.. _fine_tuning_universal:

=====================================================================
Fine-tuning universal models
=====================================================================

Often, out-of-the-box performance of a universal model is insufficient. It would be nice if we could improve its performance with little effort. Indeed, ML techniques such as transfer learning provide the means for fine-tuning the universal model for a specific application.

In the example below, we will see how the universal ANI-1ccx-gelu model can be fine-tuning to improve the accuracy of the harmonic frequencies. Below is the Jupyter :download:`notebook <materials/fine-tuning/tl_ani.ipynb>` and required files (:download:`CH3NO2_MP2_avtz.json <materials/fine-tuning/CH3NO2_MP2_avtz.json>` and :download:`ch3no2_init.xyz <materials/fine-tuning/CH3NO2_init.xyz>`) taken from the MLatom documentation and inspired by `our publication <https://doi.org/10.1021/acs.jpclett.4c03031>`__:

.. raw:: html
   :file: materials/fine-tuning/tl_ani.html

.. note::

   Fine-tuning for the universal models is supported for ANI-type models ANI-1x, ANI-1ccx, ANI-1ccx-gelu, and ANI-2x.

.. note::

    We have changed the code in MLatom to make it easier, so above scripts would need modifications, see official MLatom tutorials for information. In the future editionsn of this course, I might update this.

    .. include:: outdating.rst