Help for Neural Network Problem

[Anuj Kumar Sirohi]

 I am facing difficulties in making Neural Network for problem descried as follows:

1. Number of frames are 10000 with known corresponding Energy value

2 Each frame (box) have 216 particles with structure ( Charge_on_Particle, X_coordinate, Y_coordinate, Z_coordinate)

Now the problem is that I have target value (Energy) for whole Frame not for each particle. Kindly suggest some input format and network topology.

I am also attaching a photo of data description. Kindly suggest how can build a NNP for my problem.  

[Nongnuch Artrith]

Dear Anuj Kumar Sirohi,

You only need the total energy of each "frame" or structure to train an ANN potential.  The particle energies (or atomic energies) are not needed as input, and in fact there is no unique definition of the atomic energy.  

Note that currently aenet does not consider the charge of an input structure.

The structures (all XYZ coordinates of one "frame") and the total energies have to be provided in an extended XSF structure file format that is described in detail and with examples in the documentation: 
http://ann.atomistic.net/documentation/#structural-energy-reference-data

Best,
Nong

Thanks and Regards

Anuj Kumar Sirohi

          SC&IS(JNU)

[Anuj kumar sirohi]

Ma'am

You have considered charges in PRB Condensed Matter, 2011, 83, 153101. Can you please suggest on that basis how can I treat Coulombic Interaction.

--

[Nongnuch Artrith]

Anuj Kumar Sirohi,

In 2011, we used a different ANN potential software, we did not use aenet.  

We decided not to implement the charge training in aenet for now, as it appears to be not necessary for most systems.  From my experience, even for materials with significant ionic binding (such as transition metal oxides) it is typically sufficient and easier to increase the cutoff radius of the ANN potential a little to include most of the Coulombic contributions.  In solids, the Coulomb interactions are usually screened by the material and can be captured by intermediate-range potentials.  Around 8-12 Angstroms (or larger) are often enough.

Best,
Nong