MagpieEX Descriptors for Materials Prediction

• MagpieEX descriptors are an extended set of composition-based features that partition elements into cation and anion subsets to capture bond ionicity and charge-transfer asymmetry.
• They integrate traditional statistical moments with explicit cation–anion metrics, improving predictions of vibrational and thermal properties, including a noted 9.93% increase in phonon frequency prediction accuracy.
• The methodology employs oxidation state analysis and physical bonding parameters, offering a systematic, data-efficient approach to machine learning for materials informatics.

MagpieEX descriptors are an extended set of composition-based features for materials property prediction, introduced to augment the traditional Magpie framework with a focus on cation–anion interaction metrics. The methodology partitions elements within a chemical compound into cationic and anionic subsets, computes separate atomic property averages for each, and introduces explicit descriptors for bond ionicity and charge-transfer asymmetry. MagpieEX is designed to encode physically meaningful attributes relevant to vibrational and thermal transport properties, providing a systematic approach to capturing interatomic bonding characteristics influential in phonon behavior, dielectric response, and lattice thermal conductivity (Li et al., 31 Dec 2025).

1. Motivation for Extending Magpie Descriptors

Traditional Magpie descriptors, as established by Ward et al. (2016), represent materials by computing statistical moments—mean, range, standard deviation—of a defined set of elemental properties over the composition’s stoichiometry. These elemental properties include Pauling electronegativity, atomic (covalent/ionic) radius, valence-electron count, ionization energy, electron affinity, and atomic polarizability. While these descriptors effectively summarize bulk chemical trends, they do not distinguish between the separate roles of cations and anions or quantify the degree of ionic or covalent bonding. Bond characteristics such as polarity and charge-transfer asymmetry are critical for vibrational mode spectra and thermal transport, motivating the creation of MagpieEX. This extension leverages oxidation-state analysis to separate the constituent species and forms descriptors that are more closely aligned with the underlying physics of lattice vibrations and bond-driven phenomena.

2. Feature Set Composition: Traditional and Novel Descriptors

MagpieEX comprises the standard Magpie statistics and an additional block of cation–anion interaction descriptors. The complete feature set consists of:

A. Traditional Magpie Descriptors

For each property $$P \in \{\text{EN}, r, n_\mathrm{val}, IE, EA, \alpha\}$$:

• $P_\mathrm{mean} = \sum_i x_i P_i$, atomic-fraction-weighted mean over all species
• $P_\mathrm{range} = \max_i P_i - \min_i P_i$
• $$P_\mathrm{std} = \sqrt{\sum_i x_i (P_i - P_\mathrm{mean})^2}$$
• Additional low-order moments and fraction-of-species statistics for a total of approximately 132 dimensions in the base representation.

B. MagpieEX Cation–Anion Interaction Block

For each $P$:

• $P_\mathrm{Cat} = \sum_{i\in \mathrm{Cat}} x_i P_i$: cation mean
• $P_\mathrm{An} = \sum_{i\in \mathrm{An}} x_i P_i$: anion mean
• $\Delta P = P_\mathrm{Cat} - P_\mathrm{An}$: difference

Plus two bond-derived metrics:

• Bond ionicity ($I_\mathrm{bond}$): quantifies ionic character
• Charge-transfer asymmetry ($A_\mathrm{ct}$): quantifies electron transfer asymmetry

This results in $P_\mathrm{mean} = \sum_i x_i P_i$0 features for the average/difference block and 2 for the bond metrics, yielding 20 new descriptors. Combined with the base, the typical MagpieEX vector has about 152 dimensions.

Descriptor Block	Number of Features	Properties Included
Traditional Magpie	~132	EN, r, n_val, IE, EA, α (moments/stat)
Cation–anion interaction	18	EN, r, n_val, IE, EA, α (Cat, An, Δ)
Bond metrics	2	$P_\mathrm{mean} = \sum_i x_i P_i$1, $P_\mathrm{mean} = \sum_i x_i P_i$2

3. Mathematical Formulations

Cation/Anion Statistics:

Given a chemical formula with elements $P_\mathrm{mean} = \sum_i x_i P_i$3, atomic fractions $P_\mathrm{mean} = \sum_i x_i P_i$4, and atomic property $P_\mathrm{mean} = \sum_i x_i P_i$5, define cation and anion sets ($P_\mathrm{mean} = \sum_i x_i P_i$6, $P_\mathrm{mean} = \sum_i x_i P_i$7) via oxidation state assignment (e.g., Pymatgen bond-valence-sum):

$P_\mathrm{mean} = \sum_i x_i P_i$8

Bond Ionicity ($P_\mathrm{mean} = \sum_i x_i P_i$9):

Using the Pauling electronegativity ($P_\mathrm{range} = \max_i P_i - \min_i P_i$0), define:

$P_\mathrm{range} = \max_i P_i - \min_i P_i$1

$P_\mathrm{range} = \max_i P_i - \min_i P_i$2

Values range from 0 (covalent) to 1 (ionic).

Charge-Transfer Asymmetry ($P_\mathrm{range} = \max_i P_i - \min_i P_i$3):

With effective cation ionization energy and anion electron affinity:

$P_\mathrm{range} = \max_i P_i - \min_i P_i$4

$P_\mathrm{range} = \max_i P_i - \min_i P_i$5

Large positive $P_\mathrm{range} = \max_i P_i - \min_i P_i$6 indicates strong polarity; near zero indicates symmetric charge transfer.

4. Algorithmic Workflow

The computational process for generating MagpieEX follows these steps:

$P_\mathrm{range} = \max_i P_i - \min_i P_i$9

5. Physical Interpretations and Relevance to Lattice Dynamics

• Electronegativity (EN_Cat, EN_An, ΔEN): Determines bond polarity; large ΔEN promotes stiffer optical modes and pronounced LO–TO splitting.
• Radius (r_Cat, r_An, Δr): Governs bond length and lattice strain; pronounced mismatches modulate phonon group velocity.
• Valence Electrons (n_val_Cat, n_val_An, Δn_val): Relate to bond order, affecting the stiffness of acoustic modes.
• Ionization Energy/Electron Affinity (IE_Cat, EA_An, Δ(IE–EA), A_ct): Gauge ease of electron transfer and polarization, central for dielectric screening and phonon–electron coupling.
• Polarizability (α_Cat, α_An, Δα): Describes ion lattice response and affects anharmonic scattering rates.
• Bond Ionicity (I_bond): Correlates with the degree of ionic versus covalent bonding; higher values typically lead to more pronounced phonon scattering and reduced lattice thermal conductivity.
• Charge-Transfer Asymmetry (A_ct): Reflects the symmetry of electron transport; large asymmetry can localize vibrational modes, impacting phonon lifetimes.

By explicitly encoding these distinctions, MagpieEX vectors relate composition to vibrational frequencies, Grüneisen parameters, and phonon–phonon scattering rates.

6. Dimensionality, Computational Cost, and Integration

• Dimensionality: The standard Magpie descriptor yields ~132 features; the MagpieEX block adds 18 features for cation/anion mean and difference terms and 2 further bond metrics, resulting in a total vector of approximately 152 dimensions.
• Computational Cost: Oxidation-state assignment (e.g., with Pymatgen) scales as $P_\mathrm{range} = \max_i P_i - \min_i P_i$7 but remains negligible for formulas containing ≤10 species. Weighted averages and property lookup for each atomic attribute are $P_\mathrm{range} = \max_i P_i - \min_i P_i$8. Typical runtime per formula is a few milliseconds on a modern CPU.
• Integration Guidelines: Features should be standardized (zero mean, unit variance) across the dataset, particularly Δ-type descriptors. Special handling for cases with only cations or anions is recommended, such as setting ΔP = 0 when one subset is empty. For small datasets, dimensionality reduction (e.g., PCA, UMAP) or regularization is advised. When combining MagpieEX with learned embeddings (such as from graph neural networks), features should be normalized separately or fed into dedicated model heads to avoid scale imbalances (Li et al., 31 Dec 2025).

7. Significance and Application in Machine Learning for Materials

MagpieEX provides a physically interpretable, data-efficient extension to traditional composition-based feature sets in materials informatics. Its explicit treatment of cation–anion partitioning and introduction of bond-level metrics—ionicity and charge-transfer asymmetry—enables superior encoding of vibrational and thermal transport phenomena, particularly in lattice thermal conductivity and phonon frequency prediction tasks. The approach is compatible with both tabular foundation models (e.g., TabPFN) and graph neural network pipelines, and has demonstrated improved predictive performance over more complex, structure-based models in various MatBench tasks, with a reported 9.93% increase in phonon frequency prediction accuracy and effectiveness in modeling phonon–phonon scattering and atomic mass contrast. A plausible implication is that MagpieEX descriptors can serve as general-purpose, physically grounded inputs for accelerated discovery and characterization of functional materials in small-data regimes (Li et al., 31 Dec 2025).