🔬 materials science

PhononScore: a phonon-aware scoring function for dynamical stability

PhononScore is a computationally efficient, phonon-aware scoring function that significantly enhances the dynamical stability of candidate materials generated by crystal models, enabling high-throughput screening and closed-loop inverse design without the need for expensive explicit phonon calculations.

Original authors: Xiao-Qi Han, Ze-Feng Gao, Zhong-Yi Lu

Published 2026-07-10

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Xiao-Qi Han, Ze-Feng Gao, Zhong-Yi Lu

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a master chef trying to invent a million new recipes for a futuristic restaurant. You have a super-smart AI that can whip up endless lists of ingredient combinations. But here's the catch: most of these "recipes" are actually disasters. If you tried to cook them, the pot would explode, or the dish would instantly fall apart. In the world of materials science, these exploding dishes are called "dynamically unstable" crystals. They look great on paper, but in reality, they vibrate so wildly they can't exist.

For years, checking if a crystal recipe is safe has been like hiring a team of expert food critics to taste-test every single dish. It's accurate, but it takes forever. One expert (a computer program called MatterSim) takes about 144 seconds to check just one crystal. Another takes 272 seconds. If you have a million candidates, you'd be waiting for centuries.

Enter PhononScore, a new tool that acts like a super-fast "gut feeling" detector for crystal stability. Instead of running a full, slow taste-test, PhononScore looks at the crystal's structure and instantly gives it a "stability score." It's like having a magic wand that can scan a million recipes in seconds and tell you, "Hey, these top 100 look delicious and won't explode!"

How It Works: The Multi-Task Chef

PhononScore isn't just guessing; it's trained on a massive library of 157,463 crystal structures. Think of this library as a mix of:

100,606 AI-generated "fake" crystals (the ones that might be unstable).
46,899 real crystals from a database called Materials Project.
9,958 high-precision crystals checked by the most rigorous experts (using a method called DFT-PBE).

The model learns by doing three things at once:

Predicting the "vibe": It guesses the lowest vibration frequency (the "minimum phonon frequency"). If this number is too low (specifically below -0.1 THz), the crystal is wobbly and unstable.
Pass/Fail checks: It learns to spot if a crystal passes different levels of safety tests.
Checking the geometry: It looks at how atoms are arranged locally to see if the shape makes sense.

By combining these clues, it creates a single "PhononScore." A higher score means the crystal is more likely to be stable.

The Magic Results

The researchers tested PhononScore on a huge benchmark called PhononBench, which included crystals generated by nine different AI models. Here is what happened:

Before PhononScore: If you picked the top 100 crystals from the original AI generators, only about 30.7% were actually stable. The rest were likely to crumble.
After PhononScore: When they used the tool to re-rank the same list, the top 100 candidates jumped to 83.7% stability. That's a 2.72-fold improvement!
The Top 10: Even better, if you just look at the top 10 picks, 97.5% of them are stable.

The tool is incredibly fast. While the old methods took minutes per crystal, PhononScore takes about 0.0336 seconds per structure. That's fast enough to give feedback while the AI is still generating the crystals!

It's Not Just a Guess; It's a Calibrated Map

You might wonder, "Is it just lucky?" The paper shows that the score isn't random. There is a clear, straight-line relationship: the higher the score, the higher the chance the crystal is actually stable.

In the lowest-scoring group, less than 1% of crystals were stable.
In the highest-scoring group, nearly 90% were stable.

The authors tested this on a "hard mode" scenario where stable crystals were rare (only 5% of the pool were stable). Even then, PhononScore managed to find 6.36 times more stable crystals in the top picks than random guessing would.

A Real-World Example: The K–I–O System

To prove it understands the physics and not just the statistics, the authors looked at three specific chemicals: KIO3, K3IO5, and KIO4. They all contain Potassium (K), Iodine (I), and Oxygen (O), but in different amounts and shapes.

K3IO5 was the most unstable (vibrating at -2.434 THz).
KIO3 was in the middle (vibrating at -1.817 THz).
KIO4 was perfectly stable (vibrating at 0.000 THz).

PhononScore-DFT (the version fine-tuned with high-precision data) correctly ranked them in this exact order: K3IO5 < KIO3 < KIO4. It figured out that the specific way the oxygen atoms were arranged in KIO4 made it stable, while the other shapes caused the atoms to shake apart. This proves the tool learned the actual rules of how atoms dance together, not just memorized chemical formulas.

What It Is NOT

It's important to know what PhononScore doesn't do. It is not a replacement for the slow, expensive expert checks when you need absolute proof. It is a "surrogate" or a "scoring function." Think of it like a movie trailer: it gives you a great idea of whether a film will be a hit, but you still need to watch the full movie (run the full DFT calculation) to be 100% sure before you spend your money.

The authors also note that the model is limited by the data it was trained on. It works best for materials that behave like "harmonic" springs (where atoms vibrate predictably). It might struggle with materials that have wild, unpredictable vibrations or strong electron interactions, because the training data for those is scarce.

The Bottom Line

PhononScore is a game-changer for finding new materials. It turns a bottleneck that took hours per crystal into a process that takes milliseconds. By acting like a "stability radar," it allows scientists to filter out the millions of unstable "fake" crystals and focus their expensive computing power only on the few that are truly ready to be discovered. It's not a magic wand that solves everything, but it's the fastest, most reliable compass we've had for navigating the vast ocean of possible materials.

Technical Summary: PhononScore

Problem Statement

Recent advances in crystal generative models (e.g., MatterGen, GNoME, InvDesFlow) have enabled the rapid design of massive numbers of candidate materials. However, a critical bottleneck impedes their translation into practical discovery: a significant fraction of generated structures are dynamically unstable. Large-scale benchmarks like PhononBench reveal that under stringent criteria (minimum phonon frequency $\omega_{min} > -0.1$ THz), the average stability rate of generated structures is often below 35%, with some models performing as low as 14%.

While universal machine-learning interatomic potentials (uMLIPs) like MatterSim offer a faster alternative to Density Functional Theory (DFT) for phonon calculations, they still require minutes per structure (e.g., 144–990 seconds). This computational cost is prohibitive for real-time feedback in generative loops, active learning, or the screening of millions of candidates. There is a distinct lack of a computationally efficient, phonon-aware scoring function analogous to those used in drug discovery (e.g., docking scores, pLDDT) that can rapidly rank candidates by dynamical stability without explicit phonon spectrum calculation.

Methodology

Framework and Architecture

PhononScore is introduced as the first phonon-aware scoring function specifically designed for crystal generation. Rather than performing explicit lattice vibration calculations, it predicts a unified stability score directly from crystal structures. The architecture utilizes:

Graph Representations: Each crystal is represented as a periodic atom graph and a corresponding line graph. This jointly encodes atomic identities, bond lengths, bond angles, and local coordination environments.
Multi-Task Learning: The model employs a shared Graph Neural Network (GNN) encoder with three distinct heads to optimize complementary objectives:
1. Regression Head: Predicts the minimum phonon frequency ( $\omega_{min}$ ) using Smooth L1 loss.
2. Multi-Threshold Classification Head: Predicts the probability of satisfying multiple stability thresholds ( $\tau \in \{-0.001, -0.01, -0.1, -1.0\}$ THz) using balanced binary cross-entropy with monotonicity constraints.
3. Geometry Head: Uses a Mixture Density Network (MDN) to learn the likelihood of local geometric patterns (pairwise distances) characteristic of stable structures.
Unified Ranking Objective: The outputs of the three heads are combined into a single score ( $S = S_{reg} + \beta S_{thr} + \alpha S_{geom}$ ). A threshold-aware pairwise ranking loss is introduced to directly optimize the enrichment of dynamically stable structures, ensuring that structures satisfying stricter stability criteria receive higher scores.

Data and Training Strategy

The model is trained on a multi-fidelity phonon dataset containing 157,463 crystal structures:

Low/Medium Fidelity: 100,606 AI-generated structures and 46,899 Materials Project structures labeled with MatterSim-derived phonon frequencies.
High Fidelity: 9,958 structures with DFT-PBE phonon labels.

A two-stage training strategy is employed:

Pretraining: The model is pretrained on the large-scale MatterSim-labeled dataset to learn general stability patterns.
Fine-tuning: The pretrained model is fine-tuned on the high-fidelity DFT-PBE dataset to create PhononScore-DFT, calibrating the scores to first-principles standards.

Data splitting is performed at the formula level to prevent data leakage, ensuring that all reduced formulas in the test set are excluded from training.

Key Results

Reranking and Enrichment (PhononBench)

Evaluated on the PhononBench benchmark (9,000 candidates from 9 generative models):

Stability Rate Improvement: Reranking with PhononScore increased the average dynamical stability rate of Top-100 candidates from 30.7% (original pool) to 83.7%.
Enrichment Factor: This represents a 2.72-fold enrichment of dynamically stable structures.
Top-K Performance: The Top-10 candidates achieved a stability rate of 97.5%.
Comparison to Baselines: PhononScore outperformed a baseline ALIGNN model trained to directly regress $\omega_{min}$ , which achieved only 69.4% stability for Top-100 candidates. This indicates that optimizing for ranking and stability enrichment is more effective than simple frequency regression.

High-Fidelity Transfer and Hard Screening

DFT Transferability: On a balanced DFT-PBE test set, the fine-tuned PhononScore-DFT achieved a 93.0% stability rate for Top-100 candidates and 100.0% for Top-10.
Hard-Screening Scenarios: In imbalanced scenarios simulating rare stable structures (e.g., 5% stable fraction), PhononScore-DFT achieved a 6.36-fold enrichment for Top-10 candidates, recovering stable structures at a rate far exceeding random selection.
Monotonicity: The predicted scores exhibit a monotonic correspondence with the true probability of stability, rising from <1% stability in the lowest score bin to ~90% in the highest.

Mechanistic Validation

Stoichiometric Sensitivity: The model correctly recovered non-trivial stability orderings in the K–I–O system ( $K_3IO_5 < KIO_3 < KIO_4$ ), distinguishing stability differences arising from stoichiometry and local coordination rather than just chemical composition.
Physical Interpretation: Analysis of specific cases (e.g., $A_2BInCl_6$ perovskites) showed that the score correlates with physical stabilization mechanisms, such as the suppression of soft modes via octahedral tilting and symmetry lowering.

Significance and Claims

The paper claims that PhononScore establishes a practical foundation for high-throughput materials discovery by reformulating dynamical stability assessment as a scoring-function learning problem rather than an explicit phonon prediction task.

Efficiency: It reduces the computational cost of stability assessment to the second-level (0.0022s per structure for model forward pass), enabling real-time feedback for crystal generation, active learning, and reinforcement learning.
Role as a Surrogate: PhononScore acts as a calibrated surrogate scoring function, analogous to those in drug discovery, providing a unified and efficient evaluator for dynamical stability without explicit phonon calculations.
Integration: It serves as a model-agnostic post-processing module that can be integrated into existing AI-driven workflows to rerank candidates before expensive first-principles validation, effectively bridging the gap between generative models and high-fidelity physical simulations.

The authors note that the model's performance is fundamentally limited by the quality of the underlying supervision data (harmonic approximation) and the availability of high-fidelity data for anharmonic or strongly correlated systems, but it represents a significant step toward scalable, stability-aware materials design.