🔬 materials science

Generalizing Abell-Tersoff bond-order potential with explicit high-order many-body correlations for robust extrapolation of potential energy surfaces

This paper introduces a semiparametric interatomic potential that generalizes the Abell-Tersoff bond-order model with explicit high-order many-body correlations and chemically informed constraints, achieving interpolation accuracy comparable to machine-learning potentials while significantly improving robustness in extrapolating to unseen high-pressure and high-temperature configurations.

Original authors: Ikuma Kohata

Published 2026-08-25

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Ikuma Kohata

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

In the world of materials science, understanding how atoms stick together is the key to predicting how matter behaves. At the heart of this understanding lies a concept called the potential energy surface, which acts like a topographic map for atoms. Just as a hiker uses a map to see where the ground is high or low, scientists use this map to see where atoms are most stable and where they might move. For decades, researchers have relied on two main ways to draw these maps. The first method is like using a simple, hand-drawn sketch based on general rules of chemistry; it is fast and reliable but often too rough to capture the fine details of complex materials. The second method uses powerful computers and massive amounts of data to create a highly detailed, flexible map. While this second approach is incredibly accurate within the range of data it has seen, it often fails spectacularly when asked to predict what happens in extreme situations it has never encountered, such as crushing pressures or scorching temperatures. This limitation has been a major hurdle for simulating the most violent and exotic environments in the universe.

A researcher at the University of Tokyo has introduced a new approach that bridges this gap, offering a way to draw these atomic maps that is both detailed and robust. They developed a model called BOBOP, which stands for a generalized version of a classic bond-order potential. Instead of relying solely on vast datasets to learn how atoms interact, this model is built on a foundation of chemical intuition, specifically how the strength of a bond between two atoms changes depending on how many other neighbors are crowding around them. The researcher found that by explicitly teaching the model to recognize complex, multi-atom relationships and by hard-coding physical limits into its structure, the model could predict energy landscapes with high accuracy. More importantly, when tested on conditions far outside its training data, such as the extreme pressures found deep inside planets or the chaotic environment of liquid carbon, the model remained stable and physically realistic, whereas other modern, data-heavy models broke down and produced impossible results.

The core challenge the researcher addressed is the trade-off between flexibility and reliability. Modern machine-learning models are like brilliant students who can memorize a textbook perfectly but often fail a test if the questions are phrased slightly differently. In the context of atoms, this means that if a simulation encounters a configuration of atoms that was not in the training data, the model might invent a fake energy valley or a false barrier, causing the simulation to crash or produce nonsense. The new model avoids this by incorporating a specific mathematical structure that mimics the physical reality of bond saturation. In simple terms, as an atom gains more neighbors, the bonds it shares with each neighbor must weaken because the available electrons are spread thinner. The researcher ensured their model respects this rule by design, rather than hoping the data would teach it. They also expanded the model's ability to see beyond simple pairs of atoms, allowing it to account for how three, four, or even five atoms influence one another simultaneously. This explicit inclusion of higher-order correlations prevents the model from confusing two different atomic arrangements that might look similar to a simpler model but are physically distinct.

To test their creation, the researcher trained the model on a variety of datasets, including silicon, carbon, water, and small molecules like ethanol. They then put the model through a series of rigorous stress tests. In one experiment involving carbon, they simulated the material under pressures up to 1,000 gigapascals and temperatures near 9,000 Kelvin. Under these extreme conditions, other leading models became unstable, causing the simulated atoms to collapse into unphysical volumes or oscillate wildly. The new model, however, maintained a smooth and stable energy landscape, accurately predicting the behavior of liquid carbon where others failed. Similarly, when simulating water, the model correctly reproduced the diffusion rates and structural properties of liquid water at various temperatures, matching experimental data closely. In tests involving the shock waves that travel through diamond under high pressure, the model's predictions aligned well with the most accurate theoretical calculations, while other models diverged significantly.

The success of this work suggests that the path forward for atomic simulations does not lie solely in collecting more data or building larger neural networks. Instead, embedding known physical laws directly into the structure of the model provides a safety net that allows for reliable predictions in the unknown. By ensuring the model behaves reasonably when pushed to its limits, the researcher has created a tool that can be trusted to explore the most extreme corners of the material world. This approach offers a promising direction for designing new materials and understanding planetary interiors, where the conditions are too harsh for direct observation and too complex for traditional approximations. The study demonstrates that when machine learning is guided by the fundamental principles of chemistry, it can achieve a level of robustness that pure data-driven methods struggle to reach, opening the door to more reliable simulations of the universe's most challenging environments.

Technical Summary: Generalizing Abell–Tersoff Bond-Order Potential with Explicit High-Order Many-Body Correlations

Problem Statement
Machine-learning interatomic potentials (MLIPs) offer high accuracy and efficiency for atomistic simulations but suffer from poor robustness when extrapolating to configurations outside their training domain (out-of-distribution). While empirical potentials (e.g., Tersoff, Stillinger–Weber) exhibit robust extrapolation due to physically motivated functional forms, they lack the flexibility to accurately describe complex potential energy surfaces (PES). Conversely, flexible MLIPs often produce unphysical artifacts—such as spurious local minima, false energy barriers, and oscillatory behavior—under extreme conditions like high pressure or temperature. Existing attempts to combine empirical forms with ML (e.g., bond-order-based MLIPs) often compromise physical interpretability or fail to resolve specific degeneracies in bonding environments, limiting their reliability for multi-element systems and extrapolation tasks.

Methodology
The author introduces BOBOP (Body-Ordered Bond-Order Potential), a semiparametric interatomic potential that generalizes the Abell–Tersoff bond-order potential. The model retains an analytical functional form rooted in chemical intuition while explicitly incorporating high-order many-body correlations.

Functional Formulation: The total energy is defined as a sum of site energies, comprising repulsive ( $V_R$ ) and attractive ( $V_A$ ) Morse-type terms, and a bond-order term ( $b_{ij}$ ) that modulates the attraction based on the local environment.
Body-Ordered Extension: Unlike the original Tersoff potential, which relies on a single scalar bond-order descriptor derived from three-body angular histograms, BOBOP constructs a vector of bonding-environment descriptors ( $\zeta_{ij}$ $ζ_{ij}$ ). This vector is built from:
- Atom Basis: Products of cutoff functions, radial basis functions (exponential decay), and angular basis functions (polynomials of direction cosines).
- Explicit High-Order Features: The model explicitly constructs three-body, four-body, and five-body features. Crucially, it employs a self-interaction correction to remove contributions where neighbor indices coincide with the bond indices, ensuring features vanish when no neighbors exist.
- Nonlinear Mapping: The bond order $b_{ij}$ is obtained via a nonlinear function $\Phi(\zeta_{ij})$ designed to satisfy physical asymptotic constraints: $b_{ij} \to 1$ for isolated dimers and $b_{ij} \to 0$ for infinitely coordinated systems (bond saturation).
Training: The model is trained using a loss function minimizing errors in both potential energy and forces relative to Density Functional Theory (DFT) references. All parameters are optimized simultaneously from random initialization, without pretraining the analytical components.

Key Contributions

Resolution of Bond-Energy Degeneracy: The paper demonstrates that standard three-body descriptors (and the original Tersoff potential) suffer from "bond-energy degeneracy," where distinct geometric configurations (e.g., specific planar arrangements in methane) yield identical bond orders and energies. BOBOP resolves this by explicitly including four-body (and higher) correlations, which capture the relative orientations of multiple neighboring bonds that three-body descriptors miss.
Robust Extrapolation via Physical Constraints: By embedding physically motivated constraints (bond saturation, exponential radial decay) directly into the functional form, BOBOP avoids the unphysical oscillations and spurious minima often seen in purely data-driven MLIPs when extrapolating to unseen configurations.
Semiparametric Framework: The work establishes a framework that bridges the gap between empirical potentials and nonparametric MLIPs, offering a model that is both interpretable and capable of high-fidelity representation of complex PES.

Results
BOBOP was trained and evaluated on diverse datasets including Silicon, Carbon, Water, Ethanol, and 3BPA, and compared against leading MLIPs (GAP, ACE, DeePMD, etc.).

Carbon (Limited Diversity): Trained only on graphene and diamond, BOBOP accurately predicted cohesive energy curves for unseen structures (bc8, fcc, simple cubic) and maintained stable gradients under extreme compression and expansion. In contrast, ACE and DeePMD exhibited unphysical oscillations or instability.
Silicon: On the GAP silicon dataset, BOBOP achieved energy and force RMSEs comparable to GAP and ACE. Crucially, in Random Structure Searching (RSS) and adatom decohesion tests, BOBOP correctly identified the diamond ground state and produced smooth energy profiles, whereas GAP and ACE predicted unphysical low-energy states or exhibited spurious barriers.
High-Pressure Carbon: In NPT molecular dynamics simulations of liquid carbon at pressures up to 1000 GPa, ACE simulations became unstable and collapsed. BOBOP remained stable, reproducing radial and angular distribution functions consistent with DFT. BOBOP also accurately predicted the diamond shock Hugoniot curve at high pressures where ACE failed.
Water: BOBOP achieved interpolation accuracy comparable to local MLIPs. MD simulations of liquid water showed physically reasonable RDFs and diffusion coefficients at 300 K that matched experimental values better than ACE.
Small Molecules (Ethanol, 3BPA): BOBOP demonstrated accuracy comparable to other local models for interpolation. For out-of-distribution configurations (e.g., 3BPA at 600 K and 1200 K, or H-atom removal in ethanol), BOBOP maintained smooth, qualitatively correct energy profiles, outperforming models that exhibited large deviations or unphysical behavior.

Significance and Claims
The paper claims that BOBOP demonstrates that incorporating physically motivated constraints and explicit high-order many-body correlations into the functional form of an interatomic potential significantly improves extrapolation behavior. The study suggests that:

Physically motivated inductive biases (such as bond saturation limits and specific high-order correlation terms) are essential for controlling the extrapolation of MLIPs.
Explicit high-order correlations are necessary to resolve degeneracies in bonding environments that limit the accuracy of lower-order descriptors.
A semiparametric approach can achieve interpolation accuracy comparable to state-of-the-art MLIPs while offering superior robustness in extreme conditions (high pressure/temperature) and unseen configurations.

The author notes limitations, including the computational cost of symmetrizing high-order edge features and the current lack of explicit long-range interaction and dihedral-angle dependence, suggesting future work may involve recursive evaluation (message-passing) to address these while preserving physical interpretability.

Technical Summary: Generalizing Abell–Tersoff Bond-Order Potential with Explicit High-Order Many-Body Correlations

More like this