| Mar 21, 2026 |
Machine learning models trained on molecular quadrupole moments predict electrostatic potentials rapidly, enabling faster discovery of battery electrolyte candidates.
|
|
(Nanowerk News) Researchers at Uppsala University have shown that machine learning models trained on molecular quadrupole moments can accurately predict the electrostatic potential around battery electrolyte molecules, replacing quantum-chemical calculations that typically take days or weeks. The study, published in AI for Science (“Molecular electrostatic potentials from machine learning models for dipole and quadrupole predictions”), demonstrates that quadrupole moments are a far more effective training target than the more commonly used dipole moments for this task.
|
Key Findings
- Machine learning models trained on quadrupole moments reconstructed molecular electrostatic potentials with substantially higher accuracy than dipole-trained models.
- The improvement was consistent across two standard benchmark datasets of organic molecules, QM9 and SPICE.
- The method enables rapid screening of electrolyte and solvent candidates for energy storage devices without expensive quantum calculations.
|
|
Electrolytes control how ions travel inside a battery, how electrode interfaces form, and how stable a device remains over repeated charge cycles. Identifying better electrolyte molecules is difficult because their performance depends on subtle intermolecular forces, solvation effects, and charge distributions. Resolving these properties with standard quantum-chemistry methods is computationally expensive, which limits the number of candidate molecules that can be evaluated.
|
 |
| Molecular electrostatic potential of fluoropropylene carbonate (FPC) generated from physics-constrained PiNet2 models. (Image: Kadri Muuga and Chao Zhang, Uppsala University)
|
|
The molecular electrostatic potential, or MEP, is a central quantity in this problem. It maps how a molecule attracts or repels charge in the surrounding space and is widely used to study intermolecular interactions, chemical reactivity, and solvent design. Computing the full MEP from first principles, however, can require days or weeks of processor time for a single molecule, making large-scale screening impractical.
|
|
The Uppsala team investigated whether a neural network could learn to reconstruct the MEP from simpler multipole information. They used the PiNet2 message-passing neural network architecture and trained separate models on two targets: molecular dipole moments, which capture the simplest charge asymmetry in a molecule, and molecular quadrupole moments, which encode a richer, higher-order picture of charge distribution.
|
|
On the QM9 dataset, a standard collection of small organic molecules used to benchmark chemical property predictions, the quadrupole-trained models reproduced the electrostatic potential far more faithfully than their dipole-only counterparts. The same pattern held on the SPICE dataset, which covers a broader and more chemically diverse set of molecules. In both cases, the extra information carried by the quadrupole moment translated directly into more accurate charge models and better MEP reconstructions.
|
|
These gains have direct relevance for battery electrolyte discovery. The trained models predict electrostatic potentials in a fraction of the time a full quantum calculation requires, making it feasible to screen large libraries of candidate solvents and electrolyte additives. Characterizing the electrostatic interactions that govern solvation and interfacial behavior at this scale would be prohibitive with conventional electronic-structure methods.
|
|
The study also carries a broader lesson for machine learning in chemistry. Dipole moments are typically the default training target because they represent the leading term in the multipole expansion for neutral molecules. Quadrupole moments, though less commonly used, proved far more informative when the objective is to recover a full electrostatic landscape from point-charge representations. This suggests that careful selection of training targets can yield outsized improvements in model performance.
|
