🔬 materials science

Symbolic Predicate-Guided Language Agents for Inverse Design of Perovskite Oxides

This paper introduces ORCHESTRA, a multi-agent framework that enhances large language models for the inverse design of perovskite oxides by translating natural language rules into symbolic predicates within a domain-specific language, thereby enabling statistical validation and iterative refinement of design principles without requiring task-specific datasets or model retraining.

Original authors: Dong Hyeon Mok, Seoin Back, Victor Fung, Guoxiang Hu

Published 2026-07-20

📖 7 min read🧠 Deep dive

CC BY 4.0

Original authors: Dong Hyeon Mok, Seoin Back, Victor Fung, Guoxiang Hu

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 new dish that tastes perfect, but you don't have a recipe book. Instead of guessing blindly, you have a super-smart AI assistant who knows a little bit about cooking but sometimes makes things up. This is the world of materials science, where scientists try to invent new materials (like super-strong metals or better batteries) by working backward: instead of asking "What does this material do?", they ask "What material do I need to make this specific thing happen?"

To do this, scientists often use Large Language Models (LLMs). Think of these as the AI assistants that have read almost every chemistry textbook ever written. They are great at guessing what ingredients might work together. However, these AI chefs have a flaw: they can be confidently wrong. They might suggest a recipe that sounds delicious but is actually impossible to cook. To fix this, researchers are trying to give these AI chefs a way to check their own work using real data, turning their wild guesses into tested facts. This paper is about building a smarter kitchen where the AI doesn't just talk about recipes but actually tests them against a database of real cooking results.

The Problem: The AI Chef Who Hallucinates

In the past, scientists used computers to simulate millions of materials to find the best ones. But this takes a lot of time and requires huge amounts of data. Recently, people started using AI language models to "dream up" new materials. These AI agents are clever; they can talk through their reasoning and suggest new combinations of atoms.

However, there's a catch. Because these AI models are trained on text, they sometimes "hallucinate." They might say, "If you mix Element X with Element Y, you get a super-catalyst!" and it sounds totally logical. But if you actually try to build it, it might fail. The AI is confident, but it's wrong. Previous attempts to fix this involved having multiple AI agents argue with each other, but if they all share the same bad logic, they just end up arguing in circles.

The Solution: ORCHESTRA and the "Magic Translator"

The authors of this paper, led by Dong Hyeon Mok and colleagues, built a new system called ORCHESTRA (Operational Rule-grounded CHEmical Search Through Reasoning Agents). Think of ORCHESTRA as a kitchen with four specialized robots working together:

The Designer: This robot suggests new material recipes based on what it knows and what has worked before.
The Tester: This robot uses a fast computer simulation (a "surrogate model") to see if the new recipes actually work.
The Interpreter: This robot looks at the test results and tries to write down the "rules" of why something worked.
The Curator: This robot is the boss. It decides which rules are real and which are fake.

The secret sauce of ORCHESTRA is a Domain-Specific Language (DSL). Usually, when the Interpreter robot writes a rule, it writes it in plain English, like "Use Cobalt on the B-site." The problem is that English is vague. The DSL acts like a magic translator. It forces the AI to translate its English sentence into a strict, mathematical code that can be instantly checked against a database of thousands of real materials.

Instead of just saying, "I think Cobalt is good," the AI has to write a code that says, "Count how many materials in our database have Cobalt here, and what percentage of them actually worked." This turns a vague guess into a hard fact. If the code shows that only 1 out of 100 materials with Cobalt worked, the Curator robot knows to throw that rule in the trash.

The Experiment: Cooking with Double Perovskites

To test this system, the scientists asked the AI to design double perovskite oxides. These are a specific type of crystal structure (like a very organized Lego tower) used in things like fuel cells and solar panels. They gave the AI two main challenges:

The Easy Challenge: Find materials with a specific "band gap" (how well they conduct electricity) and "formation energy" (how stable they are).
The Hard Challenge: Find materials that are great at the Oxygen Evolution Reaction (OER). This is a tricky chemical process needed to split water into hydrogen fuel. It's hard because it depends on three different chemical steps happening at once, not just one.

They compared four different ways of designing materials:

Random: Just picking materials like rolling dice.
Background: Letting the AI guess based only on its training, with no feedback rules.
Rule-Only: Letting the AI guess and write English rules, but without the strict math check.
DSL-Guided: The full ORCHESTRA system with the math translator.

The Results: Math Beats Guessing

The results showed that the DSL-Guided team was the clear winner, especially for the hard challenge.

For the easy task (band gap/energy): The DSL-guided AI found successful materials about 68.1% of the time. The "Rule-Only" AI was close behind at 64.2%. This means that for simpler problems, even just talking about rules helps the AI.
For the hard task (OER): The DSL-guided AI found successful materials about 21.3% of the time. The other methods did much worse.

Why did the math translator make such a big difference? The authors found that the "Rule-Only" AI kept making mistakes. It would suggest a rule like "Use Lead," and because it couldn't mathematically check the data, it kept believing that rule even when the data showed it was wrong. The DSL-guided system, however, could instantly see that a rule had "zero precision" (meaning it was wrong) and delete it immediately.

The system also showed it could learn. As the AI ran more tests, its rules got better and better. The "precision" of its rules (how often they were right) went up, and the "lift" (how much better the rule was than random guessing) also increased.

What They Found in the Kitchen

When the scientists looked at the materials the DSL-guided AI successfully designed for the OER challenge, they found something cool: the AI had rediscovered real-world chemistry secrets.

It figured out that Cobalt on the B-site was essential.
It found that Ruthenium or Iridium on the neighboring site worked best.
It noticed that certain rare earth elements (like Neodymium or Praseodymium) on the A-site helped.

These are exactly the kinds of ingredients real human chemists use in the lab! The AI didn't just find random numbers; it found the actual chemical patterns that make these materials work.

The Limits: The Translator Has a Vocabulary

The paper also admits that the system isn't perfect. The "magic translator" (the DSL) has a limited vocabulary. If the AI wanted to say something very complex about the surface of the material that the translator couldn't understand, the system couldn't check it. In some cases, the "Rule-Only" AI actually performed slightly better because it could say more complex things in English, even if it couldn't check them with math.

The authors suggest that in the future, they might need a bigger translator with more words to cover all the complex chemistry. But for now, they proved that giving an AI a way to check its own math against real data makes it a much better scientist.

The Takeaway

This paper suggests that we don't need to train a brand-new AI from scratch to discover new materials. Instead, we can take a smart, existing AI and give it a "calculator" (the DSL) to ground its wild guesses in reality. By forcing the AI to translate its ideas into code that can be tested, we stop it from making up facts and help it find real, working materials faster. It's like giving a dreamer a ruler and a calculator so their dreams can actually be built.

Technical Summary: Symbolic Predicate-Guided Language Agents for Inverse Design of Perovskite Oxides

Problem Statement

The discovery of high-performance materials is increasingly driven by data- and AI-based strategies, particularly the paradigm of inverse design, where materials are constructed from target properties rather than screened from a predefined space. While generative models and high-throughput virtual screening have shown success, they often rely heavily on large, representative datasets, limiting their utility in poorly explored chemical systems where extensive computational data is unavailable.

Recent approaches utilizing Large Language Models (LLMs) offer a promising alternative by leveraging reasoning and in-context learning to propose candidates and interpret design principles without extensive task-specific training. However, LLM-based reasoning suffers from fundamental limitations: the probabilistic nature of generation can lead to hallucinations or unsupported claims that appear chemically plausible. Furthermore, self-reflection or multi-agent critique mechanisms rely solely on the intrinsic reasoning capacity of the underlying model. If incorrect reasoning accumulates during iterative design loops, it can misguide candidate generation and degrade performance compared to conventional algorithmic strategies. There is a need for a framework that grounds LLM reasoning in external, verifiable statistical evidence to mitigate these risks.

Methodology

The authors propose ORCHESTRA (Operational Rule-grounded CHEmical Search Through Reasoning Agents), a multi-agent framework that integrates LLM reasoning with a Domain-Specific Language (DSL) for chemistry. The core innovation is the translation of natural language design rules into symbolic predicates that can be executed against accumulated materials data.

Framework Architecture

The framework operates through four iterative stages involving distinct agent roles:

Design Agent: Proposes new material candidates based on chemical background knowledge, historical evaluation data, and a curated "rule store."
Calculation Node: Evaluates target properties of generated candidates using surrogate models (specifically, Machine Learning Force Fields).
Interpretation Agent: Analyzes the structure-property data to extract relationships. Crucially, it generates not only natural language design rules but also symbolic predicates (Boolean-valued expressions) defined within a predefined chemistry DSL.
Curate Agent: Evaluates the proposed rules using the symbolic predicates. It calculates statistical evidence (e.g., number of matched candidates, precision, lift) from the accumulated dataset. Rules with weak empirical support are revised or discarded, while supported rules are added to the rule store.

Symbolic Predicate Layer

The DSL allows the LLM to express rules using specific symbols and operators (e.g., elemental fractions, d-filling, charge buffering). When a rule is proposed, the corresponding predicate is applied to the dataset to generate quantitative metrics:

Precision: The rate of matched materials that follow the claimed target direction.
Lift: How much the success rate of matched materials exceeds the rate in the full dataset.

This process enables the framework to ground natural language hypotheses in data, allowing the system to validate, refine, or reject design principles iteratively.

Material System and Surrogate Models

The framework was applied to the inverse design of Double Perovskite (DP) Oxides ( $AA'BB'O_6$ ).

Target Properties: Band gap, formation energy, and Oxygen Evolution Reaction (OER) overpotential.
Surrogate Models: The Universal Models for Atoms (UMA) were used as a computationally efficient surrogate for Density Functional Theory (DFT). While UMA showed systematic errors in absolute energy values for certain lanthanide-containing systems, it demonstrated strong rank correlation (Spearman $\rho = 0.911$ ) with DFT for binding Gibbs free energies, making it suitable for comparative screening and trend-based design.
Validation: DFT calculations were used to pre-validate the MLFF performance, confirming that while absolute errors existed, the relative ordering of materials was preserved.

Key Contributions

DSL-Guided Reasoning: The introduction of a strategy to translate LLM-generated natural language rules into executable symbolic predicates, bridging the gap between semantic reasoning and statistical verification.
ORCHESTRA Framework: A multi-agent system that continuously refines a rule store based on statistical evidence, preventing the accumulation of unsupported design heuristics.
Self-Verifying Extrapolation: The ability to propose rules that extrapolate beyond the current data regime and immediately test them in subsequent iterations via the DSL, a capability absent in purely natural language strategies.
Interpretable Design Principles: The framework generates design rules that are not only statistically grounded but also chemically interpretable, revealing motifs consistent with experimental literature.

Results

The study compared four strategies: Random (no agent), Background (LLM only, no rules), Rule-only (natural language rules without DSL), and DSL-guided (full framework).

Design Performance:
- For a joint band gap/formation energy objective, the DSL-guided strategy achieved the highest hit rate (~~0.681), outperforming the Rule-only strategy (~~0.642) and significantly outperforming the Background and Random strategies.
- For the more complex OER overpotential objective (requiring simultaneous optimization of three binding energies), the DSL-guided strategy showed a clearer advantage, achieving a hit rate of ~0.213, the highest among all strategies. The Rule-only strategy struggled more with this complex target, showing a decrease in lift over time, indicating a drift toward less reliable rules.
Rule Evolution:
- In DSL-guided runs, symbolic predicates allowed for the quantitative identification and removal of unreliable rules (e.g., rules with zero precision were pruned).
- The framework successfully identified and validated rules that initially had no matching data but were confirmed in subsequent iterations as new candidates were generated.
Chemical Plausibility:
- The framework identified 198 unique DP candidates with OER overpotentials $\le$ 0.3 V.
- These candidates consistently featured Co at the B-site and Ru/Ir at the B'-site, with Nd, Pr, and Sm at the A-sites. These motifs align with known experimental high-performance catalysts.
- Extracted rules explicitly captured these relationships (e.g., "Co on B and (Ru or Ir) on B'... -> lower OER_OP"), demonstrating the framework's ability to recover literature-consistent design principles.

Significance and Claims

The paper claims that mathematical and statistical grounding can significantly enhance the reasoning capability of LLM agents in materials science. By converting natural language hypotheses into executable predicates, the framework enables:

Robustness: The ability to detect and discard unsupported rules, preventing the degradation of design performance over iterative loops.
Efficiency: Effective inverse design can be performed without large task-specific datasets or additional model training, relying instead on the LLM's reasoning augmented by statistical feedback.
Interpretability: The generation of design principles that are both statistically validated and chemically meaningful.

The authors note limitations, particularly that the performance gain depends on the complexity of the target and the expressiveness of the DSL. For targets where the governing chemical trends involve concepts outside the current DSL (e.g., complex surface-specific motifs), the DSL-guided strategy may be less expressive than a purely natural language approach. However, the study concludes that the proposed approach offers a practical strategy for grounding and improving chemical reasoning, making LLM-based inverse design more reliable for challenging targets like OER overpotential.