TASTET: a Python package for geometry-driven selection of representative atomistic structures from large configurational spaces
This paper introduces TASTET, a Python package that employs geometry-driven space-filling algorithms to efficiently select diverse, representative subsets of atomistic structures from large configurational ensembles, thereby maximizing the information gained from costly first-principles calculations without relying on prior energy estimates.
Original paper licensed under CC BY 4.0 (https://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
In the world of chemistry and materials science, researchers are constantly trying to understand how atoms arrange themselves to create new molecules or materials with specific properties. To do this, they use powerful computer simulations based on the laws of quantum mechanics, which allow them to predict how a structure will behave without building it in a lab. However, these calculations are incredibly demanding. A single simulation can take hours or even days on a supercomputer. When scientists are exploring a new system, they often generate thousands of potential structures—different ways the atoms could be arranged—because they do not know which one is the most stable or useful. Trying to run a detailed, high-precision calculation on every single one of these possibilities is usually impossible due to the sheer amount of time and computing power required.
The traditional solution has been to guess which structures are worth studying based on rough, low-cost estimates of their energy. The idea is that the most stable structures will have the lowest energy, so scientists filter out the rest and only calculate the details for the most promising candidates. But this approach has a hidden flaw: it assumes that the low-cost guesses are accurate enough to tell the difference between a truly unique structure and a minor variation. If the guess is wrong, or if it misses a subtle but important geometric difference, researchers might waste their limited computing power on redundant structures while overlooking the very configurations they need to find. The challenge, then, is not just to reduce the number of structures, but to pick a small, representative group that captures the full variety of possibilities, ensuring that no important region of the atomic landscape is left unexplored.
A team of researchers has developed a new approach to solve this problem, shifting the focus from energy estimates to the actual shape of the molecules. They created a software tool called TASTET, which acts as a smart filter for these vast collections of atomic structures. Instead of asking "which of these looks like it has the lowest energy?", the software asks "which of these looks the most different from the others?" It treats the selection process as a geometry problem. The program analyzes the local neighborhood of every atom in a structure, mapping out how the atoms are positioned relative to one another. It then compares these maps across the entire collection to measure how similar or different each structure is from every other one. Using a mathematical strategy designed to fill a space evenly, the software selects a subset of structures that are as far apart from each other as possible in terms of their shape. This ensures that the final group covers the entire range of possibilities, from the most compact arrangements to the most stretched-out ones, without wasting resources on duplicates.
The researchers tested this method on two very different types of systems to see if it could handle the complexity of real-world chemistry. The first test involved a flexible molecule containing a rhodium metal atom, surrounded by a large organic ligand. Because the ligand can twist and rotate in many ways, the computer generated nearly 950 different shapes for this single molecule. The researchers wanted to see if they could find the most important shapes using only a tiny fraction of their computing budget. They used TASTET to select just 28 structures, which is only 3% of the total. When they ran the expensive, high-precision calculations on these 28, the results revealed the main energy patterns of the entire landscape. They found that the vast majority of the energy differences came from just four specific twisting motions. By focusing on this small, diverse group, they recovered the most important energetic regions without needing to calculate the other 921 structures. When they compared this to the traditional method of filtering by low-cost energy estimates, the old method failed to capture the same diversity, often picking many structures that were essentially the same shape and missing others that turned out to be important.
The second test pushed the method even further, moving from a single molecule to a cluster of metal atoms sitting on a solid surface. This is a common setup in catalysis, where materials are used to speed up chemical reactions. The researchers generated a massive collection of 10,000 different arrangements for a cluster of twelve copper atoms on a zinc oxide surface. In this scenario, the structures were generated by an advanced search algorithm that uses a machine-learning model to predict energies. The traditional approach would be to pick the structures that the machine learning model predicted to have the lowest energy. However, the researchers found that the machine learning model was not accurate enough to make this decision reliably. When they selected 30 structures based purely on geometric diversity using TASTET, and then calculated their true energies with high-precision methods, they discovered a startling mismatch. The structure that the machine learning model thought was the best was actually the ninth best when calculated precisely, while the true best structure was ranked 28th by the model. The errors in the model were so large that a standard energy-based filter would have missed the best structure entirely.
By relying on geometry rather than the flawed energy predictions, the researchers were able to select a tiny fraction of the total pool—just 0.3%—that still contained the full range of structural behaviors. They found that the most stable structures were those where the metal cluster spread out flat against the surface, while the less stable ones stood up tall and thin. This clear relationship between shape and stability emerged only because the selection method did not bias the results toward the incorrect energy predictions of the machine learning model. The study demonstrates that when high-precision calculations are too expensive to run on everything, the best way to spend that limited resource is to ensure the selected samples are geometrically diverse. This approach allows scientists to explore the full landscape of possibilities, uncovering the true relationships between structure and energy even when their initial guesses are imperfect. The TASTET tool is now available as open-source software, offering a practical way for researchers to navigate these complex chemical spaces without getting lost in redundant data or relying on unreliable shortcuts.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.