Data-model Coevolution as the Architectural Principle for AI-Native Materials Databases

This paper proposes and validates "data-model coevolution" as a foundational architectural principle for AI-native materials databases, demonstrating through a Li-P-S ternary prototype that endogenous generation-evaluation-refinement cycles can autonomously discover novel stable phases and achieve high-precision predictive modeling with minimal first-principles cost.

The Gist

Imagine you are trying to build the ultimate library of crystal structures for a specific type of material (in this case, a mix of Lithium, Phosphorus, and Sulfur).

The Old Way: The Static Library
Traditionally, scientists built these libraries like a static archive. They would use a set of rigid rules to generate thousands of crystal shapes, calculate their properties using supercomputers, and then just "file them away." The computer models used to predict properties were like external consultants who were hired, gave their advice, and then left. The library grew by adding more files, but the "brain" (the AI model) didn't learn from the new files, and the files didn't change based on what the brain learned. It was a one-way street.

The New Way: The Self-Evolving Garden
This paper proposes a new architectural principle called "Data–Model Coevolution." Think of this not as a library, but as a living, self-tending garden.

1. The Seed (The Generator): An AI "gardener" plants seeds (generates candidate crystal structures).
2. The Soil Test (The Evaluator): Another AI "tester" checks the soil (evaluates the stability of those crystals) using a fast, smart approximation.
3. The Expert Check (The Refinement): For the most promising plants, a human-level expert (a super-accurate computer simulation called DFT) does a deep check.
4. The Growth Loop: Here is the magic: The results of the expert check don't just get filed away. They are fed back into the gardener and the tester.
   • The Gardener learns: "Oh, I shouldn't plant seeds that look like that; they don't grow well. I'll try a different shape next time."
   • The Tester learns: "I can now predict soil quality even more accurately because I've seen these new plants."

In this system, the database (the garden) and the AI models (the gardener and tester) evolve together. They are inseparable parts of the same living system.

What They Actually Did
The researchers tested this "living garden" on a complex chemical mix: Lithium, Phosphorus, and Sulfur (Li-P-S). This is a tricky system, like trying to grow a rare, exotic plant in difficult soil.

• Rapid Maturity: Within just two or three rounds of this loop, the AI models became incredibly sharp. They reached a level of accuracy where they could predict energy and forces almost as well as the slow, expensive expert simulations, but much faster.
• Filling the Gaps: The system didn't just copy what it had seen before. It discovered new, stable crystal shapes that were missing from the world's biggest existing databases (like the Materials Project).
  • It found a stable version of a crystal called Li₂PS₃ that experts knew existed in real life but had never been found in the digital databases.
  • It invented new molecular "shapes" (like rings and chains of atoms) that had never been seen in the training data but were chemically plausible.
• The "Saturation" Signal: The researchers noticed that after a few rounds, the garden stopped producing new types of basic building blocks. It had explored all the possible ways atoms could bond in that specific chemical mix. This told them, "We have covered this territory; we don't need to keep guessing."

The Result: A Universal Query Tool
Once the garden was "stabilized" (the models were trained and the data was consistent), the researchers could ask the database any question directly. They didn't need to build a new tool for every question. They could ask:

• "Which of these crystals are stable?"
• "Which ones let Lithium ions move through them quickly (good for batteries)?"
• "What do the electrons look like inside these crystals?"

The system answered all of these using the same unified framework.

The Big Picture
The paper argues that instead of building bigger and bigger piles of static data, we should build AI-native databases. These are systems where the data and the AI models grow together in a closed loop. This allows scientists to explore a specific chemical system, master it, and then use that "mature" state as a foundation to explore related systems later. It turns the database from a passive storage unit into an active, learning partner in discovery.

Technical Summary

Technical Summary: Data–Model Coevolution as the Architectural Principle for AI-Native Materials Databases

1. Problem Statement

Current computational materials databases (e.g., Materials Project, OQMD, Alexandria) operate on a data-centric architecture. In these systems, databases function as static repositories where structural entries are accumulated via predefined workflows (template filling, elemental substitution, or crystal-structure prediction). Predictive models remain conceptually external to the database state; data growth is decoupled from model updating, and models do not endogenously drive the generation of new data. This structural separation limits the continuous accumulation of system-specific understanding and is incompatible with the iterative, AI-native discovery cycles where generative models propose candidates, surrogate potentials evaluate them, and first-principles calculations refine both data and models in a closed loop.

2. Methodology

The authors propose an AI-native materials database architecture based on data–model coevolution. In this framework, structural entries and integrated predictive models jointly constitute the database state. Database growth is treated as a state transition process driven by an endogenous generation–evaluation–refinement loop.

Core Components:

• Chemical System Nodes: The framework formalizes bounded chemical systems (defined by targeted elemental combinations and functional goals) as fundamental "nodes" of database growth. The Li–P–S ternary system serves as the demonstrative prototype.
• Generative Backbone: The study utilizes MatterGen, a deep generative model, to propose candidate crystal structures within the target chemical domain. Generation is conditioned on specific energy-above-hull ($E_{hull}$) targets (0.00, 0.03, and 0.06 eV/atom).
• Surrogate Evaluation: Machine-Learned Force Fields (MLFFs) are used for rapid, near-DFT-accuracy energetic evaluation and filtering. Three architectures were benchmarked: DPA-3, MACE, and MatterSim.
• Refinement Loop:
  1. Candidate Generation: The generative model proposes structures.
  2. Filtering: MLFFs evaluate stability ($E_{hull}$).
  3. Selection: Structures satisfying the Stable–Unique–Novel (S.U.N.) criteria are selected.
  4. First-Principles Refinement: A subset of selected structures undergoes Density Functional Theory (DFT) calculations (using VASP with PBE functional).
  5. Model Update: The generative model is fine-tuned using ground-truth DFT $E_{hull}$ values. Simultaneously, the MLFF is fine-tuned on structures selected via a maximum-information-entropy-gain criterion to maximize diversity while minimizing DFT cost.

Operational Metrics:

• Local Saturation: The diversity of local chemical environments is monitored via the information entropy of local atomic features. Convergence is signaled when entropy growth saturates.
• Model Convergence: MLFF accuracy is tracked via energy and force Root-Mean-Square Errors (RMSE) on test sets.

3. Key Contributions

1. Architectural Formalization: The paper formalizes data–model coevolution as the foundational principle for AI-native databases, shifting the paradigm from static data repositories to stateful systems where models are integral components of the database state.
2. Closed-Loop Implementation: A practical implementation of a closed-loop workflow that autonomously generates, evaluates, and refines data and models within a specific chemical system (Li–P–S) without relying on predefined motif libraries.
3. Discovery of Novel Motifs: The framework autonomously rediscovered a stable Li$_2$PS$_3$ phase and diverse P–S anionic motifs (e.g., (PS$_3$)$_3^-$ trimer, (P$_3$S$_8$)$^{3-}$ ring, polymeric (PS$_4$)$_n^{n-}$ chains) that were absent from the training databases (Materials Project and Alexandria) but consistent with historical experimental observations.
4. Unified Property Querying: The stabilized "data–model state" allows for direct querying of atomistic and electronic-structure properties (phase stability, ionic transport, charge density, band structure) within a single framework, eliminating the need for separate task-specific pipelines.

4. Key Results

• Scale and Efficiency: Over seven iterations, the framework generated approximately 70,000 candidate structures, with over 10,000 meeting the S.U.N. criteria.
• Rapid Saturation: The diversity of local chemical environments saturated within two to three iterations, indicated by the convergence of information entropy and the overlap of t-SNE distributions of local structural fingerprints.
• Model Performance:
  • The DPA-3 model achieved the best performance.
  • At $N_{train} = 4050$ (approx. 4,000 DFT frames), the fine-tuned DPA-3 achieved an energy RMSE of 6.8 meV/atom and a force RMSE of 85.1 meV/Å.
  • The $E_{hull}$ prediction RMSE improved from 46.9 to 26.5 meV/atom.
  • High-fidelity models were achieved with a manageable first-principles budget, showing diminishing returns beyond early iterations.
• Property Prediction:
  • Thermodynamics: The converged node supported P–T phase stability diagrams, revealing that Li$_2$PS$_3$ and Li$_3$PS$_4$ remain stable under finite pressure (up to 2 GPa) and temperature (300–600 K).
  • Ionic Conductivity: High-throughput molecular dynamics identified 29 Li-ion conductor candidates absent from the Materials Project, with conductivity thresholds of $\ge$ 400 mS/cm.
  • Electronic Structure: An integrated EAC-Net model predicted charge densities and band structures. After fine-tuning on only 34 frames, the normalized mean absolute error (NMAE) for charge density reached $\sim$4.8 $\times$ 10$^{-3}$, accurately reproducing DFT band dispersions.

5. Significance and Claims

The paper claims that data–model coevolution serves as a practical architectural principle for AI-era materials data infrastructure. By treating databases as stateful systems where data and models evolve together, the framework enables:

• Endogenous Growth: Database expansion is driven by internal feedback loops rather than external rules.
• Scalable Knowledge Accumulation: Chemical systems are formalized as "nodes" that can be reused, extended, branched, or transferred across related chemical systems, facilitating the modular accumulation of computational materials knowledge.
• Autonomous Exploration: The system can autonomously fill gaps in existing databases by rediscovering chemically plausible motifs absent from training distributions, effectively expanding the accessible chemical bonding space.

The authors emphasize that this approach unifies database growth and model evolution, allowing for continuous, transferable knowledge accumulation across chemical-system domains. They note limitations, including that the framework ensures internal consistency within bounded systems but does not guarantee experimental synthesizability, and that it currently focuses on near-equilibrium crystalline configurations rather than transition states or extreme regimes.