PhysicsML structure

The physicsml package is built on top of the molflux ecosystem which provides a standard API for handling the different stages of machine learning pipelines (accessing datasets, featurising, splitting, building models, and computing metrics).

molflux also handles and abstracts away the boilerplate code for training torch models via the lightning integration molflux[lightning]. For more info about lightning, see their documentation.

Layers

There are three layers to physicsml models:

1. The molflux layer: handles all functionality relating to the high level API.

2. The lightning layer: handles all boilerplate code for training and inference.

3. The torch layer: handles the torch level functionality of the models.

In the following sections, we discuss each of these layers, explain what they’re responsible for, their flexibility, and how you work with them.