Page 239 - Read Online

Page 4 of 29 Liu et al. J. Mater. Inf. 2026, 6, 18

312 data points for ductility, primarily recording the maximum dynamic compressive strength and
maximum impact fracture strain. To capture intrinsic material characteristics and underlying mechanisms,
raw experimental data were transformed through feature engineering into physically meaningful descriptors,
enhancing both model accuracy and interpretability [52-54] .

In line with previously reported works , the present work proposes 17 feature descriptors related to alloy
[55]
design, electronic structure, surface characteristics, and thermodynamic stability, which can influence alloy
performance [Table 1]. Some of the features were obtained from material constitutive equations [Table 2].
Atomic size difference can serve as an indicator for quantifying the degree of lattice distortion . Mixed
[56]
enthalpy (ΔH ), mixed entropy (ΔS ), and valence electron concentration (VEC) can predict the formation
mix
mix
of solid solutions or phase stability . The presence of ΔH and ΔS has been shown to promote the
[57]
mix
mix
formation of single-phase solid solutions, inhibit the precipitation of intermetallic compounds, and enhance
solid solution strengthening effects . The shear modulus (G) represents a material’s resistance to shear
[58]
deformation, while the elastic modulus (E) indicates resistance to tensile or compressive deformation during
the elastic deformation stage. The bulk modulus (K) reflects resistance to uniform compressive deformation.
These three moduli serve as fundamental parameters for mechanical properties, collectively describing the
elastic response under various loading conditions. The equivalent molar fractions [Mo]eq and [Al]eq
quantify the contribution of β/α-stabilizing elements to the strength of titanium alloys . The introduction of
[59]
EWF is intended to reflect differences in bond strength and electronic states among different atoms . It has
[60]
been reported that EWF significantly affects the yield strength of titanium alloy materials . The electron
[37]
density at the Fermi level has been confirmed by first-principles calculations as the core electronic structure
parameter that simultaneously governs both strength and ductility . The bulk/shear modulus (B/G) ratio
[61]
determines the strength-ductility trade-off: values below 1.71 indicate strength dominance, while values
above 1.71 indicate ductility predominance [62,63] . The Poisson ratio (v) is defined as the negative ratio of
transverse strain to axial strain under uniaxial tension or compression. Materials with a higher Poisson ratio
facilitate stress release through shear mechanisms during deformation, thereby enhancing ductility . In
[64]
addition, as shown in Table 1, some descriptors can be directly calculated from atomic percentages. The
composition-weighted descriptors are defined as the atomic percentages (at%) of ten elements, namely Ti,
Al, V, Mo, Zr, Fe, Si, Sn, Nb, and Cr, and can be expressed as Equation (1) :
[65]
1 Õ
¯
= (1)

=1
where X represents the property of a given constituent element, c denotes its atomic percentage (at%), and n
i
i
is the total number of alloying elements. For non-additive quantities (such as B/G), these are treated as
heuristic indicators of overall ductility or brittleness trends. The usefulness of these descriptors is ultimately
evaluated through feature selection and model performance rather than a priori assumptions.

To identify the most relevant descriptors for target properties, a three-stage feature selection methodology
was implemented. First, mutual information (MI) and Pearson correlation analysis were combined to
address both linear redundancy and nonlinear relevance. If a feature exhibited a Pearson correlation
coefficient exceeding 0.8 [Equation (2)], high linear redundancy was inferred, and a priority ranking based
on MI was employed to systematically eliminate low-MI features [Equation (3)] :
[55]
Í
=1 ( − ¯ )( − ¯ )
= p Í p Í (2)
2 2
=1 ( − ¯ ) · =1 ( − ¯ )
Õ Õ ( , )
( ; ) = ( , )log( ) (3)
( ) ( )
∈ ∈
where r denotes the Pearson correlation coefficient, I represents mutual information, and corr(x, y) quantifies