search

jparkhill / TensorMol

Tensorflow + Molecules = TensorMol
http://blogs.nd.edu/parkhillgroup
GNU General Public License v3.0
271 stars 75 forks source link

Explanation of units in Tensormol 0.1 publication... #25

Closed Jussmith01 closed 6 years ago

Jussmith01 commented 6 years ago

We've noticed a few issues with units on errors in the Tensormol paper, which is making it hard to compare results. Can you provide (or point to them in the code) the equations used to compute the energy, force, and dipole errors? Our primary confusion is where the per atom part of the equation comes in. From your cost function it appears to be inside the squared part. Is this true from the reported errors? Thanks so much!

kunkinger commented 6 years ago

Hi Justin,

If you go to line 4881, 4891 and 4893 in TFNetworks/TFMolInstanceDirect.py, you can see "natom" in those lines. The "natom" is actually the 1/(number of atoms in each molecule) as you can see from line 1902 or line 1936 in Containers/TensorMolData.py

As you can see from the code, the denominator is inside the squared part. Sorry for the confusion caused by the naming.

-Best, Kun

On Fri, Mar 9, 2018 at 2:54 PM, Justin Smith notifications@github.com wrote:

We've noticed a few issues with units on errors in the Tensormol paper, which is making it hard to compare results. Can you provide (or point to them in the code) the equations used to compute the energy, force, and dipole errors? Our primary confusion is where the per atom part of the equation comes in. From your cost function it appears to be inside the squared part. Is this true from the reported errors? Thanks so much!

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/jparkhill/TensorMol/issues/25, or mute the thread https://github.com/notifications/unsubscribe-auth/AHG2WBKlfqcjNCPd8DYURrNWMP3QcU_Sks5tct3egaJpZM4SkyKY .

jparkhill commented 6 years ago

Justin- Hope all's well down there. Units are def awk af although I wouldn't call it an issue.

We're just simple men of simple pleasures, and quite certain the ANI networks are more throughly trained, on more data and welcome any reports that your charges/dipoles are more accurate. I would add that our published networks are under-trained simply because we just didn't even feel like waiting longer. The longer this field can avoid becoming a pissing match the better...

Best- John

Dom1L commented 6 years ago

I need to chime into this one...

I think the forces are actually in J mol-1 A-1. The conversion factor JOULEPERHARTREE that is defined in TensorMol/PhysicalData.py is the conversion factor from Hartree to J/mol and not Hartree to Joule.