[Code Idea]: Improve channel coupling

[galegozi]

Summary

This idea aims to improve the coupling between channels. The basic idea is to create a tighter coupling between certain channels and loosen the coupling between others.

Motivation

Imagine if there is copper and steel together in a simulation. In the current model, there is no way to tell it that the copper channels are related to each other, and that the steel channels are related to each other. Therefore, the embedding generated will be a mush between the two materials. The goal with this idea is to delay the mush by manually grouping relevant channels into an initial embedding, creating a "copper embedding" and a "steel embedding", in this case.

Affected Area

This would impact the encoding and decoding of the model, where the embedding is created (into packed) and where it is opened (out of packed).

Proposed Approach

The following steps would be used when inferencing with the model. For simplicity assume only copper and steel are part of the simulation:

1. A constant size embedding is generated for both copper and steel. Although it is possible to have an embedding engine for each material, it would be better to have one engine that can use any material. This would likely require something like [Code Idea]: Generate smart channel label embeddings #53 for best results. The idea is that if there is copper density, copper temperature, and steel density, this stage would combine the copper density and temperature to create a "copper" embedding, and resize the steel density to create a "steel" embedding.
2. The two are then merged together into a single constant size embedding. Note: this should also be able to take a variable amount of channels, in case there are different amounts of materials in the simulation. In this case it would take the copper and steel embedding and merge them together to create a merged embedding.

This would make the embedding take two steps; however, it could lead to a smarter embedding.

Alternatives Considered

The current approach can be used assuming it handles coupling properly; however, that is not known for sure.

Additional Context