Very high atomic forces with accurate potential in MD simulation

[Nicholas Martinez]

I have managed to generate a potential with low RMSE using 5000 structures, but when I simulate classical MD with it, the atomic forces are very high and the atoms rapidly disseminate--it is not at all representative of the system.

Can someone provide some guidance about how I can influence how aenet learns these atomic forces, especially for atoms very close together?

Thank you

[Alexander Urban]

This sounds like a symptom of overfitting.  When the ANN potential reproduces the energy exceedingly well, it can introduce “wiggles” in the potential energy that affect the forces.  

See also Figure 8 in Miksch et al., Mach. Learn.: Sci. Technol. 2, 2021, 031001, https://doi.org/10.1088/2632-2153/abfd96

Alex

--
Alexander Urban
Assistant Professor
Department of Chemical Engineering
Columbia University
812 Mudd Building, 500 W120th Street, New York, NY 10027
212-854-0105
https://urban-group.cheme.columbia.edu

Columbia Electrochemical Energy Center
https://ceec.engineering.columbia.edu

--
You received this message because you are subscribed to the Google Groups "aenet" group.
To unsubscribe from this group and stop receiving emails from it, send an email to aenet+un...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/aenet/790010cf-cf11-483f-8d9e-1db46fac7c99n%40googlegroups.com.

[Nicholas Martinez]

Ah okay, thank you for the insight