Legacy Module

The legacy module contains the older versions of the KLIFF descriptors, neural network training utilities, and associated functionalities.

Tip

Although the module is named legacy it is not being deprecated and is still actively maintained.

In fact, this might be better solution for users who are looking for a more performant but less flexible neural network potentials, and certain UQ functionalities.

The major difference between the legacy module training and the new module training is that the legacy training method was more imperative, and the new training method is more declarative. In the newer training method you specify the model, the loss function, and the optimizer, and the training loop is handled by the KLIFF. KLIFF automatically loads the data, performs the desired splits, and trains the model. In the legacy training method, you write the python code to manually call the required KLIFF modules your self.

DUNN vs TorchML

The biggest advantage of the legacy module is the DUNN model driver support, which is not available in the new module. The DUNN model driver is a descriptor-based neural network model driver, which is much more performant than the new module model drivers. This is because the DUNN model driver uses Eigen3 based C++ backend for neural network evaluation, and analytical derivatives. The TorchML model driver uses the libtorch backend, which is not as performant but much more flexible. For appropriate use cases, the DUNN model driver can be orders of magnitude faster than the TorchML model driver.

Below table illustrates the difference DUNN and TorchML model drivers:

	TorchML	DUNN
Backend	libtorch	Eigen3 (C++)
Forces	Automatic Differentiation	Analytical Derivatives
Flexibility	Any arbitrary torch model (e.g. GNN)	Just DNN with enumerated non-linearities
Performance	Flexible but slower	Faster
UQ	No UQ support yet	Ensemble average

So decide as per your requirements which model driver you want to use.