MTS for baseline correction of ML potentials

[stefano ferrero]

Dear i-PI developers,

I have a doubt concerning the use of the MTS algorithm when one aims to apply a baseline to an ML potential.

From what I can see in the literature ( V. Kapil, J. VandeVondele, and M. Ceriotti, JCP 144(5), 054111 (2016) and O. Marsalek and T. E. Markland, JCP 144(5), (2016)),
one usually chooses 2 methods: the lower level of theory will describe the fast varying part of the potential whereas the higher level of theory describes the slowly varying part, aiming to arrive at the accuracy of the latter with the speed of the former.

 
In my case the ML potential has both the speed and the accuracy I aim for, and therefore I would imagine using it for both the inner and outer loop of the MTS
but I would like to apply a baseline correction to it with a tight-binding method to have more stable trajectories.

Can I achieve this baseline correction with this input part in the MTS forces? 

<forces>
<force forcefield="ML_Potential">
<mts_weights>[0,1]</mts_weights>
</force>
<force forcefield="Tight-Binding_Potential">
<mts_weights>[-1,0]</mts_weights>
</force>
<force forcefield="ML_Potential">
<mts_weights>[1,0]</mts_weights>
</force>
</forces>

Thank you very much for your help,

Stefano

[Michele Ceriotti]

In this scenario the most common application is to use DFTB as a baseline, and then train a correction (some call these "delta models") that however must be trained on the difference between the DFT target and DFTB. 

This http://doi.org/10.1021/acs.jctc.0c00362 is a good example, and I believe there should be example inputs somewhere that you can take as guidance. 

If you want to also do something fancier, here https://aip.scitation.org/doi/10.1063/5.0036522 we show how to use uncertainty quantification on the ML model to build a "variable baseline" model that falls back to DFTB when the ML model is close to blowing up. 

Hope this helps

M

[stefano ferrero]

Dear Michele,

Thank you very much for the answer. Now I have a more clear picture. 

I am just left with an operational doubt looking at the input file for the work you mentioned (http://doi.org/10.1021/acs.jctc.0c00362).

In the forces section (hereafter) I can recognize a committee of 5 delta models equally weighted acting on the outer loop with the DFTB baseline, whereas the "lammps1d" acting in the inner loop and in the outer correction step I suppose it would be an NN-potential trained "classically" on energy and forces and not a "delta-model". Is this correct?  

 <forces>

<force forcefield="driver-lammps1" weight="0.2">

<mts_weights>[1.0,0]</mts_weights>

</force>

<force forcefield="driver-lammps2" weight="0.2">

<mts_weights>[1.0,0]</mts_weights>

</force>

<force forcefield="driver-lammps3" weight="0.2">

<mts_weights>[1.0,0]</mts_weights>

</force>

<force forcefield="driver-lammps4" weight="0.2">

<mts_weights>[1.0,0]</mts_weights>

</force>

<force forcefield="driver-lammps5" weight="0.2">

<mts_weights>[1.0,0]</mts_weights>

</force>

<force forcefield="driver-dftb">

<mts_weights>[1.0,0]</mts_weights>

</force>

<force forcefield="driver-lammps1d">

<mts_weights>[-1.0,0.0]</mts_weights>

</force>

<force forcefield="driver-lammps1d">

<mts_weights>[-0.0,1.0]</mts_weights>

</force>

</forces>   

Thank you very much for your help and sorry for bothering again,

Stefano

[Mariana Rossi]

Hi Stefano,

In that publication they use a full NN (trained directly on the DFT data) in order to enable efficient multiple-time-stepping and the committee of NNs for the baseline correction (trained on the difference between DFT and DFTB+). The first one is indeed the one called "driver-lammps1d" in the example above. Depending on what is in the `nmts` block, that force will be corrected by the baseline+delta-committee every n steps.

Michele can correct any statements above, but I think that's all.

Cheers,

Mariana

[stefano ferrero]

Dear Mariana,

Thank you very much for your answer.

Best,

Stefano