🔬 materials science

LATHE: LAnguage-driven Toolkit for Hypothesis-based crystal Editing

This paper introduces LATHE, a language-driven toolkit that enables fine-grained, differentiable editing of crystal structures based on seven geometric and symmetry constraints, facilitating a closed-loop multi-agent system that successfully translates natural-language hypotheses into targeted materials design.

Original authors: Qianyu Zheng, Shuyi Jia, Victor Fung

Published 2026-10-05

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Qianyu Zheng, Shuyi Jia, Victor Fung

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 science is the art and science of understanding how the arrangement of atoms determines the behavior of the world around us. From the hardness of a diamond to the ability of a solar cell to capture light, these properties emerge from the precise geometry of crystal structures, which are repeating patterns of atoms held together by chemical bonds. For decades, scientists have struggled to design new materials with specific traits, such as a particular color or electrical conductivity, because the number of possible atomic arrangements is so vast that testing them one by one is impossible. Traditional methods often rely on brute-force computer simulations that are slow and expensive, while newer artificial intelligence tools have shown promise but often lack the precision to make small, controlled adjustments to an existing structure. The challenge has been to find a way to take a known crystal and tweak its shape—stretching a bond here, tilting an angle there—with the same certainty a sculptor has when refining a statue, all while ensuring the material remains physically stable.

A team of researchers at the Georgia Institute of Technology has developed a new system called LATHE that bridges this gap between human intuition and atomic precision. The system acts as a geometric toolkit that allows scientists to describe their desired changes in plain English, such as "lengthen this bond" or "increase this angle," and then automatically translates those words into the mathematical instructions needed to reshape the crystal. Unlike previous methods that try to generate entirely new structures from scratch or rely on large language models that often misunderstand spatial relationships, LATHE treats the crystal structure as a flexible object that can be edited directly. It focuses on seven fundamental geometric properties: the distance between atoms, the angles where bonds meet, the twist of atomic chains, the local neighborhood of each atom, the overall size and shape of the repeating unit, and the symmetry of the pattern. By turning these physical concepts into adjustable targets, the system can push a crystal toward a specific shape using a process similar to how a ball rolls down a hill to find the lowest point, but in this case, the "hill" is a landscape of geometric constraints.

The researchers built this toolkit to work alongside a large language model, which serves as the interface for human users. When a scientist types a request, the language model acts as a translator, breaking down the natural language instruction into a precise set of rules that the LATHE engine can execute. For example, if a user asks to increase the distance between two specific atoms, the system identifies those atoms, calculates the current distance, and then uses a mathematical optimization process to move the atoms until the new distance is reached. Crucially, the system does not just chase the geometric target; it also keeps an eye on the energy of the structure to ensure the atoms do not end up in an impossible or unstable configuration. This is achieved by using a machine learning model trained on known materials to act as a safety check, gently guiding the atoms away from configurations that would cause the crystal to fall apart. The result is a system that can faithfully carry out complex, multi-step instructions, such as changing the coordination number of an atom or adjusting the symmetry of the entire crystal, with a success rate of over 87 percent, even when the instructions are vague or ambiguous.

To prove the system works, the researchers tested it on a wide variety of tasks, ranging from simple adjustments like changing a single bond length to complex challenges like altering the symmetry of a crystal to a higher level. In tests involving bond lengths, angles, and other geometric features, the system achieved near-perfect success rates, meaning it almost always reached the exact target specified by the user. When they added the energy safety check, the system continued to succeed in meeting the geometric targets while also ensuring the final structure was physically realistic and stable. The researchers then took this capability a step further by embedding it in a closed-loop system where two artificial intelligence agents work together. One agent acts as a hypothesis generator, proposing ideas for how to change a material to achieve a desired property, such as a specific band gap for a solar cell. The second agent, equipped with the LATHE toolkit, translates those ideas into actual structural changes and runs the optimization. In a test case involving a material called Cu2CdGeS4, the team set a goal to triple the material's band gap, a measure of how well it can conduct electricity. Starting from a value of 0.309 electron volts, the system successfully navigated through a series of structural modifications, including tilting bond angles and changing the number of neighbors around specific atoms, to reach a target of 1.0 electron volts within a tolerance of 0.1. This was achieved in just thirteen to nineteen cycles of hypothesis and testing, demonstrating that the system can solve non-trivial design problems without human intervention.

The significance of this work lies in its ability to make the process of materials discovery more interpretable and controllable. In many previous approaches, artificial intelligence models acted as black boxes, producing results without explaining how they got there or why a specific change was made. With LATHE, every step of the process is tied to a clear, human-readable hypothesis. If the system changes a bond length, it is because a specific agent proposed that change to achieve a specific goal. This transparency allows scientists to trust the results and understand the reasoning behind them. The system also avoids the common pitfalls of other methods, such as the inability to make fine-grained edits or the tendency to generate structures that look correct but are physically impossible. By grounding the optimization in physical laws and using a differentiable approach that allows for smooth, continuous adjustments, the researchers have created a tool that can be applied to a broad class of functional material design tasks. The success of the multi-agent system in reaching a target band gap suggests that this approach could be scaled up to tackle more complex challenges in the future, potentially accelerating the discovery of new materials for energy, electronics, and other critical technologies. The work demonstrates that by combining the reasoning power of language models with the precision of geometric optimization, it is possible to create a new paradigm for computational materials discovery that is both powerful and understandable.

Technical Summary: LATHE – A Language-Driven Toolkit for Hypothesis-Based Crystal Editing

Problem Statement

The design of materials with targeted functional properties (e.g., specific band gaps, coordination geometries, or crystal symmetries) faces a significant bottleneck: the gap between qualitative structural intuition and quantitative geometric control. Existing computational frameworks fall short in addressing this challenge in two primary ways:

Large Language Models (LLMs): While capable of reasoning and generating hypotheses, LLMs lack spatial awareness and operate at a coarse granularity within token space. They struggle to execute precise, continuous geometric operations (e.g., bond stretching, angle tilting, symmetry enforcement) required for structure–property reasoning.
Generative and Optimization Models: Current approaches, including equivariant diffusion models, autoregressive language models, and gradient-based structure generation, generally focus on de novo generation or end-to-end property targeting. They lack a controllable, physically grounded mechanism for editing an existing structure in response to specific geometric hypotheses.

Consequently, there is no existing framework that allows a materials scientist to input a natural-language hypothesis about modifying a specific site or motif and receive a faithfully edited crystal structure that satisfies those constraints while remaining physically plausible.

Methodology

The authors present LATHE (LAnguage-driven Toolkit for Hypothesis-based crystal Editing), a geometric toolkit that bridges this gap by expressing crystallographic properties as differentiable objectives.

1. Differentiable Geometric Optimization

LATHE formulates crystal structure editing as a continuous optimization problem over two differentiable tensors: atomic fractional coordinates ( $F$ ) and the lattice matrix ( $L$ ). Cartesian positions are recovered as $X = F \cdot L$ . The toolkit defines loss functions for seven classes of properties:

Bond Length, Angle, and Dihedral Angle: Computed using the Minimum Image Convention (MIC) with numerically stable formulations (e.g., atan2 for angles) to avoid gradient singularities.
Coordination Environment: Uses a soft counting scheme based on Gaussian-smeared radial distribution integration to handle the non-differentiability of integer neighbor counts. Auxiliary penalties prevent unphysical geometries (e.g., atomic collapse).
Symmetry (Space Group): A composite loss term ( $L_{sg}$ ) includes symmetry compliance (matching transformed atomic positions to matching species via Hungarian matching), lattice shape constraints, volume regularization, and penalties for degenerate lattices or overlaps. To address convergence difficulties, atomic positions are initialized onto valid Wyckoff sites of the target space group.
Lattice Parameters and Volume: Optimized by minimizing deviations in edge lengths and interaxial angles, with volume loss formulated as absolute or relative squared deviations.
Energy Regularization: An auxiliary Machine Learning Interatomic Potential (MLIP), specifically MatterSim, is incorporated as a differentiable energy term ( $L_{energy}$ ). This acts as a regularizer to ensure optimized structures remain physically plausible without overriding the primary geometric objectives.

2. Optimization Algorithms

Single-Objective: Uses Adam-based gradient descent with cosine learning rate scheduling and early stopping once constraints are satisfied.
Multi-Objective: Employs Project Conflicting Gradients (PCGrad) to resolve destructive interference when gradients from competing property losses point in conflicting directions.
Space Group Optimization: Utilizes Basin-Hopping with random perturbations to escape local minima (e.g., supergroups) that gradient descent alone cannot escape.

3. Agent Integration and Workflow

LATHE is exposed via a Model Context Protocol (MCP) server, allowing external LLM agents to interact with it as a service.

LATHE-Equipped Agent: A Pydantic AI agent that translates natural-language instructions into validated JSON optimization configurations. It performs a multi-stage workflow: inspecting the input structure, consulting the MCP schema for appropriate loss functions, assembling a configuration, and verifying physical plausibility before execution.
Multi-Agent System: The framework integrates the LATHE-equipped agent with a Hypothesis Agent. The Hypothesis Agent proposes geometric modifications in natural language based on structure-property relationships. The LATHE-equipped agent translates these into executable configurations. The resulting structure is evaluated, and the (hypothesis, modification, outcome) triple is fed back to close the loop. This system supports both "manipulation" (geometric-only) and "design" (geometric + energy minimization) modes.

Key Contributions

Differentiable Geometric Toolkit: The first framework to express seven distinct classes of crystallographic properties as differentiable loss functions over fractional coordinates and lattice parameters, enabling direct gradient-based editing of existing structures.
Language-to-Geometry Bridge: The integration of LATHE with an MCP server and a multi-agent system that translates qualitative, natural-language hypotheses into executable, physically grounded structural modifications.
Interpretable Design Trajectory: A paradigm where every accepted structural change is associated with a human-readable hypothesis and a measured outcome, creating an auditable loop for computational materials discovery.
Robust Optimization Strategies: The implementation of PCGrad for multi-objective stability and Basin-Hopping for escaping local minima in symmetry-constrained optimization.

Results

Single-Property Benchmarks

LATHE was evaluated on 120 test cases per property type (78 for space group).

Constraint Satisfaction: In property-only settings, LATHE achieved near-perfect pass rates: 100% for bond length, bond angle, dihedral angle, coordination environment, and volume; 94.2% for lattice parameters; and 89.7% for space group.
Energy Awareness: When an auxiliary MLIP energy loss was added, pass rates remained largely preserved (e.g., 93.6% for space group, 90.0% for coordination environment). Crucially, energy-aware optimization significantly reduced the distance ( $|\Delta E|$ ) to local energy minima compared to property-only optimization, confirming the structures are physically plausible.

Agent Performance

Translation Accuracy: On a benchmark of 40 prompts (including "informative" prompts with deliberate ambiguity), the LATHE-equipped agent achieved an 87.5% success rate in translating natural language into valid executable configurations.
Baseline Comparison: This outperformed direct CIF generation by LLMs (45.0%) and the state-of-the-art coding agent Claude Code (72.5%). Most failures (96%) were due to optimization infeasibility rather than translation errors, indicating the language-to-configuration layer is robust.

Inverse Design Case Study

In a band-gap inverse-design task (targeting 1.0 ± 0.1 eV for $Cu_2CdGeS_4$ ):

The multi-agent loop successfully reached the target tolerance in 9 out of 10 independent runs.
The system converged in a median of 13 iterations (design setting) and 16 iterations (manipulation setting).
The "directional gate" (requiring that geometric intent survives MLIP relaxation) effectively filtered out brittle changes, guiding the search toward stable, energy-favorable structures.

Significance and Claims

The paper claims that LATHE establishes a new paradigm for interpretable, closed-loop computational materials discovery. By bridging natural-language hypothesis generation with physically grounded gradient-based structural editing, the framework allows for:

Fine-grained control: Moving beyond coarse elemental substitution to precise geometric operations (bond stretching, angle tilting, symmetry changes).
Iterative Design: Enabling workflows where scientists can investigate the impact of specific structural motifs on target properties without starting from scratch.
Scalability: The modular agent architecture allows the system to be immediately applicable to a broad class of functional material design tasks.

The authors emphasize that the system does not require task-specific fine-tuning; instead, it leverages in-context learning and a structured feedback loop to navigate complex structure-property landscapes. While the current framework is limited by the local nature of gradient-based optimization, it provides a foundational step toward systems that can reason about structure, execute changes with physical fidelity, and explain their trajectory in natural language.