gabor1
commented
4 years ago The positions MUST be in Angstroms, otherwise the force (in eV/A) is not the precise derivative of the energy. Make sure the forces are really forces and not gradients (as given by some quantum chemistry packages).
On your soap command:
-
(n_max,l_max) = (14,14) is complete overkill, and also slows everything down. set it to (12,6) to get high accuracy results, (8,4) for something quicker
-
you need to add the "add_species=T" command, otherwise all atomic species are ignored
-
This is further down, once you get the accuracy that you want, but if you want a potential that can be used for high temperature MD, you will need to do something to enforce atomic repulsion at close approach. Either an explicitly fitted 2b potential (can be added to the gap string, look at published papers), or some other baseline. In this case I recommend not using the "e0_method=average" command, but actually explicitly compute the isolated atom energies and add them to the training set (they will be picked up and used for e0 for each species)
Let me know how you get on!
-- Gábor
Gábor Csányi Professor of Molecular Modelling Engineering Laboratory, University of Cambridge Pembroke College Cambridge
Pembroke College supports CARA. A Lifeline to Academics at Risk. http://www.cara.ngo/
On 26 Jun 2020, at 13:00, vvassilevg notifications@github.com wrote:
I would like to fit with GAP the PES of molecules using SOAP. As a test, I am using a glycine molecule (500 training points). So far, I get high Mean Absolute Errors on the training set (the errors are of course even higher for the test set) for both Energies and Forces (above 0.07 eV and 0.7 ev/A, respectively).
I have tested different parameters for the gap fit command. One example is:
gap_fit at_file=train.xyz \ gap={soap cutoff=5.0 \ covariance_type=dot_product \ zeta=2 \ delta=0.016 \ atom_sigma=0.3 \ l_max=14 \ n_max=14 \ n_sparse=2000 \ sparse_method=cur_points} \ force_parameter_name=forces \ e0_method=average \ default_sigma={0.001 0.2 0.0 0.0} \ do_copy_at_file=F sparse_separate_file=F \ gp_file=gap_soap.xml
I have tried different values for l_max, n_max and cutoff.
An example of a molecule in my train.xyz files is:
10 Lattice="200.0 0.0 0.0 0.0 200.0 0.0 0.0 0.0 200.0" Properties=species:S:1:pos:R:3:forces:R:3 energy=-7735.046780 pbc="T T T" N -5.400440 5.468773 2.837348 0.197830 0.001017 0.374624 H -6.365580 4.176578 3.913868 0.136175 0.068701 -0.293331 H -3.910505 6.045725 3.925944 -0.120680 0.003343 -0.049358 C -4.498210 4.417443 0.471706 -0.433094 1.062071 -0.682114 C -2.707543 2.205386 0.405318 0.102343 -0.550705 -0.104142 O -1.920056 1.231402 -1.521152 -0.009198 0.153232 0.169461 O -2.067417 1.333725 2.758112 -0.292697 0.513378 -0.084358 H -0.893916 -0.033471 2.454792 0.283477 -0.392345 -0.057309 H -6.148205 4.006446 -0.739463 0.074871 -0.442791 0.154181 H -3.631109 5.954848 -0.713001 0.060973 -0.415900 0.572343
Energy is in eV, forces in eV/A, and in this case, positions are in a.u., but I have also trained using A.
Is there anything missing (or wrong) when setting the gap_fit command and the soap descriptor?
I will be very grateful for any help you can provide.
Best regards,
— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub, or unsubscribe.
vvassilevg
mcaroba
I would like to fit with GAP the PES of molecules using SOAP. As a test, I am using a glycine molecule (500 training points). So far, I get high Mean Absolute Errors on the training set (the errors are of course even higher for the test set) for both Energies and Forces (above 0.07 eV and 0.7 ev/A, respectively).
I have tested different parameters for the gap fit command. One example is:
I have tried different values for l_max, n_max and cutoff.
An example of a molecule in my train.xyz files is:
10 Lattice="200.0 0.0 0.0 0.0 200.0 0.0 0.0 0.0 200.0" Properties=species:S:1:pos:R:3:forces:R:3 energy=-7735.046780 pbc="T T T" N -5.400440 5.468773 2.837348 0.197830 0.001017 0.374624 H -6.365580 4.176578 3.913868 0.136175 0.068701 -0.293331 H -3.910505 6.045725 3.925944 -0.120680 0.003343 -0.049358 C -4.498210 4.417443 0.471706 -0.433094 1.062071 -0.682114 C -2.707543 2.205386 0.405318 0.102343 -0.550705 -0.104142 O -1.920056 1.231402 -1.521152 -0.009198 0.153232 0.169461 O -2.067417 1.333725 2.758112 -0.292697 0.513378 -0.084358 H -0.893916 -0.033471 2.454792 0.283477 -0.392345 -0.057309 H -6.148205 4.006446 -0.739463 0.074871 -0.442791 0.154181 H -3.631109 5.954848 -0.713001 0.060973 -0.415900 0.572343
Energy is in eV, forces in eV/A, and in this case, positions are in a.u., but I have also trained using A.
Is there anything missing (or wrong) when setting the gap_fit command and the soap descriptor?
I will be very grateful for any help you can provide.
Best regards,