Agentic inverse design of spatially programmed materials

Materials scientists have long treated the internal structure of a solid as a fixed landscape, determined by the chemical ingredients mixed together and the heat or pressure applied during creation. While they can tune a material's properties by changing its composition or applying external forces like magnetic fields, they have struggled to write a specific, custom pattern directly into the material's interior. Imagine trying to carve a unique design inside a block of stone without breaking the surface; for decades, this has been nearly impossible because the tools to shape matter from the inside out are blunt and imprecise. The goal of "spatially programmed materials" is to overcome this by arranging atoms or defects in a deliberate, three-dimensional pattern, turning the depth of a material into a design variable just like its color or hardness. This approach promises to unlock new physical behaviors, such as guiding light or electricity in ways that uniform materials cannot, but the path to creating these structures has been blocked by a difficult problem: figuring out exactly how to build them.

The challenge lies in the gap between a desired design and the physical steps needed to achieve it. To create a specific internal pattern, a researcher must choose a sequence of actions, such as firing beams of ions at a target, but predicting the result of those actions is complex. Traditionally, a human expert would have to guess the right combination of beam energy, angle, and duration, run a computer simulation to see the result, and then guess again if the outcome was wrong. This trial-and-error loop is slow and relies heavily on human intuition. In a new study, researchers at the Massachusetts Institute of Technology have shown that an artificial intelligence agent can close this loop on its own. By connecting a large language model to a physics simulator, they created a system that can look at a desired internal pattern and automatically figure out the exact sequence of steps needed to build it, without needing a human to guide every move.

The researchers tested this system using ion implantation, a technique where beams of charged atoms are fired into a material to alter its properties. They gave the AI agent a simple goal: create a specific depth profile of ions or empty spaces within a solid. In one experiment, the researchers drew a rough sketch of a desired pattern by hand—a shape that rose to a plateau, stepped up, and then leveled off again. They handed this sketch to the agent. The agent first converted the drawing into a digital target, then began to plan. It selected a type of ion, in this case bismuth, and a target material, silicon. It then ran a physics simulation to see how ions of different energies would travel through the silicon. The agent realized that a single beam energy would only create a small, localized bump in the pattern. To fill in the entire shape, it needed to combine many different beam energies.

The agent began to iterate. It chose a set of energies, ran the simulation, and compared the result to the target sketch. Where the simulation fell short, the agent added new energies or adjusted the existing ones. It did this repeatedly, refining its plan until the simulated pattern matched the hand-drawn target with high precision. In the end, the agent determined that it needed to fire ions at 28 different energy levels, ranging from 2.0 to 195.8 million electron volts, to recreate the shape. The final result was a match that was off by less than 8 percent, a level of accuracy that would have required a human expert many hours of manual calculation and adjustment to achieve.

The system proved equally capable with more complex designs. The researchers asked the agent to create a "superlattice," a pattern of repeating peaks and valleys, and a "Fibonacci superlattice," a pattern that follows a mathematical sequence where the spacing between peaks changes in a specific, non-repeating way. For the repeating pattern, the agent successfully spaced five peaks exactly 5 micrometers apart. For the Fibonacci pattern, it arranged the peaks to follow the correct mathematical ratio, preserving the unique order even as the pattern grew deeper into the material. The agent also tackled a different kind of challenge: creating a specific pattern of damage, or vacancies, rather than just adding ions. It successfully designed a recipe to create a gradient of empty spaces across a boundary between two different materials, silicon and germanium, showing that it could control not just where ions land, but the damage they leave behind.

One of the most practical tests involved the real-world limits of laboratory equipment. In a real accelerator, there is a maximum energy limit; the machine cannot fire ions faster than a certain speed. The researchers asked the agent to design a pattern that extended deeper than this limit allowed. Instead of failing or ignoring the constraint, the agent adapted. It realized it could not reach the deepest parts of the target, so it focused its efforts on the section it could reach. It then converted its list of discrete energy steps into a single, smooth ramp of changing energy, a plan that a human operator could follow to sweep the beam from low to high energy continuously. This demonstrated that the agent could not only solve abstract puzzles but also generate practical, executable instructions for real machines.

The researchers also pushed the system to its absolute limits by asking it to write the letters "MIT" into the depth of a material. The agent had to figure out how to arrange the ions so that, when viewed from the side, the concentration of atoms would form the shape of the letters. It succeeded in writing the "M" and the "I" clearly, but the "T" remained blurry. This failure was not a flaw in the agent but a fundamental limit of physics: the ions spread out slightly as they travel, a phenomenon known as straggling, which blurs sharp edges at certain depths. The agent respected this physical reality, showing that it understands the laws of nature it is trying to manipulate.

This work marks a shift in how materials are designed. Instead of a human guessing the steps to build a structure, the AI agent acts as a bridge, translating a visual or mathematical idea directly into a physical recipe. The system does not invent new physics; it uses established simulation tools to explore the vast space of possible combinations far faster than a human could. By automating the inverse design process, the researchers have shown that the spatial structure of a material can become a programmable variable. This opens the door to creating materials with custom internal architectures, potentially leading to new types of electronic devices, radiation-hardened components, or quantum materials with properties that have never been seen before. The agent's ability to handle complex, multi-step tasks with minimal human input suggests a future where the design of matter is limited only by the imagination of the designer and the laws of physics, rather than the difficulty of the calculation.