Energetically Driven Structure Matching for Autonomous Total X-ray Scattering Experiments

In the modern quest to discover new materials, scientists are increasingly turning to laboratories that run themselves. These "self-driving laboratories" use robots to mix chemicals, powerful computers to analyze results, and artificial intelligence to decide what to do next, all in a continuous loop. The goal is to speed up the process of finding materials with specific properties, such as better batteries or more efficient catalysts. However, a significant hurdle remains: knowing exactly what the robot has actually created. When a machine synthesizes a tiny particle, the computer might predict a perfect shape, but the real particle could be slightly different due to the messy realities of heat, surface tension, and atomic stress. If the system cannot accurately read the true structure of the material it just made, it cannot learn effectively or guide the next step with precision. This gap between the computer's idealized model and the physical reality of the atom is where the next breakthrough in autonomous science must happen.

A team of researchers at the Technical University of Denmark and the IT University of Copenhagen has developed a new way to bridge this gap, specifically for experiments involving gold nanoparticles. These tiny particles, often just a few nanometers wide, are crucial for many technologies, but their internal arrangement of atoms determines how they behave. The researchers focused on a technique called total X-ray scattering, where a beam of X-rays hits the particles and bounces off in a pattern that reveals their atomic structure. In previous autonomous experiments, the system would compare the experimental X-ray pattern to a library of computer-generated patterns based on perfect, geometric shapes like spheres or octahedrons. While fast, this method often led to mistakes. The computer would tell the robot it had made a specific shape, only to find later that the particle was actually a different size or form because the computer model ignored how atoms shift and stretch when they are free on the surface of a tiny particle.

To solve this, the team introduced a method that brings energy and heat into the computer models. Instead of just looking at static, perfect shapes, they used a sophisticated computer program based on machine learning to simulate how the atoms in the gold particles would actually relax and move. They tested three levels of detail. The first was the simple, idealized shape, which is quick to calculate but often wrong. The second involved letting the computer "relax" the atoms, allowing them to shift into a more stable, lower-energy position, which corrected many of the size errors. The third and most detailed level involved simulating the particles as if they were vibrating with heat, creating a dynamic cloud of possible shapes rather than a single frozen image. When they compared these three approaches against real X-ray data from gold nanoparticles, the results were clear. The simple models consistently underestimated the size of the particles and sometimes confused one shape for another. The models that accounted for atomic relaxation were much better, but the models that included the thermal motion of atoms matched the real experimental data almost perfectly.

The researchers discovered that the simple models failed because they treated the particles as rigid blocks, ignoring the fact that surface atoms are less constrained and often pull inward or shift to relieve stress. This distortion changes the way X-rays scatter, leading the computer to guess the wrong size or shape. By using the machine-learned models to account for these shifts, the system could distinguish between similar-looking shapes, such as a decahedron (a ten-sided shape) and an octahedron (an eight-sided shape), which the simple models often mixed up. Furthermore, the team found that simulating the particles in a vacuum at a slightly lower temperature produced results nearly identical to simulating them in water, which is a major practical advantage. It means the computer can run these complex, heat-aware simulations much faster without needing to model the surrounding liquid, making the process feasible for real-time use.

To make this high level of accuracy work within the tight time limits of an autonomous experiment, the team designed a "funnel" strategy. Imagine a process that starts with a wide net and quickly narrows down to the most promising candidates. First, the system uses the fast, simple models to scan a broad range of possibilities and identify a few likely shapes and sizes. It then takes only those top candidates and runs the more detailed, energy-aware relaxation simulations to refine the list. Finally, it applies the most computationally expensive, heat-simulated models only to the very best remaining options. This hierarchical approach allows the system to get the most accurate structural information possible without waiting hours for the computer to finish. In tests, this method could identify the best-fitting structure within a fifteen-minute window, a timeframe that matches the cycle of a real-world autonomous synthesis experiment.

The implications of this work extend beyond just gold nanoparticles. By proving that energetically driven structure matching can run in parallel with synthesis, the researchers have shown a path toward laboratories that do not just follow a pre-set plan but can adapt their goals based on what they actually find. If the system realizes that a specific shape is impossible to make under current conditions, or that a slightly different shape is energetically more stable, it can update its target in real time. This moves autonomous science from simply trying to hit a static target to exploring the landscape of what is physically possible. The study demonstrates that by combining rapid screening with targeted, high-fidelity simulations, scientists can finally give self-driving laboratories the ability to see the true atomic structure of the materials they create, turning a blind guess into a clear, informed decision.