🔬 materials science

Data-Efficient Training of Linear ACE Potentials through Leverage-Guided Subset Selection of ASSYST Structure Pools

This paper demonstrates that leverage-guided, label-free subset selection of ASSYST structure pools can significantly reduce the DFT labeling workload (by 2–3x) required to train accurate linear Atomic Cluster Expansion potentials while maintaining competitive energy, force, and defect-level fidelity compared to random and other baseline sampling strategies.

Original authors: Aynour Khosravi, Marvin Poul, Jörg Neugebauer, Chad Sinclair

Published 2026-07-22

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Aynour Khosravi, Marvin Poul, Jörg Neugebauer, Chad Sinclair

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

The Great Material Hunt: Why We Need Smarter Maps

Imagine you are an architect trying to build a skyscraper, but instead of blueprints, you have to calculate the physics of every single brick, bolt, and beam from scratch every time you want to add a new floor. That is essentially what scientists face when they try to simulate how materials behave. To understand why a metal bends, why a battery degrades, or how a new alloy might hold up under pressure, they need to know the "potential energy surface"—a giant, invisible map that tells them how much energy is stored in every possible arrangement of atoms.

The most accurate way to draw this map is using a super-precise method called Density Functional Theory (DFT). Think of DFT as a master cartographer who can draw a map with perfect detail, but it takes them days to draw just a single square mile. If you want to simulate a whole city (or a chunk of metal with billions of atoms), you'd need to wait longer than the universe has existed. To solve this, scientists use "Machine-Learned Interatomic Potentials" (MLIPs). These are like student cartographers who learn from the master's maps. Once trained, they can draw the map almost instantly, allowing us to simulate massive systems. But here's the catch: to train the student, the master still has to draw thousands of those tiny, expensive square miles first. The question is: how many square miles does the master actually need to draw before the student gets it right?

The Paper's Big Idea: Finding the "Golden Spots"

This paper tackles that exact problem. The researchers, working with a tool called ASSYST (which automatically generates a massive library of random, weird, and wonderful atomic structures), asked a simple question: If we have a pool of 20,000 potential structures, do we really need to pay the "DFT tax" to label all of them? Or can we pick a smaller, smarter subset that teaches the computer model just as well?

They tested a clever strategy called Leverage-Guided Selection. Imagine you are trying to learn the layout of a new city. You could walk down every single street (random sampling), or you could look at a map and pick the streets that connect the most unique neighborhoods (leverage selection). The paper's method uses math to find the "most unique" atomic arrangements in the pool—those that fill in the biggest gaps in the model's knowledge—without ever needing to know the expensive energy values first. They call this "label-free" because it only looks at the shape of the atoms, not their calculated energy.

What They Found:
The results were surprisingly efficient. For pure metals like Aluminum and Copper, the researchers found that by using this "smart picking" method, they could train a model using only 30–40% of the data that a random pick would require to reach the same level of accuracy. In other words, they achieved the same high-quality map by doing only one-third of the expensive work. This is a 2 to 3 times reduction in the computational cost.

The "Gotcha" and the Nuance:
The paper is careful to note that this isn't magic. While the "smart picking" worked wonders for pure elements, the results for complex alloys (mixtures of Aluminum and Copper) were a bit more nuanced.

For ordered alloys: Once the model saw enough different types of chemical environments, picking 25% of the data with the smart method gave results just as good as picking 50% randomly.
For dilute alloys (a little bit of one metal in another): The smart method shined even brighter, cutting the required data in half while actually improving the accuracy for specific defects like missing atoms (vacancies).

What They Ruled Out:
The study explicitly argues against the idea that you just need to pick the "hardest" or "weirdest" looking structures based on how wrong a previous guess was (a method called "active learning" based on energy or force errors). They found that while picking based on big errors helped a little at the start, it was unstable and didn't cover the whole "map" as well as their leverage method. The paper suggests that simply looking for the biggest mistakes isn't as effective as looking for the most informative gaps in the geometry.

How Sure Are They?
The authors are very confident in these numbers because they tested them rigorously. They didn't just guess; they ran the simulations five times with different random seeds to make sure the results weren't a fluke. They also validated their models against completely independent sets of data (structures they hadn't seen before) and checked specific, difficult scenarios like grain boundaries and vacancies. The paper suggests that this "geometry-first" approach is a robust, statistically sound way to save time and money, but it notes that for extremely complex, multi-component alloys, more testing is needed to see where the limits lie.

In short, the paper shows that if you want to teach a computer to understand materials, you don't need to show it every single example. You just need to show it the right ones. By using a mathematical compass to find the most unique atomic shapes, scientists can build better models with significantly less expensive computing power.

Technical Summary: Data-Efficient Training of Linear ACE Potentials through Leverage-Guided Subset Selection

Problem Statement
The development of machine-learned interatomic potentials (MLIPs), particularly those based on the Atomic Cluster Expansion (ACE), is often constrained by the high computational cost of generating large Density Functional Theory (DFT) training datasets. While frameworks like the Automated Small SYmmetric Structure Training (ASSYST) can systematically generate diverse pools of candidate structures (often tens of thousands), the question of how many configurations must be labeled with expensive DFT calculations to achieve reliable accuracy remains unresolved. Existing data-efficient strategies often rely on active learning (requiring repeated retraining) or label-driven heuristics (requiring DFT labels to rank candidates), which do not fundamentally reduce the initial labeling cost. Furthermore, previous studies on data selection have not thoroughly investigated whether reducing training set size compromises the transferability of the resulting potentials to physically distinct regimes, such as defect energetics or alloy systems.

Methodology
The authors propose a label-free, geometry-based subset selection strategy for training linear ACE potentials. The core methodology relies on statistical leverage scores derived from the ACE design matrix, which quantify the influence of specific configurations on the regression fit without requiring access to quantum-mechanical labels (energies or forces).

Data Generation: The study utilizes pre-generated ASSYST structure pools for elemental Aluminum (Al), Copper (Cu), and Al–Cu alloys. These pools contain small periodic cells (8–10 atoms) generated by sampling random crystal prototypes, relaxing them, and applying controlled perturbations.
Model Architecture: Linear ACE models are employed, where the site energy is expressed as a linear combination of symmetry-adapted basis functions. The fitting problem is formulated as a Tikhonov-regularized Bayesian linear regression.
Leverage-Based Selection:
- Configuration-Level Scores: Instead of selecting individual observables (e.g., a single force component), the authors aggregate row-level leverage scores into a configuration-level metric ( $L_R$ ). This score depends solely on the geometry of the weighted design matrix and the regularization prior.
- Sampling Strategies: Two CUR-type subsampling modes are implemented:
  - Energy-CUR: Selects configurations based on leverage scores derived from an energy-only feature matrix.
  - Block-CUR: Selects configurations based on leverage scores derived from the full block of energy and force data, treating each configuration as a correlated block of rows.
- Baselines: The leverage-guided methods are compared against random sampling, energy-based selection (ranking by DFT energy residuals), and force-based selection (ranking by DFT force residuals).
Validation Protocol: An iterative protocol is used where training sets are initialized with 10% of the pool and expanded. Performance is evaluated on held-out test sets and independent DFT validation sets comprising bulk, surface, grain boundary, and defect structures.

Key Contributions

Label-Free Efficiency: The paper demonstrates that leverage scores can be computed before any DFT labeling occurs, enabling a "filter-then-label" workflow that avoids the computational cost of labeling the entire candidate pool.
Transferability Preservation: Unlike some previous data selection studies that focused only on fitting metrics, this work explicitly validates whether reduced training sets maintain accuracy on out-of-distribution properties, specifically defect energetics (vacancies, grain boundaries) and alloy formation energies.
Systematic Comparison: The study provides a controlled comparison of selection strategies across elemental and alloy systems using a fixed ACE architecture, isolating the impact of the selection strategy from hyperparameter tuning.

Results

Data Efficiency: For elemental Al and Cu, leverage-guided subsets (specifically block-CUR) achieve plateau-level energy and force accuracy using only 30–40% of the labeled configurations required by random sampling. This corresponds to an effective 2–3× reduction in DFT labeling costs.
Convergence Behavior: Leverage-guided methods exhibit smoother and more reproducible learning curves with lower run-to-run variability compared to random or force-based selection. Block-CUR consistently yields the lowest force errors, reaching saturation faster than energy-only or heuristic baselines.
Defect and Alloy Performance:
- In Cu, leverage-based selection significantly outperforms baselines in predicting vacancy formation energies and grain boundary energies, reducing errors from ~0.25 eV (baselines) to ~0.04 eV.
- In Al–Cu alloys, while all strategies perform similarly for ordered intermetallics once sufficient chemical diversity is present, leverage-guided selection (specifically CUR at 25% data) achieves superior accuracy in the dilute regime (Cu in Al), reducing vacancy formation errors significantly compared to 50% baseline models.
Independent Validation: Validation against independent DFT test sets (not used in training or selection) confirms that the reduced datasets do not degrade predictive fidelity for bulk, surface, and defect structures.

Significance and Claims
The authors claim that descriptor-space–guided, label-free subsampling is a practical and statistically grounded strategy for reducing the DFT workload in linear ACE potential development. The paper argues that:

Data efficiency and transferability are not mutually exclusive: High-quality, transferable potentials can be trained on substantially smaller subsets if the selection is guided by the geometric structure of the descriptor space rather than scalar residuals.
Geometry drives selection: By prioritizing configurations that span the descriptor space (high leverage), the method effectively captures broken-symmetry and low-coordination environments (critical for defects) that are often sparsely sampled by random or residual-based methods.
Workflow Optimization: The approach enables a shift from "label-then-filter" to "filter-then-label," fundamentally altering the cost structure of MLIP development by eliminating the need to compute DFT labels for the vast majority of generated candidate structures.

The paper concludes that while the method is demonstrated on linear ACE models, the principle of geometry-driven subset selection offers a general pathway for improving data efficiency in MLIPs trained on systematically generated structure pools.

The Great Material Hunt: Why We Need Smarter Maps

The Paper's Big Idea: Finding the "Golden Spots"

Technical Summary: Data-Efficient Training of Linear ACE Potentials through Leverage-Guided Subset Selection

More like this