiable wireless channel model that can include a geometric algebra tokenizer and geometric algebra transformer (also called a geometric algebra backbone), which can be trained to provide accurate and fast wireless channel prediction. In some examples, the model can be used to predict a wireless channel between a transmitter (Tx) and a receiver (Rx) in a building environment or other environment in novel scenes not seen in training data. The model can jointly model relationships between the transmitter, the receiver, the three-dimensional scene and the channel between the transmitter and the receiver. The model can also enable inverse problems like completing a scene or identifying a position and/or orientation of a wall in a three-dimensional scene or inferring another physical feature of the scene. The model can also be used to predict a position of the transmitter and/or receiver or predict other characteristics of the transmitter and/or receiver. An extension of the model can be to use a diffusion model to predict distributions. Further, the model can be an equivariant wireless channel model meaning it can take advantage of symmetries within the three-dimensional space. In some aspects, the techniques described herein relate to a processor-implemented method of processing data using a geometric algebra transformer, the processor-implemented method including: encoding, at a wireless geometric algebra tokenizer, one or more of three-dimensional scene data, a transmitter position, a receiver position and link information as a sequence of tokens which are geometric algebra representations; and processing, at a neural network, the sequence of tokens to generate output including N geometric algebraic multivectors and M scalars to jointly model a relationship between the three-dimensional scene data, the transmitter position, the receiver position and wireless channels. In some aspects, the techniques described herein relate to an apparatus to predict link properties between a transmitter and a receiver in a three-dimensional space, the apparatus including: one or more processor; and a computer-readable medium storing instructions which, when executed by the one or more processor, cause the one or more processor to be configured to: receive a three-dimensional geometry, a transmitter position and a receiver position; and predict, based on a neural network wireless channel model, network link properties related to one or more channel between the transmitter position and the receiver position. In some aspects, the techniques described herein relate to an apparatus to infer a scene property i