Structure-Agnostic Material Property Prediction

38 concepts27 papers2 proposals

The Structural Bottleneck in Materials Discovery

Traditional materials informatics operates under a structural paradigm: to predict how a material behaves, you must first know exactly where its atoms sit in three-dimensional space. Properties like thermal conductivity, mechanical strength, or electronic bandgaps are deeply tied to crystal lattices and bonding networks.

However, acquiring these 3D coordinates creates a severe computational and experimental bottleneck. For undiscovered or hypothetical compounds, finding the stable 3D geometry requires expensive quantum mechanical simulations like Density Functional Theory (DFT), as noted by Wei-Chih Chen et al. in Machine learning and evolutionary prediction of superhard B-C-N compounds. For real-world synthesized materials, such as heavily disordered alloys or doped semiconductors, the exact atomic arrangements are often messy, dynamic, or entirely uncharacterized, as discussed by Gyoung S. Na et al. in A public database of thermoelectric materials and system-identified material representation for data-driven discovery.

To bypass this bottleneck, a powerful alternative direction has emerged: structure-agnostic property prediction. The core intuition is that a material's chemical formula alone contains a massive amount of implicit structural and physical information. Elements carry intrinsic periodic traits—such as electronegativity, ionic radii, and valence electron configurations—that fundamentally dictate and constrain how they can organize in nature. Rather than explicitly computing the final 3D shape, models can learn to map the chemical "recipe" directly to macroscopic properties.

Learning Latent Chemistry

How do we represent elements without manually engineering physical descriptors? One approach is to let the data reveal the underlying relationships. In Matrix- and tensor-based recommender systems for the discovery of currently unknown inorganic compounds, Atsuto Seko et al. demonstrated that collaborative filtering and Tucker decomposition—the same mathematics powering commercial recommender systems—can identify stable new materials. By treating chemical elements like users and items, the model learns latent elemental properties based solely on their co-occurrence in known crystal databases, completely free of explicit physical descriptors.

Stoichiometric Feature Aggregation

To turn a full chemical formula into a single numerical vector for machine learning, researchers use stoichiometric embeddings, a concept reviewed by James Damewood et al. in Representations of Materials for Machine Learning. The mechanism operates in three distinct steps:

  1. Elemental Vectorization: Each element in a compound is mapped to a dense vector. This can be based on periodic trends, such as the structure-free "Mendeleev encodings" utilized by Zixin Zhuang et al. in Classification of battery compounds using structure-free Mendeleev encodings. Alternatively, vectors can encode statistical geometric preferences; Andrij Vasylenko et al. developed Local Environment-induced Atomic Features (LEAFs) in Digital features of chemical elements extracted from local geometries in crystal structures to capture how elements statistically prefer to arrange themselves in space.
  2. Stoichiometric Weighting: The vector for each constituent element is multiplied by its fractional concentration or stoichiometric ratio within the formula.
  3. Global Pooling: The weighted vectors are aggregated (typically via summation) to generate a single global descriptor for the material. As shown by Andrij Vasylenko et al., this simple aggregation bridges the gap between composition and crystal structure, enabling highly accurate structure-agnostic property prediction.

Mapping the Compositional Landscape

Can a model operating purely on elemental fractions capture highly non-linear physical phenomena? Remarkably, the thermodynamic and mechanical landscapes of materials are often smooth and highly continuous when projected onto compositional coordinates.

For example, in A seven-facet polyhedron captures the composition-only formation-energy landscape of inorganic solids, Justin Tahmassebpur et al. revealed that the bulk formation-energy landscape of inorganic solids can be accurately approximated by a simple composition-only "max-affine" model. This minimalist model generalizes to predict defect energetics and elemental mixing without any structural inputs.

For mechanical properties, Tomoko Fukuda et al. showed in Development of a Neural Network Simulator for Structure-Activity Correlation of Molecules: Neco. (6). Estimation of Mechanical Properties of Cr-Mo Steel, Ni Steel, Ni-Cr Steel and Ni-Cr-Mo Steel. that neural networks can map the non-linear relationships of alloying elements directly to the mechanical properties of structural steels, bypassing physical microstructure testing entirely.

Bridging the Gap to Experimental Synthesis

Structure-agnostic modeling is uniquely suited for high-throughput screening and experimental workflows where 3D structures are fundamentally inaccessible.

Consider the challenge of doping, where adding trace elements breaks crystal symmetry and makes traditional 3D simulations computationally prohibitive. In Predicting thermoelectric properties from chemical formula with explicitly identifying dopant effects, Gyoung S. Na designed a specialized neural network (DopNet) that maps chemical formulas directly to physical properties while explicitly capturing dopant effects, bridging the gap between theoretical modeling and practical experimental synthesis.

Furthermore, when exploring entirely unknown chemical spaces, composition-only models allow researchers to prioritize experimental efforts before any structures exist. In Discovery of a Low Thermal Conductivity Oxide Guided by Probe Structure Prediction and Machine Learning, Quinn Gibson et al. successfully discovered an oxide with record-low thermal conductivity by training machine learning models to screen millions of compositions based solely on formula-level inputs.

Similarly, for complex porous materials, Zhonglin Cao et al. developed MOFormer: Self-Supervised Transformer model for Metal-Organic Framework Property Prediction, a structure-agnostic Transformer that predicts properties using text-based connectivity strings rather than computationally expensive 3D geometry optimizations. By shifting the analytical burden from explicit 3D geometry to learned elemental embeddings and stoichiometric aggregation, structure-agnostic property prediction unlocks rapid exploration across vast, uncharted chemical spaces.

History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_composition_based_property_prediction_d7fc68cc,
  title = {composition-based property prediction},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/composition-based-property-prediction/rdr_d7fc68cc5b0d4b309145dbe108b3afb6},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_composition_based_property_prediction_d7fc68cc_v1,
  title = {composition-based property prediction},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/composition-based-property-prediction/rdr_d7fc68cc5b0d4b309145dbe108b3afb6/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake