Key Findings
The development of the qNEP (charge-aware neuroevolution potential) allows for the simulation of sodium metal electrode-organic electrolyte interfaces over timescales significantly longer than previous methods. In a system of roughly 1,000 atoms, the potential remained stable for 100 nanoseconds, representing a two to three order of magnitude improvement over prior lithium and sodium simulations. For larger systems of approximately 27,000 atoms, simulations lasting tens of nanoseconds were achieved, with the longest single-GPU run taking eight days.
Technical Details
The qNEP strategy utilizes three primary pillars:
- High-temperature ab initio molecular dynamics sampling to broaden descriptor space coverage and enhance ergodicity.
- Injection of short ab initio simulations at critical nodes during active learning to prevent unphysical energy pathways at bifurcation points.
- A training set incorporating inorganic products (sodium fluoride, sodium oxide) and organic decomposition intermediates.
The model was trained using the GPUMD package on four RTX 4090 GPUs over approximately two days. It contains 22,633 trainable parameters for six elements, based on a training set of 5,035 configurations totaling 683,526 atoms. Accuracy metrics show root mean square errors of 10.50 meV per atom for energy and 251.89 meV per angstrom for forces on the training set. This force error is roughly 38 times smaller than that of the ReaxFF force field (9,680 meV per angstrom). Speed-wise, qNEP is one order of magnitude faster than ReaxFF and six orders of magnitude faster than ab initio molecular dynamics.
Background & Context
Sodium batteries offer a cheap, abundant alternative to lithium-ion systems, but their performance is limited by the solid electrolyte interphase (SEI)—a passivation film formed during initial charging. Direct observation of SEI formation is difficult because conventional techniques like X-ray photoelectron spectroscopy provide only snapshots. Previous simulations were either too short (ab initio) or lacked the transferability and accuracy to capture single-electron reduction pathways (ReaxFF).
Significance & Outlook
By incorporating trainable atomic charges, qNEP can track charge transfer and distinguish chemical states, a feature missing in standard machine learning potentials. This allows researchers to simulate and differentiate the decomposition pathways of various electrolytes, including carbonates (ethylene carbonate, dimethyl carbonate, propylene carbonate) and ethers (dimethoxyethane, diglyme).
Get our weekly technology intelligence — free
Receive an infographic that lets you judge at a glance whether each field’s analysis report is worth reading.
Subscribe Free — Weekly Tech Intelligence
By subscribing, you’ll receive Troy-Technical’s weekly technology intelligence newsletter.
- Your email and selected fields are used only to deliver the newsletter.
- We never share your information with third parties.
- You can unsubscribe anytime via the link in each email.
See our Privacy Policy for details.
Takes about a minute · Unsubscribe anytime
