📄 chemistry

The Illusion of Chemical Continuity Limits Machine Learning Generalization

This paper challenges the "Scaling Hypothesis" in materials informatics by demonstrating that machine learning models fail to generalize for bandgap predictions because the electronic bandgap manifold is topologically disjoint due to discrete orbital quantization, necessitating a shift from simple coordinate-based regression to methods that directly integrate quantum mechanical Hamiltonians.

Original authors: Oleksandr Voznyy, Zhibo Wang, Alexander Davis, Rajarshi Dutta, Alex-Cristian Tomut, Kanishk Yadav, Ihor Neporozhnii, Sjoerd Hoogland

Published 2026-08-12

📖 9 min read🧠 Deep dive

CC BY 4.0

Original authors: Oleksandr Voznyy, Zhibo Wang, Alexander Davis, Rajarshi Dutta, Alex-Cristian Tomut, Kanishk Yadav, Ihor Neporozhnii, Sjoerd Hoogland

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

Imagine you are trying to teach a robot to predict the weather. You might think that if you give the robot a million years of weather data, it will eventually learn the perfect formula to tell you exactly what the sky will do tomorrow. This is the dream of "materials informatics," a field where scientists use massive amounts of data and powerful computer brains (machine learning) to predict how new materials will behave. The big idea, or "Scaling Hypothesis," is that the universe of chemicals is like a smooth, continuous hill. If you know the shape of the hill in one spot, you should be able to guess the shape everywhere else just by looking at enough data points. This works beautifully for some things, like predicting how much energy it takes to build a molecule. But for a specific property called the "bandgap"—which is essentially the magic switch that decides if a material is a conductor, an insulator, or a semiconductor (like the silicon in your phone)—the robots keep hitting a wall. No matter how much data they are fed, their predictions are often off by a wide margin, sometimes missing the mark by as much as 0.4 to 0.5 electron volts (eV). Since visible light spans a range of about 1 eV, this error is huge; it's like trying to tune a radio and landing on static instead of the song you want.

This paper investigates why the robots are failing at this specific task. The authors, a team of researchers from the University of Toronto and other institutions, argue that the problem isn't that the robots aren't smart enough or that they haven't seen enough data. Instead, they suggest that the "smooth hill" the robots are trying to climb doesn't actually exist for bandgaps. They found that the chemical space for these properties is actually more like a shattered mosaic of disjoint islands. When the researchers tried to train their models on different types of crystal structures (like rocksalt vs. zincblende), adding more data didn't help; it actually made the predictions worse. They discovered that the fundamental nature of electrons and atoms creates sharp, sudden jumps in properties that cannot be smoothed out by simple math. The paper suggests that to fix this, we can't just throw more data at the problem; we need to teach the robots the actual quantum physics rules (the Hamiltonian) that govern these jumps, rather than just asking them to guess based on patterns.

The Illusion of a Smooth Hill

For a long time, scientists believed that the world of chemistry was a giant, continuous landscape. Imagine a smooth, rolling hill where if you know the height at one point, you can easily guess the height a few steps away. This idea, called the "Scaling Hypothesis," is the foundation of modern materials discovery. The belief was that if we just collected enough data points—millions of crystal structures and their properties—our artificial intelligence (AI) models could learn to walk this hill perfectly. They could predict how a new, never-before-seen material would act just by looking at its neighbors.

This strategy has been a huge success for some properties. For example, predicting "formation energy" (how much energy it takes to build a material from scratch) is like walking on a smooth, well-paved path. The AI models can do this with near-perfect accuracy, matching the results of the most complex physics simulations. But when it comes to predicting the "bandgap," the AI hits a brick wall. The bandgap is a critical property that determines how a material interacts with light and electricity. It's the difference between a material that blocks electricity (an insulator), one that conducts it freely (a metal), and one that can be switched on and off (a semiconductor).

Despite a decade of trying and feeding the models massive datasets, the error in predicting bandgaps stays stubbornly high, hovering around 0.4 to 0.5 eV. To put that in perspective, the entire range of colors we can see with our eyes is only about 1 eV wide. An error of 0.5 eV means the AI might predict a material is transparent when it's actually opaque, or that it glows red when it should be blue. The question the paper asks is simple: Why does the AI fail so badly at this one specific task when it's so good at everything else?

The Shattered Mosaic

The authors of this paper decided to stop assuming the hill was smooth and start looking at the ground with a microscope. They used a comprehensive dataset of binary semiconductors (materials made of two elements) and state-of-the-art AI models to map out the landscape. What they found was shocking: the landscape for bandgaps isn't a smooth hill at all. It's a collection of disjoint islands.

Imagine you are trying to teach a robot to recognize the difference between a "rock" and a "feather." If you show it a million rocks and a million feathers, it learns the pattern. But what if you suddenly show it a "rock-feather" hybrid? If the robot has only ever seen pure rocks and pure feathers, it might get confused. The authors found that different crystal structures (the way atoms are arranged) are like these different islands. They found that models trained on one type of structure, like "zincblende," failed miserably when asked to predict properties for another type, like "rocksalt," even if the chemical ingredients were the same.

In fact, adding more data from different structures didn't help; it made things worse. This is called "negative learning." It's like trying to teach a student to solve math problems by mixing up algebra, geometry, and calculus all in one lesson without explaining the rules. The student gets confused and starts getting everything wrong. The paper shows that when you mix data from different crystal structures, the AI tries to average them out, creating a prediction that is wrong for everyone. The "latent space" (the internal map the AI builds in its brain) shows that these structures occupy completely separate, non-overlapping regions. There is no smooth path connecting them.

The Quantum Cliff

Why is the landscape so broken? The paper traces the problem back to the fundamental nature of electrons. In the quantum world, electrons live in specific, discrete "orbitals" (like steps on a ladder). You can't stand halfway between two rungs. When you change the ingredients of a material slightly, the energy levels of these electrons don't slide smoothly; they jump.

The authors ran a simulation where they slowly changed one element into another (a process called "alchemical interpolation"). They watched what happened to the total energy of the material versus the bandgap. The total energy behaved like a smooth slide, changing gently as the atoms shifted. But the bandgap? It behaved like a cliff. As soon as the electron count changed just a tiny bit, the bandgap would suddenly collapse or jump to a completely different value. This happens because the electrons have to rearrange themselves into new, distinct energy levels, and this rearrangement creates a sharp "cusp" or break in the data.

Because of these sharp jumps, the periodic table (the chart of all elements) isn't a smooth gradient for bandgaps. It's a collection of isolated domains. The paper shows that even within a single group of elements, like the halides or chalcogenides, the trends are messy and unpredictable. For example, while the bandgap for sodium (Na) compounds might decrease smoothly as you change the partner element, magnesium (Mg) compounds might do something completely different, breaking the pattern entirely. The AI, which relies on finding smooth patterns to make guesses, has nothing to grab onto.

The Illusion of Continuity

The core argument of the paper is that the "Big Data" approach has hit a limit. The prevailing belief was that if we just gathered enough data, the AI would eventually figure out the physics. But the authors argue that you can't solve a problem of quantum discontinuity with a problem of statistical continuity. The AI is trying to draw a smooth line through a set of points that are actually on different, disconnected islands. No matter how many points you add, if the islands don't touch, the line will always be wrong.

The paper explicitly rules out the idea that the problem is just a lack of data or a lack of model complexity. They tested this by using the most advanced models available (like coGN and CGCNN) and by training them on massive datasets. The results were the same: the error floor remained. They also ruled out the idea that the problem was just the complexity of the crystal structures, because even when they simplified the task to look only at the "gamma point" (a specific spot in the energy map), the errors persisted.

The authors suggest that the solution isn't to collect more data, but to change the question. Instead of asking the AI to predict the bandgap directly (which is a jagged, discontinuous property), we should ask it to predict the underlying physics that creates the bandgap. They propose that we need to teach the AI to learn the "Hamiltonian" (the mathematical description of the quantum system) directly. If the AI can learn the smooth, continuous rules of how electrons interact, it can then calculate the jagged bandgap from those rules. It's the difference between trying to memorize the weather forecast for every day of the year versus learning the laws of thermodynamics that actually cause the weather.

Conclusion: A New Path Forward

This paper serves as a reality check for the field of materials science. It suggests that the dream of a universal, data-driven model that can predict any property of any material is flawed when it comes to electronic properties like bandgaps. The "illusion of chemical continuity" is just that—an illusion. The quantum world is discrete, and our current methods of treating it as a smooth, continuous landscape are fundamentally mismatched.

The authors don't claim to have solved the problem yet. Instead, they have identified the boundary condition: simple regression on atomic coordinates cannot resolve the intrinsic discontinuities of eigenvalue phenomena. They suggest that the future of the field lies in integrating quantum mechanics more deeply into machine learning, perhaps by having the AI predict the continuous Hamiltonian matrix elements instead of the final property. Until we do that, the AI will continue to struggle with the sharp, jagged edges of the quantum world, no matter how much data we throw at it. The paper ends with a clear message: we need to stop trying to smooth out the cliffs and start learning how to climb them.

Technical Summary: The Illusion of Chemical Continuity Limits Machine Learning Generalization

Problem Statement
The field of materials informatics operates under the "Scaling Hypothesis," which assumes that chemical space is a continuous manifold. This premise suggests that with sufficient data, deep learning models can approximate any structure-property relationship, effectively learning the universal functional mapping from crystal structure to physical properties. While this approach has achieved near-Density Functional Theory (DFT) accuracy for thermodynamic properties like formation energy, it fails to generalize for electronic bandgaps. Despite a decade of architectural innovation and exponential data growth, the Root Mean Squared Error (RMSE) for bandgap prediction remains stubbornly high (approximately 0.4–0.5 eV), with outliers reaching 7 eV. This error magnitude is critical, as it exceeds the energy range of visible light (~1 eV), rendering predictions inadequate for practical applications. The prevailing view attributes this stagnation to data scarcity or insufficient model complexity; however, this paper challenges that assumption, positing that the fundamental limitation lies in the topological nature of the property itself rather than the data or the model.

Methodology
To isolate the root cause of the error floor, the authors conducted a systematic investigation using a comprehensive dataset of binary semiconductors and state-of-the-art Graph Neural Networks (GNNs), specifically Connectivity Optimized Nested Line Graph Networks for Crystal Structures (coGN) and the Crystal Graph Convolutional Neural Network (CGCNN).

Dataset Curation and Standardization: To decouple structural variation from intrinsic property trends, the study utilized full combinatorial sets of binary materials. To eliminate complexities arising from indirect bandgaps and k-space anisotropy, the electronic gap was standardized to the gamma-point gap ( $\Gamma$ -gap). The dataset included materials from the Materials Project, supplemented by DFT-calculated structures (using VASP with PBE functionals) to fill gaps in zincblende binary compounds, particularly for alkali and alkaline earth elements.
Topological Analysis: The authors analyzed the "roughness" of the learned manifold by comparing target values against the average of their nearest neighbors in the latent space. They employed Uniform Manifold Approximation and Projection (UMAP) and Principal Component Analysis (PCA) to visualize the latent space embeddings of different crystal structures (rocksalt, zincblende, CsCl).
Transferability Testing: Models were trained on various subsets (e.g., gamma gaps from the Materials Project, zincblende-only datasets, and combined datasets) to assess whether increasing structural diversity improved generalization. Specific "hard" cases, such as MgO and NaF in the zincblende structure, were used as fixed leave-one-out test sets to evaluate extrapolation capabilities.
Alchemical Interpolation: To probe the continuity of the property space between discrete atoms, the authors performed ab initio simulations of "alchemical" transmutation (e.g., varying electron count from ZnSe to GaSe under a rigid-band approximation). This compared the smoothness of total energy evolution against the evolution of the electronic bandgap.
Statistical Correlation: The study employed Pearson and Spearman correlation coefficients, linear regression residuals, and Kolmogorov-Smirnov tests to evaluate trends within and between crystal families and oxidation states.

Key Results

Topological Disjointness: Unlike formation energies, which exhibit a smooth, continuous manifold where nearest neighbors correlate strongly with target values, bandgaps display a topologically disjoint landscape. In the latent space, different crystal structures (rocksalt, zincblende, CsCl) occupy distinct, non-overlapping regions with no continuous path connecting them.
Negative Learning: Increasing the dataset size by adding structurally diverse data (e.g., adding rocksalt data to a zincblende model) does not improve accuracy. Instead, it results in "negative learning," where the model's prediction accuracy for unseen chemistries degrades. The model is forced to interpolate across a void, averaging incompatible structure-property relationships and leading to regression-to-the-mean behavior.
Lack of Intra-Group Trends: Even within a single crystal structure class (e.g., zincblende), no consistent monotonic trends exist across oxidation states or element rows. Statistical analysis revealed weak or non-existent correlations between residuals of different groups, and property distributions were found to be statistically disjoint.
Discrete Quantum Nature: The alchemical interpolation simulations demonstrated that while total energy follows a smooth, differentiable curve, the electronic bandgap exhibits sharp, non-differentiable discontinuities. As fractional electrons are added, the Fermi level jumps sharply past the conduction band minimum due to the quantized nature of atomic orbitals and sudden changes in HOMO identity.
Architecture Independence: The failure to generalize persists across different model architectures, including advanced GNNs like coGN. This confirms that the limitation is not a lack of model complexity but an intrinsic property of the data topology.

Key Contributions

Identification of the "Illusion of Continuity": The paper establishes that the assumption of a continuous chemical manifold is invalid for electronic eigenvalue phenomena like bandgaps. The periodic table, when treated as a continuous vector space for these properties, is actually a collection of discrete, disjoint manifolds.
Demonstration of Negative Transfer: The study provides empirical evidence that aggregating distinct crystal structures into massive heterogeneous datasets can actively harm model performance for bandgap prediction, contradicting the "Big Data" paradigm in this specific domain.
Root Cause Isolation: By systematically eliminating confounding factors (structural variation, indirect gaps, model architecture), the authors isolate the discrete quantization of atomic orbitals and resulting band crossings as the fundamental barrier to ML generalization.
Methodological Shift Proposal: The authors argue that the current approach of direct property regression on atomic coordinates is insufficient. They propose a shift toward learning the underlying quantum mathematics, specifically suggesting the direct entrainment of Hamiltonian matrix elements or electron density as descriptors, which evolve more smoothly than the resulting bandgap.

Significance and Claims
The paper claims to establish a boundary condition for the "Big Data" paradigm in materials informatics. It asserts that simple regression on atomic coordinates cannot resolve the intrinsic discontinuities of eigenvalue phenomena. The significance of this work lies in its challenge to the prevailing dogma that more data and complex models will eventually solve the bandgap prediction problem. Instead, it posits that the periodic table representation is too coarse-grained to capture the abrupt transitions of quantum states.

The authors conclude that until ML models integrate explicit quantum mechanical principles—moving from property interpolation to Hamiltonian learning—bandgap predictions will remain constrained by the inherent topological fragmentation of chemical space. The work suggests that the current error floor is not a technical hurdle to be overcome by scaling, but a fundamental limitation arising from the non-differentiable nature of the target property relative to the input descriptors.