All-in-one: Learning across quantum chemical levels. Better than transfer learning!

The explosion of quantum chemical datasets (see our overview of them) satisfies the appetite of those data-hungry machine learning potentials while raising another critical question: how to learn data in different fidelities?

Here, we propose the all-in-one (AIO) ANI model, which is able to handle an arbitrary number of QC levels. The idea behind it is simple. We added the level of theory as an additional input to the model, and thus, the model gained the ability to distinguish data from various fidelities and infer their correlations.

If you are still suffering from finding the best set of hyperparameters for transfer learning, AIO may solve your problem since similar accuracy is achieved with AIO while no expertise in fine-tuning is required and all parameters will be saved in just one model.

Another advantage of AIO is its combination with the delta-learning strategy, where any pairs of corrections from low level to high level can be obtained, as long as AIO has already seen them.

More detailed discussions can be found in our preprint at ChemRxiv. The code and the foundational models are available at https://github.com/dralgroup/aio-ani. They will be integrated into the universal and updatable AI-enhanced QM (UAIQM) library and made available in the MLatom package so that they can be used online at the XACS cloud computing platform.