MASON: Materials Atomistic Structure builder Operated in Natural language

In the world of materials science, researchers often need to build tiny, perfect models of matter to understand how things work at the atomic level. Imagine trying to predict how a new battery will hold a charge or how a drug will interact with a cell; scientists do this by creating digital versions of atoms arranged in specific patterns. These patterns, known as crystal structures, are the blueprints for everything from the silicon in computer chips to the metals in jet engines. To study them, scientists use powerful computers to run simulations, but before the computer can start working, a human must first construct the model. This is a delicate task. If the atoms are placed even slightly wrong, the entire simulation fails, and the results become meaningless. For decades, this step has required a researcher to write complex computer code or manually manipulate 3D shapes, a process that is slow, prone to human error, and difficult for the new generation of artificial intelligence to handle on its own.

A team of researchers has now introduced a new tool called MASON, designed to bridge this gap between human intent and digital construction. The system allows scientists to describe the structure they need using ordinary English, and an artificial intelligence agent translates those words into a precise, error-free atomic model. Instead of asking the AI to write code that might contain mistakes, MASON provides the AI with a dedicated set of specialized tools it can use directly. Think of it as giving a builder a set of pre-made, perfectly calibrated bricks and a hammer, rather than asking them to mix the cement and shape the bricks from scratch every time. The researchers demonstrated that this approach works reliably, allowing an AI to take a simple request like "build a model of carbon monoxide sticking to a platinum surface" and produce a complex, scientifically valid structure ready for simulation.

The core challenge the researchers addressed is that while modern AI can hold a conversation and reason through problems, it struggles with the specific, rigid rules of crystallography. In the past, if a scientist wanted an AI to build a model, the AI would have to write a program to do it. This is risky because a tiny error in the code could result in atoms overlapping or a structure that doesn't exist in reality. The MASON system changes this dynamic by acting as a specialized server that sits between the AI and the construction process. When a researcher asks a question, the AI breaks the request down into a sequence of steps and calls upon MASON to perform each one. MASON then executes the operation, such as cutting a slice of a crystal or adding a missing atom, and immediately reports back exactly what it built. This feedback loop ensures that if a step goes wrong, the AI sees the result and can correct its next move before the model is finalized.

The system is organized into five main categories of tools that cover the entire lifecycle of building a model. First, it can analyze existing structures to check if they are valid, counting the atoms and measuring the distances between them. Second, it can create larger versions of a crystal or introduce specific flaws, like removing an atom to simulate a defect. Third, it can slice a block of material to create a flat surface, which is essential for studying how chemicals interact with metals. Fourth, it can place molecules onto those surfaces in precise locations. Finally, it can combine two different materials to see how they fit together, a process known as creating a heterostructure. In every case, the tool does the heavy lifting of the geometry, ensuring the atoms are spaced correctly and the symmetry is maintained.

To prove the system works, the researchers walked through a complete example of building a model for carbon monoxide adsorbed on a platinum surface. The process began with the AI asking MASON to read the basic structure of bulk platinum and verify its shape. Next, the AI instructed MASON to cut a thin slice of this material, specifically a layer four atoms thick, and to fix the bottom two layers in place to mimic a solid block. The system then expanded this slice to create a larger surface area and identified the specific spots where a carbon monoxide molecule could attach. Finally, the AI directed MASON to place the molecule exactly two angstroms above a specific hollow spot on the surface. Throughout this entire sequence, which involved more than ten separate tool calls, the AI received constant updates on the dimensions and arrangement of the atoms. The final result was a complete, three-dimensional model that passed every geometric check, ready for a supercomputer to run a simulation.

The researchers also tested the system on a wide variety of other challenges to ensure its reliability. They successfully built models of graphene nanoribbons, twisted layers of carbon atoms, and complex interfaces between silicon carbide and silicon. In each instance, the AI used the natural language request to guide the construction, and MASON provided the precise execution. For example, when asked to build a model of a silicon crystal with a missing atom, the system correctly identified how the symmetry of the crystal changed and reported the new arrangement of the remaining atoms. When asked to combine two different materials, the system calculated the best way to align their atomic grids to minimize stress, presenting the options to the AI for selection. In all eighteen examples tested, the structures built by the AI were mathematically consistent and geometrically sound, matching the results of manual calculations.

This work represents a significant step toward making scientific research more autonomous. By equipping AI agents with reliable, specialized tools, the researchers have removed the need for the AI to write its own code for basic structural tasks. This reduces the chance of error and allows the AI to focus on the scientific reasoning rather than the mechanical details of construction. The system is open-source, meaning other scientists can use it to build their own models, and it is designed to work with any AI framework that supports the Model Context Protocol. The ultimate goal is not just to build models faster, but to create a workflow where a researcher can state a scientific question in plain language, and the AI can handle the entire process of setting up the experiment, from the initial atomic arrangement to the final data analysis. While the current version focuses on building structures, the same approach could eventually be applied to running the simulations and analyzing the results, moving the field closer to a future where scientists define the goals and the machines handle the execution.