🔬 materials science

In the age of big data and AI: do we still need chemical bonds?

This paper argues that chemical bonding remains essential in the age of AI, demonstrating that deep learning models naturally recover bonding concepts from their latent spaces and that explicit bonding descriptors significantly improve materials property predictions, thereby reframing chemical bonding as a learnable, compressible representation of electronic structure.

Original authors: Mohammed Al-Fahdi, Aakash Naik, Katharina Ueltzen, Luc Walterbos, Tom Demeyere, Philipp Benner, Janine George

Published 2026-11-03

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Mohammed Al-Fahdi, Aakash Naik, Katharina Ueltzen, Luc Walterbos, Tom Demeyere, Philipp Benner, Janine George

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

Chemistry is the science of how matter holds together and changes. At its heart lies a simple, powerful idea: the chemical bond. For over a century, scientists have used this concept to explain why atoms stick together to form molecules, why some materials are hard while others are soft, and how substances react with one another. In the world of organic chemistry, which deals with carbon-based life and plastics, these bonds are often visualized as clear lines connecting atoms, much like a map of a city. However, in the vast and complex world of inorganic solids—materials like metals, ceramics, and minerals—the picture becomes blurry. Atoms in these materials do not always connect in neat, predictable pairs. Sometimes they share electrons in a collective cloud, or interact in ways that defy simple drawing. Because of this ambiguity, chemists have developed many different ways to describe these connections, none of which is perfectly unique or universally correct.

Today, a new force has entered the laboratory: artificial intelligence. Powerful computer models can now predict the properties of materials with incredible speed and accuracy, often by looking only at the positions of atoms and their types, without ever explicitly asking "what kind of bond is this?" This has led to a provocative question: if machines can learn the rules of materials without understanding bonds, do we still need the concept of the chemical bond at all? Is it an outdated relic, or does it hold a key that the machines have not yet found?

A team of researchers from Germany set out to answer this question by putting the relationship between chemical bonds and artificial intelligence to the test. They did not simply ask whether the machines could predict properties; they asked what the machines were actually learning. They examined two of the most advanced types of artificial intelligence used in materials science, known as foundation models. These models are trained on massive databases of crystal structures and their energies, learning to predict how atoms interact. The researchers wanted to see if, hidden inside the complex mathematics of these models, there was a representation of chemical bonding that matched the descriptions chemists have used for decades.

The team first looked at a model called Allegro. This model breaks down the total energy of a material into contributions from pairs of atoms, effectively creating a "bond energy" for every connection it sees. The researchers compared these machine-generated bond energies with traditional, quantum-mechanical measures of bond strength and bond order, which are calculated using complex physics equations. They found a striking connection. The machine's internal bond energies correlated strongly with the traditional measures of how strong a bond is and how many electrons are shared. In fact, the machine had learned to recognize these chemical concepts without ever being told what they were. It had discovered them on its own, simply by trying to predict the energy of the material.

They repeated this investigation with a second model called MACE. This model looks at the local environment around each atom rather than just pairs. The researchers checked if the model's internal understanding of an atom's surroundings matched the chemical concept of electric charge. Again, the machine had learned the pattern. The internal numbers the model used to describe an atom's environment aligned closely with the electric charge of that atom, a fundamental property that dictates how it interacts with its neighbors. The machine had not been programmed with the rules of charge or bonding; it had inferred them from the data.

However, the researchers did not stop at showing that machines learn these concepts. They asked the reverse question: if we give the machines the explicit language of chemical bonds, does it help them become better? To test this, they built a new version of a standard machine-learning model and fed it explicit chemical data. They provided the model with numbers representing bond strength and bond order, derived from quantum calculations, alongside the usual information about atom positions. They then asked the model to predict difficult physical properties, such as how stiff a material is (its bulk modulus) or how resistant it is to being scratched (its shear modulus).

The results were clear. The model that received the explicit chemical bond information made significantly better predictions. For properties like stiffness and resistance to shearing, the error in the predictions dropped by fifteen percent compared to models that only saw the positions of atoms. This proved that while machines can learn to mimic the effects of bonding, giving them the actual language of chemistry provides a shortcut. It allows them to understand the material more deeply and make more accurate guesses, especially for properties that depend heavily on how electrons are shared between atoms.

The study suggests that the relationship between chemistry and artificial intelligence is not a competition where one replaces the other. Instead, they are partners. The machines are capable of discovering the hidden patterns of bonding on their own, but they do not always express these patterns in a way that humans can easily read or use. Conversely, the traditional concepts of chemical bonds, while sometimes ambiguous, provide a compact and meaningful way to describe the complex dance of electrons. When these human-friendly concepts are fed into the machines, they act as a guide, helping the models focus on the most important physical details.

The researchers conclude that the age of big data and artificial intelligence does not render the chemical bond obsolete. Instead, it offers a new way to evaluate and refine these concepts. If a machine can learn the rules of bonding from data, and if those rules help the machine predict the behavior of new materials, then the concept of the bond is validated as a useful and powerful tool. The goal is no longer to find the single, perfect definition of a bond, but to find the description that best helps us understand and design the materials of the future. In this new era, chemical bonding is not just a static rulebook; it is a learnable language that connects the abstract world of quantum mechanics to the tangible reality of the materials we build and use.

Technical Summary: In the Age of Big Data and AI: Do We Still Need Chemical Bonds?

Problem Statement
The rapid advancement of deep learning (DL) in materials science, particularly through graph neural networks (GNNs) and machine-learned interatomic potentials (MLIPs), has raised a fundamental question: are explicit chemical bonding concepts still necessary? Modern models achieve remarkable predictive accuracy for energies, forces, and material properties using only compositional and structural inputs (atomic identities and interatomic distances), often without explicit bonding descriptors. This success challenges the utility of traditional bonding heuristics, which are inherently ambiguous (lacking a unique quantum-mechanical definition) and computationally expensive to derive. The paper addresses whether AI renders chemical bonding obsolete or if bonding concepts remain essential for interpretability, data efficiency, and physical insight.

Methodology
The authors employ a multi-faceted approach combining data-driven analysis, latent space probing, and architectural modification:

Latent Space Analysis of Foundation Models: The study investigates whether pretrained foundation MLIPs implicitly learn bonding information.
- Allegro: The authors extract edge-wise energies ( $E_{ij}$ ) and 1,024-dimensional edge embeddings ( $x_{ij}$ ) from a pretrained Allegro model. They correlate these with quantum-chemical descriptors: Integrated Crystal Orbital Hamilton Populations (ICOHP), Integrated Crystal Orbital Bond Indices (ICOBI), and bond distances.
- MACE: Similarly, node embeddings from a pretrained MACE model are analyzed against atomic charges (Mulliken, Löwdin) and site-wise ICOHP/ICOBI statistics.
- Linear Probing: To quantify the information content within these latent spaces, the authors train linear regression models (probing) to predict bonding descriptors (ICOHP, ICOBI, charges) directly from the MLIP latent representations without further fine-tuning of the base model.
Architectural Integration (ESF-MEGNet): To test if explicit bonding information improves performance, the authors developed ESF-MEGNet (Electronic-Structure-Featurized MEGNet). This architecture integrates quantum-chemical descriptors (Löwdin charges as node features; ICOHP and ICOBI as edge features) into a standard MEGNet graph neural network.
- Control Experiments: Performance is compared against a baseline MEGNet and two control variants: ESF-MEGNet-zero (bonding features set to zero) and ESF-MEGNet-rand (bonding features replaced by random values). This isolates the contribution of the bonding information from architectural changes.
Data Sources: The analysis utilizes a database of ~13,800 crystalline materials from the Materials Project, featuring automated LOBSTER calculations for bonding descriptors. Specific case studies include ThCr2Si2-type compounds (e.g., LaRu2P2) to illustrate algorithmic ambiguities in bond detection.

Key Results

Implicit Learning of Bonding: Both Allegro and MACE latent spaces encode chemically meaningful bonding information despite never being explicitly trained on bonding descriptors.
- Allegro: Edge energies ( $E_{ij}$ ) show strong Pearson correlations with ICOHP ( $r=0.82$ ) and ICOBI ( $r=-0.82$ ). Linear probing of edge embeddings ( $x_{ij}$ ) yields $R^2 > 0.95$ for predicting ICOHP and ICOBI, indicating that bond energy and bond order are linearly accessible from the latent space.
- MACE: Node embeddings correlate strongly with atomic charges ( $dCor \approx 0.83-0.85$ ) and summed ICOHP ( $r=0.72$ ). Linear probing achieves $R^2$ values of 0.97 (charges), 0.95 (ICOHP), and 0.95 (ICOBI).
Performance Gains with Explicit Features: Incorporating explicit bonding descriptors into the ESF-MEGNet architecture significantly improves property prediction compared to the baseline MEGNet and control models.
- Bulk and Shear Moduli: ESF-MEGNet reduced prediction error (MAE) by 15% for both bulk and shear moduli.
- Other Properties: Improvements were also observed for Vickers hardness (12% reduction in MAE), average mean-squared displacement (4% reduction), and phonon DOS peak frequency (9% reduction).
- Interpretability: The performance gains were attributed to the added bonding information, as the control models (zero/random features) performed similarly to the baseline, confirming that the improvement stems from the chemical content, not just increased model complexity.
Ambiguity in Bonding Detection: The paper demonstrates that no single bonding algorithm is universally correct. In ThCr2Si2 structures, geometric algorithms failed to detect metal-metal bonds, while quantum-chemical (ICOHP-based) algorithms overestimated interlayer P-P bonding in specific pressure regimes. This highlights that different descriptors capture complementary aspects of electronic structure.

Significance and Claims
The paper argues that the age of AI does not render chemical bonding obsolete; rather, it shifts the perspective from seeking a "true" definition of a bond to evaluating the utility of bonding descriptors.

Mutual Benefit: The relationship between chemical bonding and machine learning is one of mutual benefit. While foundation models implicitly learn bonding representations, explicit bonding descriptors provide a low-dimensional, interpretable compression of electronic structure that improves predictive accuracy and model trustworthiness.
Bonding as a Learnable Representation: The authors propose that chemical bonding should be understood as a "learnable representation of electronic structure." Different bonding models (geometric, orbital-based, density-based) represent alternative compressions of the same underlying information. Their value is determined by their predictive power, dimensionality, and complexity.
Future Directions: The work suggests that future foundation models could be trained to simultaneously predict energies, forces, and bonding descriptors (e.g., ICOHP), bridging the gap between high-accuracy prediction and chemical interpretability. It also posits that joint-embedding predictive architectures (JEPAs) could be used to learn robust representations by combining multiple, complementary bonding analyses.
The "Materials Genome": The authors conclude that if a "materials genome" exists, it must be encodable through a representation of chemical bonding. If no bonding representation can capture the necessary information to explain and predict material behavior, the current understanding of the chemical bond may require revision.

In summary, the paper demonstrates that while deep learning can implicitly recover bonding concepts, explicitly incorporating these concepts into machine learning architectures enhances both the performance and interpretability of materials models, validating the continued necessity of chemical bonding in the era of big data.

Technical Summary: In the Age of Big Data and AI: Do We Still Need Chemical Bonds?

More like this