In materials science, there is a fundamental asymmetry between how easily we can describe a material and how accurately we can predict its behavior. Knowing a substance's chemical formula (its composition) is trivial. However, knowing exactly how its atoms are arranged in three-dimensional space (its crystal structure) is incredibly difficult, often requiring expensive quantum-mechanical simulations or grueling laboratory crystallography.
To accelerate discovery, early materials machine learning (ML) models relied heavily on composition-only inputs. By mapping a simple list of elements and their ratios directly to a target property, these models bypassed the structural bottleneck entirely.
But this shortcut hides a critical mathematical flaw: compositional degeneracy. A single chemical formula does not map to a single material; it maps to an entire landscape of potential structural arrangements, known as polymorphs. Because a material's physical properties are dictated by its spatial structure, composition-only models face an insurmountable physical blind spot.
The One-to-Many Mapping Problem
To understand why composition-only models struggle, consider a classic example: carbon. A model looking only at the element carbon cannot inherently distinguish between soft, lubricating graphite and ultra-hard, transparent diamond.
As noted by Lee et al., deterministic machine learning models typically map a chemical composition to a single static vector. This mathematical formulation assumes a one-to-one relationship between formula and property. In reality, polymorphism introduces severe aleatoric uncertainty. Identical chemical formulas can manifest as entirely different phases with vastly divergent electronic, mechanical, and chemical behaviors.
This structural uncertainty spans the entire spectrum of chemistry. It complicates the characterization of simple organometallic compounds (Bacsa et al.), introduces phase ambiguity in complex layered perovskites (Yang et al.), and alters the solubility and conductivity of molecular crystals (Dzamba et al.). In the lab, controlling which polymorph crystallizes requires precise intervention, such as using polymer microgels (Whaley et al.) or tracking real-time phase evolution via in situ NMR (Harris). If physical systems require this much control to isolate a phase, an ML model operating purely on formulas is fundamentally unequipped to resolve the differences.
The Illusion of Database Sufficiency
If composition-only models are fundamentally blind to structure, why did they initially achieve such high accuracy on benchmark datasets?
The answer lies in a systemic data bias. In a study analyzing structure-property relationships, Aspuru-Guzik et al. introduced an information-theoretic framework to explain this paradox. They revealed that popular materials databases suffer from severe polymorph scarcity. Most entries in these datasets represent only the most stable, lowest-energy "ground state" structure for any given formula.
Because these benchmarks lack structural variety, the inherent degeneracy of chemical compositions is underrepresented. The datasets artificially enforce a one-to-one mapping that does not exist in nature. Consequently, the high performance of composition-only models is an illusion maintained by curated data; when deployed to find novel materials or metastable phases, these models rapidly fail because they cannot generalize to alternative spatial configurations.
Where Composition Fails to Predict Performance
When predicting properties that are highly sensitive to atomic configurations, composition-only models break down completely.
- Superconductivity: The critical temperature ($T_c$) at which a material becomes a superconductor depends entirely on fine-grained electron-phonon interactions. Tran et al. demonstrated that predicting superconductivity from composition alone reduces physical reliability because atomic arrangement dictates the underlying quantum mechanics. To fix this, Sommer et al. developed the 3DSC dataset, explicitly pairing chemical formulas with approximate 3D structures. They proved that injecting even rough structural information drastically improves an ML model's ability to predict $T_c$ accurately.
- Ionic Conductivity: For solid-state batteries, finding materials that allow lithium ions to flow freely is essential. However, Hargreaves et al. showed that composition-only models fail to differentiate between polymorphs of the same lithium electrolyte. Because ion transport pathways are entirely structural, a model blind to the lattice cannot determine if a material will be a superb conductor or a complete insulator.
- Organic Electronics: In devices like OLEDs and organic transistors, macroscopic performance depends on how molecules pack together. As DeLongchamp points out, minor variations in molecular side chains can trigger completely different assembly motifs, making device performance hyper-sensitive to structural shifts that composition metrics fail to capture.
Infusing Structural Awareness into Fast Models
To build truly reliable models, researchers must bridge the gap between composition and structure. One option is Crystal Structure Prediction (CSP), but hunting for the lowest-energy 3D arrangement from scratch is computationally exhausting. Even advanced shortcuts, like using machine learning to swap elements into existing structural templates (Liu et al.) or using recommender systems to screen stable compositions (Seko et al.), ultimately still require expensive downstream structural refinement.
The frontier of the field focuses on a smarter alternative: cross-modal representation learning. The goal is to give fast, composition-based models "structural sight" during training, so they can make accurate predictions at test time without needing explicit 3D coordinates.
[Chemical Formula] ───► ( Composition Encoder ) ──┐
├──► [Contrastive Loss]
[3D Crystal Lattice] ──► ( Structure Encoder ) ───┘
The authors of CLCS (Lee et al.) achieved this by setting up a contrastive learning framework between composition and structure encoders. By forcing the model to align the latent representations of formulas with their corresponding 3D lattices, the composition encoder learns to implicitly capture structural information.
Taking this further, Na developed CroMEL (Cross-modality material embedding loss). CroMEL aligns the embedding spaces of heterogeneous descriptors, allowing a model to leverage massive 3D simulation databases during training. When evaluating a new material where the crystal structure is entirely unknown, the composition model can project the formula into a structurally aware latent space, resolving polymorphic ambiguity.
Other architectures address the problem by changing how the model outputs or receives data. To handle the hidden structural diversity of perovskites, Tuoc et al. abandoned single-value scalar predictions in favor of probabilistic deep learning, mapping formulas to property distributions that naturally reflect polymorphic variance. Alternatively, Lee et al. designed DOSTransformer, which uses learnable structural prompts to feed crystal system context directly into a multi-modal transformer. This allows the model to explicitly separate different phases of the same atomic species when predicting complex electronic properties like the Density of States (DOS).
By moving away from naive composition-to-property mappings and embracing cross-modal structural alignment, modern materials machine learning is overcoming its oldest blind spot—ensuring that models finally understand not just what a material is made of, but how its atoms stand in space.
Key Foundational Work
Spin, charge and lattice correlation in thiazyl radicals and their molecular compounds (2007) — contributed 2 concept elements to this direction.
Abundant Polymorphism in a System with Multiple Hydrogen-Bonding Opportunities: Oxalyl Dihydrazide (2006) — contributed 2 concept elements to this direction.