Distilling Large-Scale Machine Learning Interatomic Potentials to Predict Long-Timescale Electrolyte Transport Properties

Ishan Amin, Yue Jian, Nitesh Kumar, Christine Zhang, Jianwei Lai, Casey S. Mezerkor, Aaron M. Melemed, Jonathan R. Thurston, Neil P. Dasgupta, J. David Bazak, Ethan J. Crumlin, Samuel Blau, Aditi S. KrishnapriyanarXiv 2026 · arXiv

This paper introduces a "knowledge distillation" workflow that allows researchers to predict the transport properties of battery electrolytes with near-quantum accuracy at a fraction of the computational cost. By transferring the accuracy of a large, slow "teacher" machine learning model into a fast, system-specific "student" model, the authors achieved a 40x speedup, enabling the long-timescale simulations necessary for high-throughput electrolyte screening.

TL;DR

The authors developed a method to distill large-scale, pretrained machine learning interatomic potentials into fast, specialized models that run 40x faster. This allows for the 20-nanosecond simulations required to accurately calculate electrolyte transport properties (like viscosity and conductivity), outperforming classical force fields without requiring any new, expensive ab initio data.

What Problem Does This Paper Address?

Designing next-generation batteries requires a deep understanding of how ions move through liquid electrolytes. Key metrics like ionic conductivity, viscosity, and self-diffusivity—collectively known as transport properties—determine a battery's power density and charging speed.

However, calculating these properties is a "Goldilocks" problem:

  1. Classical Force Fields are fast enough to run long simulations but often lack the chemical accuracy to model complex ion-ion interactions correctly.
  2. Ab Initio Molecular Dynamics (AIMD) is highly accurate but computationally grueling; it is too slow to reach the multi-nanosecond timescales needed for transport properties to "converge" (settle into stable, reliable values).
  3. Large-scale Machine Learning Interatomic Potentials (MLIPs) offer near-quantum accuracy but are still too heavy and slow for high-throughput screening of the vast space of possible electrolyte chemistries.

Methods

The researchers proposed a Teacher–Student Distillation workflow:

  1. Data Generation: They start with a "Teacher" model—a large, pretrained MLIP called UMA-OMol. They run a very short (100 ps) reference trajectory for a specific electrolyte system.
  2. Distillation & Robustness: They sample snapshots from this trajectory. Crucially, they "rattle" half of these snapshots with random Gaussian noise and have the Teacher relabel them with energy, force, and stress targets. This "noising" teaches the Student model how to handle configurations it might encounter off the equilibrium path, preventing simulation crashes.
  3. Student Training: They train a much smaller, faster "Student" model (based on the PaiNN architecture) on this specific dataset.
  4. Production: The Student model then runs four independent 20 ns simulations. Because it is ~40x faster than the Teacher, it can complete in days what would have taken the Teacher months.
**Figure 1:** The distillation workflow, showing data generation, production simulations, and property convergence.

Datasets

The study validated the method across 16 diverse electrolyte systems, including:

  • Salts: NaPF₆, NaOTf, and a LiPF₆ baseline.
  • Solvents: Monoglyme (DME), diglyme (DEGDME), tetraglyme (TEGDME), and propylene carbonate (PC).
  • Concentrations: Ranging from dilute (0.1 M) to concentrated (1.0 M).

The models were trained solely on labels provided by the UMA-OMol teacher (pretrained on the OMol25 dataset) and evaluated against in-house experimental measurements of density, viscosity, and transport.

Key Insights and Ideas

  • Noise is Mandatory: A major "aha" moment was finding that students trained only on smooth MD trajectories are unstable. Adding "rattled" structures to the training set provides a "restoring force" that keeps the long-term simulations stable.
  • Faithful to a Fault: The Student models reproduce the Teacher's behavior almost perfectly. However, this means they also inherit the Teacher's biases, such as a systematic ~10% overestimate of mass density.
  • The Convergence Requirement: The authors demonstrate that simulations shorter than 5 ns are too noisy to provide reliable transport data. The 20 ns window enabled by distillation is essential for accuracy.
  • Cutoff Limitations: The Teacher's 6 Å radial cutoff was identified as a source of error in low-ionicity systems (like NaOTf in DME), where long-range ion pairing is significant.

Key Results

  • Speed: The PaiNN student models achieved a median throughput of 9.5 ns/day, a 41x speedup over the UMA-OMol teacher (0.23 ns/day).
  • Accuracy vs. Experiment: The distilled models significantly outperformed the standard classical baseline (OPLS-AA).
  • Ionic Conductivity: The student halved the error of OPLS-AA (36% vs 67% median deviation).
  • Shear Viscosity: The student was more than twice as accurate as OPLS-AA (93% vs 204% median deviation).
  • Trend Tracking: Despite the inherited density bias, the student models correctly captured chemical trends, such as how diffusivity decreases as glyme chain length increases.
FIG. 3: Ionic conductivity and shear viscosity vs. experiment at 298 K, for the distilled student models and the OPLS-AA force field. Each point is the mean over four independent 20 ns replicas, and error bars show ±1 s.d. across the replicas. (a,b) Ionic conductivity σ for the PaiNN student and OPLS-AA. Conductivity is only computed for the concentrated (≥0.5 M) systems, since at 0.1 M the box holds too few charge carriers to converge the collective estimator. The student roughly halves the median absolute percentage deviation of OPLS-AA (MdAPD; 36% vs. 67%), with a comparable log–log correlation. (c,d) Viscosity η against experiment for the two force fields. Again the student tracks experiment much more closely than OPLS-AA, with an MdAPD less than half that of OPLS-AA and a log–log fit slope closer to one.

Limitations and Future Work

The primary limitation is the inherited bias from the teacher model, specifically the density overshoot and the lack of explicit long-range electrostatics due to the 6 Å cutoff.

The authors note that their pipeline is "architecture agnostic." As better "foundation" MLIPs are released—those with better density predictions or explicit long-range interactions—they can be dropped into this distillation workflow as new Teachers to immediately produce even more accurate Student models for electrolyte design.

Summary updated

Abstract

Accurately predicting the transport properties of liquid electrolytes is critical for next-generation battery design and requires force fields that are both quantum-mechanically accurate and fast enough for long-timescale simulation. Recent large-scale pretrained machine learning interatomic potentials (MLIPs) approach density-functional-theory accuracy, but remain too slow to access the necessary timescales to converge transport properties. We provide a knowledge distillation procedure to transfer the accuracy of a large-scale "teacher" MLIP into fast, system-specific "students", without any additional ab initio data. Our distilled models run ~40x faster, reproduce teacher-predicted structural observables, and produce stable 20 ns trajectories in days, rather than the months the teacher would require. Across a diverse set of sodium- and lithium-ion electrolytes, student-predicted self-diffusivities, ionic conductivities, and shear viscosities track experiment and outperform classical force-field predictions, establishing distillation as a practical route to accurate, long-timescale simulations necessary for high-throughput electrolyte screening.
BibTeX
@misc{amin2026distilling,
  author = {Ishan Amin and Yue Jian and Nitesh Kumar and Christine Zhang and Jianwei Lai and Casey S. Mezerkor and Aaron M. Melemed and Jonathan R. Thurston and Neil P. Dasgupta and J. David Bazak and Ethan J. Crumlin and Samuel Blau and Aditi S. Krishnapriyan},
  title = {{Distilling Large-Scale Machine Learning Interatomic Potentials to Predict Long-Timescale Electrolyte Transport Properties}},
  year = {2026},
  eprint = {2610.39513},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2610.39513},
}
Figure