# US 2021/0081776 A1 · Configuration of an optical switch fabric using machine learning

> Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt

- Number: US 2021/0081776 A1
- Document ID: US20210081776A1
- Status: Published application
- Kind: A1
- Application number: 16/569,496
- Filed: 2019-09-12
- Published: 2021-03-18
- Earliest claimed: 2019-09-12
- Assignees: International Business Machines Corporation
- Applicants: International Business Machines Corporation
- Inventors: NICOLAS DUPUIS; BENJAMIN GILES LEE
- Web page: https://patentgrep.com/patent/US20210081776A1

## CPC

- G06N 3/0675: {using electro-optical, acousto-optical or opto-electronic means}
- G06N 3/04: Architecture, e.g. interconnection topology
- G06N 3/0499: Feedforward networks
- G06N 3/08: Learning methods
- G06N 3/084: Backpropagation, e.g. using gradient descent
- G06N 3/09: Supervised learning
- H04Q 11/0005: {Switch and router aspects}
- H04Q 11/0062: {Network aspects}
- H04Q 2011/0039: {Electrical control}
- H04Q 2011/0049: {Crosstalk reduction; Noise; Power budget}
- H04Q 2011/005: {Arbitration and scheduling}
- H04Q 2011/009: {Topology aspects}
- G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- H04Q: SELECTING (switches, relays, selectors H01H; wireless communication networks H04W)

## Abstract

An optical switch fabric comprises two or more optical switch elements. The optical switch elements are configured in a topology. A switch control has a plurality of bias control signals. The switch control can address one or more of the optical switch elements and can apply one of the bias control signals to bias of the addressed optical switch element to establish a switch setting. The topology and switch settings determine how each of one of the inputs is connected to each of one of the outputs of the optical switch fabric. The switch settings are determined by a machine learning process which includes a model creation. The model can be made to adapt dynamically during optical switch fabric operation.

## Claims

### Claim 1

An optical switch fabric comprising:
two or more optical switch elements, each optical switch element having a bias, one or more optical switch element inputs and one or more optical switch element outputs, one or more optical switch element outputs from a first optical switch element connected to one or more inputs of a second optical switch element so that the optical switch elements are optically connected in a topology;
one or more inputs to the optical switch fabric, each input connected to one of the optical switch element inputs of an optical switch element at an input of the topology;
one or more outputs from the optical switch fabric, each output connected to one of the optical switch element outputs of an optical switch element at an output of the topology; and
a switch control with a plurality of bias control signals, the switch control capable of addressing one or more of the optical switch elements, being an addressed optical switch element, and switch control applying one of the bias control signals to the bias of the addressed optical switch element to establish a switch setting, where the switch setting is determined by a machine learning process and the switch settings and topology determine how each of one of the inputs is connected to each of one of the outputs.

### Claim 2 (depends on 1)

A switch fabric, as in claim 1, where the machine learning process is performed by a deep neural network (DNN).

### Claim 3 (depends on 2)

A switch fabric, as in claim 2, where the DNN is trained with a dataset, the dataset having a plurality of records, each record having output information for the optical switch fabric and switch settings for each of the optical switch elements.

### Claim 4 (depends on 3)

A switch fabric, as in claim 3, where each record further comprises other information.

### Claim 5 (depends on 4)

A switch fabric, as in claim 4, where the other information includes one or more of the following: operating temperature of the optical switch fabric, humidity, and ambient temperature.

### Claim 6 (depends on 1)

A switch fabric, as in claim 1, where one or more of the optical switch elements is a 2×2 Mach-Zehnder switch (MZS).

### Claim 7 (depends on 6)

A switch fabric, as in claim 6, where the bias control signal is applied to an optical phase shift control of the 2×2 MZS to control the switch setting.

### Claim 8 (depends on 6)

A switch fabric, as in claim 6, where the machine learning process is performed by a deep neural network (DNN) trained with a dataset of a plurality of records, each record having output information, disambiguating information about the output information, and optical phase shifts as switch settings for each of the MZSs.

### Claim 9 (depends on 8)

A switch fabric, as in claim 8, where the disambiguating information includes one or more of the following: derivatives of the output information, and representative information about the derivatives of the output information.

### Claim 10 (depends on 1)

A switch fabric, as in claim 1, where the model parameters are updated during operation of the switch fabric.

### Claim 11 (depends on 10)

A switch fabric, as in claim 10, where the model parameters are updated by changes to one or more outputs of a machine learning process.

### Claim 12

A training process comprising the steps of:
a. inputting learned feature information into one or more inputs of a machine learning system;
b. comparing one or more outputs of the machine learning system to a desired label to determine a difference;
c. using one or more of the differences to determine changes to one or more model parameters in the machine learning system;
d. determining whether the differences are within a validation accuracy;
e. if the differences are within the validation accuracy, using existing model parameters in the machine learning system as a model of an optical switch fabric;
f. if the differences are not within the validation accuracy, backwardly propagating the changes to the model parameters through the machine learning system;
g. repeating steps b through g until the differences are within the validation accuracy.

### Claim 13 (depends on 12)

A process, as in claim 12, where the machine learning system is a Deep Neural Network (DNN).

### Claim 14 (depends on 12)

A process, as in claim 12, where the learned feature information includes output values of an optical switch fabric and the labels are a switch setting for each of one or more optical switch elements connected in a topology to form the optical switch fabric.

### Claim 15 (depends on 14)

A process, as in claim 14, where the learned feature information further comprises other parameters.

### Claim 16 (depends on 14)

A process, as in claim 14, where the output values are updated by a data acquisition system monitoring an operation of the optical switch fabric and steps a through g are repeated with the updated output values to produce a new set of model parameters.

### Claim 17 (depends on 14)

A process, as in claim 14, where the switch setting is an optical phase shift for a Mach-Zehnder switch (MZS).

### Claim 18 (depends on 17)

A process, as in claim 17, where the learned feature information further comprises disambiguating information about the output values.

### Claim 19 (depends on 12)

A process, as in claim 12, where the learned feature information includes a switch setting for each of one or more optical switch elements connected in a topology to form the optical switch fabric and the labels are output values of an optical switch fabric associated with switch settings.

### Claim 20

An optical switch fabric system comprising:
an optical switch fabric comprising:
two or more optical switch elements, each optical switch element having a bias, one or more optical switch element inputs and one or more optical switch element outputs, one or more optical switch element outputs from a first optical switch element connected to one or
more inputs of a second optical switch element so that the optical switch elements are optically connected in a topology;
one or more inputs to the optical switch fabric, each input connected to one of the optical switch element inputs of an optical switch element on an input of the topology;
one or more outputs from the optical switch fabric, each output connected to one of the optical switch element outputs of an optical switch element on an output of the topology;
a switch control with a plurality of bias control signals, the switch control capable of addressing one or more of the optical switch elements being an addressed optical switch element and switch control applying one of the bias control signals to the bias of the addressed optical switch element to establish a switch setting, where the switch setting is determined by a machine learning process and the switch settings and the topology determining how one of the inputs is connected to one of the outputs;
an optical switch model running on a processing unit;
a data acquisition unit monitoring the outputs of the optical switch fabric, and providing one or more new output information to the optical switch model;
machine learning process using the new output information to re-create the model to form a re-created model, the re-created model having update model parameters.
