AQVolt26: High-Temperature r$^2$SCAN Halide Dataset for Universal ML Potentials and Solid-State Batteries

The paper introduces AQVolt26, a high-temperature r$^2$SCAN dataset of over 320,000 lithium halide configurations, demonstrating that while foundational machine learning potentials provide a strong baseline, their reliability for dynamic screening of solid-state electrolytes critically depends on augmentation with targeted high-temperature data rather than general near-equilibrium relaxation information.

The Gist

The Big Picture: The Quest for the "Perfect Battery"

Imagine you are trying to build a car that never catches fire, charges in seconds, and lasts forever. That is the dream of the Solid-State Battery. Unlike the lithium-ion batteries in your phone today (which use a flammable liquid inside), solid-state batteries use a solid block of material to move energy.

One of the most promising materials for this "solid block" is a Halide (a salt-like crystal). Think of halides as a dance floor where lithium ions are the dancers. For the battery to work, these dancers need to zip across the floor quickly.

The Problem: The "Soft" Dance Floor

The problem is that halide dance floors are wobbly.

• The Challenge: At room temperature, the floor is stable. But to make the lithium dancers move fast, you have to heat the floor up. When it gets hot, the floor gets "soft" and distorted. The dancers start jumping wildly, and the floor stretches and twists in crazy ways.
• The Old Way: Scientists used to try to predict how these dancers move using complex math (called DFT). It's like trying to calculate the exact path of every single dancer in a stadium by hand. It's accurate, but it takes forever.
• The New Way (AI): Scientists built AI models (Machine Learning Potentials) to act as a "crystal ball" that predicts how the dancers will move instantly. But here's the catch: Most of these AI models were trained on calm, room-temperature dance floors. When you ask them to predict what happens on a hot, chaotic, distorted dance floor, they get confused and make wild guesses.

The Solution: AQVolt26 (The "Heat Training" Dataset)

The authors of this paper realized that to teach the AI how to handle a hot, wobbly halide battery, they needed to show it examples of exactly that chaos.

They created AQVolt26, which is essentially a massive training manual for the AI.

• The Analogy: Imagine you are training a pilot. You can't just teach them to fly in perfect, sunny weather. You have to simulate storms, turbulence, and engine failures so they don't crash when the real storm hits.
• What they did: They generated 322,000+ specific computer simulations of lithium halides. They didn't just look at them when they were calm; they heated them up to extreme temperatures (up to 1,500°C) and watched them twist, stretch, and distort. They then used a super-accurate math method (r2SCAN) to label exactly what happened in every single one of these chaotic moments.

The Results: Teaching the AI to Dance in the Heat

Once they fed this "Heat Training" data (AQVolt26) into their AI models, the results were impressive:

1. Stability: The old AI models would crash (literally, the simulation would break) when the temperature got high. The new models, trained on AQVolt26, stayed stable. They knew how to handle the "wobbly floor."
2. Accuracy: The new models could predict how fast the lithium ions would move (ionic conductivity) much better than before. This is crucial because if the AI thinks the battery is fast when it's actually slow, you waste time building a bad battery.
3. The "Goldilocks" Lesson: The paper also found something interesting about mixing data.
   • If you only train on calm data, the AI is good at room temperature but crashes in heat.
   • If you train on chaotic data (AQVolt26), the AI becomes a master of heat but gets slightly "clumsy" at room temperature.
   • The Fix: The best approach is a mix. Use the "Heat Training" (AQVolt26) to teach the AI how to survive the storm, but keep a little bit of "Calm Training" (standard data) to keep it precise when things are quiet.

Why This Matters

This paper is a blueprint for the future of battery discovery.

• Before: Scientists had to guess which materials might work, or run slow, expensive simulations that often failed when things got hot.
• Now: They have a specialized "Heat Training" dataset that allows AI to safely and accurately simulate the extreme conditions inside a next-generation battery.

In short: The authors built a simulator for extreme heat so that AI can learn to design batteries that are safe, fast, and powerful, without needing to physically build and test thousands of dangerous prototypes first. They taught the AI to dance in the fire so we can build better batteries for our electric cars and phones.

Technical Summary

1. Problem Statement

The discovery of safe, high-energy-density solid-state electrolytes (SSEs) is critical for next-generation batteries, particularly for electric vehicles. Halide-based SSEs are a promising candidate due to their high ionic mobility and electrochemical stability. However, their discovery is hindered by the computational challenges of modeling their dynamic softness.

• The Challenge: Halides possess highly polarizable anions and frustrated structural arrangements, creating shallow, flat potential energy surfaces (PES). To accurately simulate ion transport, researchers must run molecular dynamics (MD) at elevated temperatures (>1000 K), which pushes the system into highly distorted, off-equilibrium regimes.
• The Limitation of Current Models: Universal Machine Learning Interatomic Potentials (MLIPs) trained on foundational datasets (e.g., Materials Project, MatPES) often fail in these regimes. While they perform well near equilibrium, their energy predictions degrade significantly under the high-strain, high-temperature conditions required to probe ion hopping. Furthermore, existing datasets often lack sufficient coverage of lithium halide configurations at these extreme states.

2. Methodology

The authors developed AQVolt26, a specialized dataset and training framework designed to bridge the gap between universal models and the specific needs of halide electrolytes.

A. Dataset Generation (AQVolt26)

• Scope: The dataset contains 322,656 r2SCAN single-point calculations for lithium halide systems.
• Source Material: Generated from 4,911 unique Li-halide candidates (including novel "stuffed" structures and oxyhalides) sourced from the Materials Project and MatPES.
• Sampling Strategy:
  1. Surrogate-Driven Exploration: A state-of-the-art r2SCAN ML force field was used to run NpT MD simulations at temperatures ranging from 300 K to 1500 K, generating over 200 million configurations.
  2. Dimensionality Reduction: The 2DIRECT (2-Stage Dimensionality-Reduced Encoded Clusters with Stratified) sampling method was applied. This reduced the 200M structures to a diverse, representative subset of ~186k structures for DFT labeling, maximizing feature space coverage while minimizing computational cost.
• DFT Settings: Calculations were performed using VASP with the r2SCAN meta-GGA functional (using PBE pseudopotentials as a proxy due to VASP limitations), ensuring high-fidelity electronic structure data.

B. Model Training

• Architecture: The authors utilized the eSEN (equivariant Smooth Energy Network) architecture, a state-of-the-art graph-based model.
• Training Protocol: A multi-stage training approach was employed:
  1. Pre-training: Direct force regression on a broad dataset (MatPES + MP-ALOE).
  2. Fine-tuning: Conservative force prediction (gradient of energy) with FP32 precision.
• Co-training Strategy: The study systematically compared models trained on:
  • Foundational data only (MatPES/MP-ALOE).
  • Foundational data + AQVolt26 (off-equilibrium halide data).
  • Foundational data + AQVolt26 + Materials Project (near-equilibrium relaxation data).

3. Key Contributions

1. AQVolt26 Dataset: The creation of the largest off-equilibrium dataset for solid-state electrolytes, specifically targeting the high-temperature, distorted configurations of lithium halides.
2. Demonstration of Domain-Specific Sampling: The paper proves that while universal baselines are robust, they require targeted, high-temperature domain data to reliably model dynamically soft materials.
3. Analysis of Data Composition: The study clarifies the distinct roles of different data types:
   • Off-equilibrium data (AQVolt26): Essential for dynamic stability and force accuracy in distorted regimes.
   • Near-equilibrium data (Materials Project): Improves relaxation accuracy but can degrade extreme-strain robustness if not balanced correctly.
4. Open Release: The dataset and models are made available via Hugging Face to accelerate community research.

4. Key Results

• Force and Energy Fidelity:
  • Models trained only on foundational data showed severe degradation in energy prediction for off-equilibrium halide structures (errors of 52–85 meV/atom).
  • Co-training with AQVolt26 reduced these errors to 15–24 meV/atom for strained configurations, while maintaining reasonable force predictions.
• Dynamic Stability (MD Simulations):
  • In heating ramp tests (300 K to 2100 K), AQVolt26-trained models maintained >80% survival rates for Li-halide systems, significantly outperforming models trained without the specific halide data.
  • The models preserved physical guardrails (preventing unphysical atom loss or volume explosion) up to 700 K and beyond.
• Physical Consistency (Strain Tests):
  • Under extreme hydrostatic strain (±20% volume change), the AQVolt26-trained model exhibited a 0.2% failure rate, outperforming several state-of-the-art baselines (e.g., MatPES-MACE at 14.8% failure).
  • The model predicted stiffer, more rigid bonds, better capturing the steep changes in the PES for perturbed structures.
• Ionic Conductivity:
  • The AQVolt26-trained model achieved a Mean Absolute Error (MAE) of 0.6 mS/cm for room-temperature conductivity in Li-halides, compared to 4.2 mS/cm for the baseline MatPES-TensorNet.
  • It reduced false positives (overestimation of conductivity) common in ML force fields.
• Trade-offs:
  • Including near-equilibrium Materials Project data improved relaxation benchmarks but slightly degraded performance in extreme-strain regimes and increased variance in formation energy predictions for the halide subset.

5. Significance and Impact

• Paradigm Shift in Dataset Design: The paper argues that for "dynamically soft" materials (like halides, sulfides, and hydrides), domain-specific configurational sampling is as critical as model architecture. Universal models cannot simply rely on broad, near-equilibrium datasets; they require explicit exposure to high-energy, distorted phase spaces.
• Accelerated Discovery: By providing a robust, physically consistent potential for high-temperature halide dynamics, AQVolt26 enables the reliable screening of solid-state electrolytes for ion transport, a task previously too risky for general-purpose MLIPs.
• Strategic Data Integration: The findings offer a blueprint for future MLIP development: use foundational datasets for a robust base, augment with targeted high-temperature/off-equilibrium data for dynamic stability, and use near-equilibrium data selectively as a task-specific complement rather than a universal additive.

In summary, AQVolt26 represents a critical step toward making machine learning a viable tool for the high-throughput discovery of next-generation solid-state battery materials by solving the specific challenge of modeling high-temperature, anharmonic lattice dynamics.