Molecular dynamics (MD) simulations are a powerful tool to study liquids and interfaces, yet for many years they have suffered from a strict trade-off between accuracy and efficiency. Ideally, one would use ab-initio MD where the potential energy surface is computed at each step. However, these simulations are very costly and are typically restricted a few hundred atoms. Neural network potentials (NNPs) are an attempt to fix this trade-off by fitting the ab-initio interatomic potential energy surfaces using neural networks, allowing for AIMD-quality simulations at a significantly reduced cost. While NNPs have proven useful, generating NNPs is still difficult, requiring careful fitting of a large number of ab-initio energies which must adequately sample the available phase space. Even choosing the training data can be challenging, especially when studying phase transitions, as any implicit bias for a given phase or density can introduce an erroneous bias into the fitted potential energy surface. This bias can manifest in the NNP as each input for the neural network is typically normalized using the average and standard deviation of the input over the training data. To remedy this issue, we introduce volume-scaled NNPs, wherein each NNP input is scaled by a constant related to the volume of the potential cutoff. This scaling improves training in a more robust and physically meaningful manner, while also avoiding bias associated with the training data. Using Argon as a test case, we show that volume-scaling significantly improves the performance of the NNP, yielding an accurate thermal expansion coefficient and melting point, despite the training set using only crystalline structures with a majority having 1-4 atoms per unit cell. We show that the same performance cannot be achieved with the standard normalization procedure. We anticipate that volume-scaled NNPs will be an important advance in machine-learned potentials.

Volume scaling for neural network potentials

Macromolecular Chemistry: The Second Century

[Author Benefits](https://scimeetings.acs.org/?utm_source=pubs_content_marketing&utm_medium=website&utm_campaign=0320_MCF_NPI_Launch_Spring_Homepage&ref=pubs_content_marketing)

[How to use SciMeetings](https://storage.googleapis.com/dl.morressier.com/acs/SciMeetings-ACSSpring2020-HowToGuide_M3020-Final.pdf)

[Frequently Asked Questions](https://storage.googleapis.com/dl.morressier.com/acs/SciMtng_ACSNMSp2020_FAQ_M3120_Final.pdf)

ACS Spring 2020 National Meeting & Expo

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.