🔬 materials science

IRIS: an agentic eye on symmetry-derived paths and energy landscapes of structural phase transitions

The paper introduces IRIS, an agentic workflow that automates the construction of interpretable energy landscapes and Landau free energy models for structural phase transitions by converting plain-language requests and crystal structures into symmetry-adapted computational results through deterministic verification gates and language model-guided recovery.

Original authors: Sukriti Manna, Sraddha Agrawal, Ramamoorthy Ramesh, Panchapakesan Ganesh, Pierre Darancet, Subramanian K. R. S. Sankaranarayanan

Published 2026-11-03

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Sukriti Manna, Sraddha Agrawal, Ramamoorthy Ramesh, Panchapakesan Ganesh, Pierre Darancet, Subramanian K. R. S. Sankaranarayanan

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

Materials scientists often study how the tiny building blocks of a solid rearrange themselves to create new properties. Imagine a crystal as a rigid city of atoms. Sometimes, when conditions change, the entire city shifts its layout, moving from one stable arrangement to another. This is a structural phase transition. These shifts are the engine behind some of the most useful materials we have, such as those that generate electricity when squeezed or change their magnetic properties with a flick of a switch. To understand and predict how these materials behave, researchers must map the energy landscape of the transition. Think of this landscape as a topographical map of hills and valleys, where the valleys represent stable arrangements of atoms and the hills represent the energy barriers that must be overcome to switch from one state to another. Creating an accurate map of this terrain has traditionally been a slow, manual process that requires deep expertise in crystallography and complex computer simulations. A single mistake in the setup—such as a slight error in how the atoms are ordered or how the simulation is run—can produce a smooth, beautiful map that is completely wrong, leading scientists to draw false conclusions about the material's capabilities.

A team of researchers has now introduced a new system called IRIS, which acts as an automated guide to build these energy maps with a level of precision and reliability that was previously difficult to achieve. The system is designed to take a simple request from a scientist, such as "map the energy path for this specific material," and turn it into a detailed, verified energy landscape. Instead of relying on a single artificial intelligence to do everything, the researchers built a workflow where a large language model acts as a coordinator, but every single step of the process is checked by strict, unchangeable rules. The model proposes what to do next, but it cannot make a final decision on whether a result is correct. Before any data moves forward, a series of automated gates verify that the atomic structures are physically possible, that the calculations have converged properly, and that the final mathematical model fits the data. If a step fails, the system does not guess; it consults a list of known fixes or asks a human expert for guidance, ensuring that no errors slip through unnoticed.

The researchers tested this system on six different materials, ranging from simple crystals to complex compounds used in advanced electronics. In one notable test, the system successfully mapped the energy landscape of a material called bismuth ferrite, which involves six different ways the atoms can move simultaneously. This is a task so complex that it usually requires a team of experts to manage the many variables. IRIS handled the entire process, generating thousands of atomic structures, running the necessary computer simulations, and fitting the final energy model without any human intervention after the initial request. In a separate live test on a material called barium titanate, the system fetched the starting structures from a public database, performed the symmetry analysis, submitted the calculations to a high-performance computer cluster, and collected the results in just 34 minutes. Throughout this entire process, every single gate passed on the first try, and the system produced a complete, verified map of the material's behavior.

To ensure the system was truly robust, the researchers deliberately injected 53 different types of errors into the workflow, such as incorrect settings or incomplete data. The automated gates caught every single one of these faults. When the gates were turned off, the system allowed 14 of those errors to pass silently, resulting in incorrect maps. This experiment proved that the safety checks, not just the artificial intelligence, are what make the system reliable. The language model proved excellent at interpreting the scientist's request and deciding which known fix to apply when a problem arose, but it was the deterministic gates that prevented wrong results from being accepted. The system successfully separated the competition between different atomic movements from the locking of their directions, revealing physical insights that were previously hidden. By combining the flexibility of language-based requests with the rigor of automated verification, IRIS transforms the construction of physical models from a specialized, error-prone task into a reusable, automated capability, allowing scientists to explore the energy landscapes of new materials with confidence and speed.

Technical Summary: IRIS – An Agentic Workflow for Symmetry-Derived Energy Landscapes

Problem Statement

Structural phase transitions underpin the functionality of ferroelectrics, multiferroics, and correlated oxides. Understanding these transitions requires constructing interpretable energy landscapes in the space of symmetry-adapted order parameters. These landscapes provide the coefficients for Landau and phase-field models, define switching pathways, and classify transitions. However, generating these landscapes from first principles is currently a bottleneck that relies heavily on specialist expertise. The process involves complex steps: crystallographic analysis to identify irreducible representations (irreps), reliable generation of symmetry-derived structural paths, extensive density-functional theory (DFT) sampling, and model fitting.

Existing workflow engines (e.g., AiiDA, atomate) automate execution and provenance but lack the domain knowledge to make critical crystallographic decisions, such as determining the dimensionality of the sampling grid or verifying that generated structures match the intended symmetry. Furthermore, mismatches in conventions (origin, atom ordering, normalization) can produce smooth but physically incorrect landscapes. While Large Language Model (LLM) agents show promise in scientific automation, they often produce fluent but physically wrong outputs and lack reliable self-correction mechanisms. The challenge is to automate this complex, multi-step scientific workflow while ensuring rigorous physical correctness and explicit verification.

Methodology: The IRIS Framework

The authors introduce IRIS (Irrep Route and Interpolation System), an agentic workflow designed to convert plain-language requests and crystal structures into symmetry-adapted energy landscapes and fitted Landau models. The core design principle is verifier-guided reasoning: an LLM may propose actions, but deterministic checks (gates) decide whether a result is accepted.

Architecture

IRIS is implemented as a gated state graph using the LangGraph library, comprising seven agent nodes ( $f_0$ – $f_6$ ) and a recovery node. Each node is followed by a deterministic gate ( $G_0$ – $G_6$ ) that validates the output before the workflow proceeds.

$f_0$ Intake (LLM): Parses plain-language requests into a structured configuration (system, parent/child structures, analysis type, budget). It is the only node that directly calls an LLM.
$f_1$ Crystallographer: Performs symmetry-mode analysis. It standardizes structures, establishes the group-subgroup relationship ( $H \subset G$ ), and decomposes the distortion into irrep modes using ISODISTORT.
$f_2$ Path Planner: Selects active modes based on explicit policies (R1–R5) and builds a certified grid of structures. It ensures every frame has a known space group derived from the group-subgroup tree.
$f_3$ Runner: Submits DFT calculations (e.g., VASP) to an HPC cluster and monitors job completion.
$f_4$ Reviewer: Checks energy convergence and flags SCF-limited runs.
$f_5$ Analyst: Fits invariant Landau polynomials to the energy data and locates minima and barriers.
$f_6$ Skeptic: Independently re-derives every reported number from the raw data to verify consistency.

The Gate and Recovery Mechanism

The workflow's robustness relies on deterministic gates ( $G_i$ ) that return either pass or fail(c) with a specific error code.

Recovery Node: If a gate fails, the workflow pauses and routes to a recovery node. The recovery logic follows a strict hierarchy:
1. Stored Rules: A rule table ( $R$ ) attempts to apply a pre-defined fix (e.g., resubmitting a failed job, removing a rigid translation).
2. Sub-agent (LLM): If no rule applies, a restricted LLM sub-agent chooses a fix from a pre-listed menu of options generated by the code. The LLM cannot invent new actions.
3. Human Intervention: If the sub-agent cannot choose or the menu is empty, the workflow pauses for human input. The human's decision is then stored as a new rule.
Verification: The LLM never routes the workflow, submits calculations, or accepts results. It only interprets requests or selects from a constrained set of recovery actions. Every result is re-checked by the same gate after a fix is applied.

Key Contributions

Agentic Workflow with Explicit Verification: IRIS demonstrates a framework where LLMs are used for interface and recovery selection but are strictly bounded by deterministic crystallographic and physical checks. This prevents the propagation of "fluent but wrong" errors common in autonomous agents.
Symmetry-Adapted Structure Generation: The system automates the generation of energy landscapes on high-dimensional order-parameter spaces (up to six dimensions) by ensuring every generated structure is certified against the group-subgroup tree and symmetry constraints.
Recovery Benchmarking: The authors provide a rigorous evaluation of the agentic layer using 53 injected faults across six material systems. They demonstrate that deterministic gates are essential for catching errors, while stored rules and constrained LLM sub-agents effectively handle recoverable failures.
Prospective Automation: The system successfully executed a "live" campaign on BaTiO3, fetching structures from the Materials Project, running symmetry analysis, submitting DFT jobs, and producing a fitted landscape without human intervention beyond the initial request and submission confirmation.

Results

The paper validates IRIS on six compounds with increasing complexity:

CsSnI3 & Ca3Ti2O7: Validated structure generation on 2D irrep planes. IRIS generated 961 and 1,372 frames respectively, all with correct space groups. Certification caught errors (e.g., rigid translations, mis-scaled vectors) that symmetry checks alone missed.
Hexagonal Manganites (InMnO3, ErMnO3): Successfully modeled the improper ferroelectric ground state, capturing the U(1)-like "Mexican hat" potential and the trilinear coupling between the trimerization mode and polarization.
BiFeO3: Constructed a six-dimensional energy landscape coupling polar ( $\Gamma^-_4$ ) and tilt ( $R^-_5$ ) order parameters. The system identified that biquadratic couplings ( $\lambda_1 P^2 R^2$ ) compete the magnitudes, while $(P \cdot R)^2$ locks the axes. It ranked homogeneous switching pathways, showing that rotating polarization with the tilt following is energetically cheaper than reversing polarization at a fixed tilt.
HfO2: Demonstrated limitations in mapping polymorph families where phases are not in a simple group-subgroup relation, highlighting the need for a common parent basis for global energy surfaces.
BaTiO3 (Live Run): In a prospective run, IRIS processed requests for polar paths and a 2D plane, performing 96 DFT calculations in 34 minutes total. All gates passed on the first attempt, and the system produced fitted landscapes consistent with literature.

Evaluation of the Agentic Layer:

Fault Injection: Of 53 injected faults, all were caught by deterministic gates. Without gates, 14 errors resulted in silent, wrong results.
Recovery: Stored rules alone solved 96% of benchmark cases. When rules failed, LLM sub-agents (restricted to a menu) chose the correct fix in 85–97% of cases (with admissible menus).
Intake: LLMs outperformed rule-based parsers and keyword lookups in interpreting plain-language requests, especially for non-specialist or terse inputs, achieving 92–95% accuracy in generating valid configurations.

Significance and Claims

The paper claims that IRIS makes the construction of physical models a reusable computational capability. By grounding automation in explicit verification (gates) and limiting the LLM's role to proposal and constrained selection, the system achieves a level of reliability suitable for scientific discovery.

Scientific Insight: The generated landscapes provide physical insights, such as separating the competition of polar/tilt magnitudes from the locking of their axes in BiFeO3, and confirming that the domain barrier in hexagonal manganites arises from polar coupling rather than angular terms.
Automation Framework: IRIS provides a framework for scientific automation where the "safety" of the workflow is enforced by code (gates and rules) rather than relying on the LLM's internal reasoning. This addresses the risk of agents taking risky actions or propagating errors.
Limitations: The authors acknowledge that the system currently handles pairwise transitions well but faces challenges with polymorph families requiring a unified global basis (as seen in HfO2). Additionally, the system defaults to fixed-lattice calculations, with strain coupling handled via explicit interpolation or relaxed anchors, and does not yet automatically handle complex spin-lattice couplings requiring spin-orbit interaction.

In summary, IRIS represents a shift from "autonomous" agents that plan and execute freely to "assisted" agents that operate within a rigorously verified, deterministic framework, enabling the automated construction of complex, symmetry-derived energy landscapes.