An active learning workflow for predicting misfit volume in body-centered cubic refractory high-entropy alloys Authors: Shunshun Liu, Prasanna V Balachandran Published: 2026 ABSTRACT AbstractRefractory high-entropy alloys (RHEAs) exhibit exceptional high-temperature mechanical properties. However, a mechanistic understanding of their yield strength requires accurate determination of the misfit volume descriptor (δ), which quantifies the local volume change due to the size and electronic heterogeneity of constituent elements in the solid solution. Traditional approximations, such as Vegard’s Law, fail to capture local atomic relaxation and electronic structure effects in these compositionally complex systems. An active learning workflow is developed that integrates density functional theory calculations with ensemble machine learning (ML) to efficiently predict the δ-descriptor across 126 equiatomic quinary body-centered cubic (BCC) RHEAs. Partial dependence plots and symbolic regression reveal that atomic size mismatch directly influences the δ-descriptor, while electronegativity variations provide electronic structure compensation. Incorporating ML-predicted δ-values into the Maresca-Curtin mechanistic model achieves good correlation with experimental yield strengths across diverse BCC RHEA compositions.Keywords: Refractory high-entropy alloys, Misfit volume, Body-centered cubic, Active learning, Machine learning, Mechanistic models, Yield strengthSubject terms: Materials science, Mathematics and computing, Physics Introduction Refractory high-entropy alloys (RHEAs) have emerged as an important class of structural materials for high-temperature, load-bearing applications due to their combination of high melting points and exceptional strength1–3. These alloys are composed of multiple refractory elements in equiatomic or near-equiatomic proportions and exhibit mechanical strength retention at high homologous temperatures, where conventional structural materials have been shown to rapidly degrade. This makes RHEAs particularly attractive in applications such as aerospace propulsion, nuclear energy systems, and advanced power generation4–6. The fundamental microstructural mechanisms governing the high-temperature mechanical properties of RHEA remain actively debated. Multiple research groups have identified several key factors contributing to these exceptional properties3,6–15. In single-phase, body-centered cubic (BCC) RHEAs, solid-solution strengthening has been shown to result from modulus mismatches, atomic size differences, and electronic structure variations among constituent elements, which induce severe lattice distortions. These distortions generate high friction stress and act as strong barriers to dislocation motion, leading to high dislocation density and the activation of multiple slip systems3,14. Chemical short-range order (SRO) is also considered critical for explaining the observed resistance to plastic deformation in BCC-based RHEAs8,11,12. In multi-phase RHEAs, precipitation strengthening from secondary phases provides additional hardening at elevated temperatures6,7,10. Furthermore, improved control of interstitial elements and grain boundaries has been demonstrated to enhance the mechanical properties of RHEAs13,15. Plastic deformation in pure BCC metals is primarily controlled by the high Peierls barrier associated with screw dislocations16, which dictates the strong temperature dependence of the yield strength (). However, literature on BCC alloys have revealed a more complex mechanism where the edge dislocations have a decreasing mobility with increasing solute concentration17. Recent studies on BCC HEAs also suggest that edge dislocation is competing with screw dislocation or dominating the strengthening in the BCC alloy system especially at elevated temperature3,18–20. This shift in strengthening mechanism offers significant advantages for predictive modeling. Screw dislocation models often require complex first principles calculations and empirical parameter tuning, as exemplified by the Suzuki-Rao model21,22. In contrast, the mechanistic framework for edge dislocation strengthening, such as the Maresca-Curtin model, is largely parameter-free23,24. This provides a direct link between atomic descriptors from first principles calculations (e.g., density functional theory, DFT) and the macroscopic property. The predictive accuracy of the Maresca-Curtin mechanistic model depends on several physical descriptors. Among these descriptors, the alloy misfit volume (δ) holds particular physical significance. The δ-descriptor quantifies local effective-volume fluctuations resulting from atomic-size mismatch and electronic charge transfer in metallic alloys25–27. These collective effective-volume fluctuations manifest as local lattice distortions induced by solute atoms within the alloy matrix. Despite the physical significance of δ, efficient computation using high-throughput DFT remains challenging. Traditionally, this descriptor is approximated by Vegard’s Law, which assumes a linear rule-of-mixture based on the properties of the constituent pure elements28,29. Although convenient, this linear approximation frequently fails in alloy systems where complex electronic interactions and charge transfer induce significant composition dependence25,30. This failure is especially pronounced in BCC-based RHEAs containing 3d, 4d, and 5d elements simultaneously, where deviations from Vegard’s Law are significant31–33. In addition to its role in solid-solution strengthening, the δ-descriptor has been shown to affect radiation damage tolerance and diffusion kinetics. For instance, increased misfit volume in CoCrNi enhances grain boundary absorption of self-interstitial atoms compared to pure Ni or dilute NiCr alloys34. The addition of Pd to NiCoCrFe raises the activation energy for interstitial motion while reducing that for vacancy diffusion, a phenomenon attributed to the larger misfit volume of Pd35. Furthermore, alloys with increased local lattice distortion have been shown to immobilize irradiation-induced defects, such as interstitials and vacancies, which enhances irradiation tolerance36. Taken together, these findings demonstrate that δ serves as a fundamental atomic-scale descriptor with broad physical relevance in compositionally complex alloys. An alternative computational approach uses root-mean-square atomic displacements (RMSAD) as a structural proxy that correlates with δ27,37,38. However, Nöhring et al. contend that this correlation relies primarily on empirical observations rather than a detailed mechanical explanation of dislocation-solute interactions27. RMSAD is calculated from a single DFT-relaxed structure at a specified equiatomic composition by quantifying atomic deviations from ideal lattice positions. Several phenomenological descriptors for solid-solution strengthening, including valence electron concentration (VEC)39, electronegativity difference ()26,40, and 41, have also been proposed in the literature. For instance, was introduced by Oh et al. as a proxy for atomic-level pressure fluctuations, which is the same physical quantity that the δ-descriptor quantifies from first principles26. Unlike the descriptors above, only the δ-descriptor maintains a rigorous connection to dislocation mechanics and solute-dislocation interactions27. The calculation of δ using DFT requires evaluating the misfit volumes for each solute element. This approach involves constructing and relaxing multiple supercells with systematically varied compositions near the target alloy, resulting in significantly greater computational demands. This work addresses existing computational limitations by developing an active learning (AL) workflow to rapidly predict the δ-descriptor in equiatomic, quinary RHEAs using DFT calculations. Yin and Curtin33 define the misfit volume of solute element n () in a multicomponent alloy as the volume change computed from replacing an average matrix atom with a solute atom of type n in a large SQS supercell, accounting for full atomic relaxation. The δ-descriptor is then calculated as the collective contribution of all elemental misfit volumes (see Equation 1 in the Methods section for details). The AL workflow, shown in Fig. 1, efficiently samples the compositional space and constructs a surrogate machine learning (ML) model to predict the δ-descriptor. This approach substantially reduces computational cost by minimizing the number of DFT calculations. Symbolic regression analysis of the ML predictions provides interpretable, closed-form analytical expressions that quantify the physical contributions of atomic-size mismatch and electronegativity variations to δ. Incorporating the ML-predicted δ-descriptor into the Maresca-Curtin edge dislocation strengthening model shows good correlation with experimental . The DFT-derived δ-values provide a more rigorous physical basis than Vegard’s law, explicitly accounting for local atomic relaxation and electronic structure effects. The alloy misfit volume (δ) prediction workflow. The quinary RHEA design space (S) consists of five elements selected from three 3d elements (Ti, V, Cu), three 4d elements (Zr, Nb, Mo), and three 5d elements (Hf, Ta, W). Consequently, the total number of possible RHEA compositions is . Each composition is represented by three descriptors: the standard deviation of atomic volume (sdAtomicVolume, ), electronegativity (sdElectronegativity, ), and the number of unfilled d-orbital electrons (sdNdUnfilled, ). Ten alloy compositions were randomly chosen from this design space to form the initial training set, T. Eleven special quasi-random structures (SQS) were generated for each composition. Atomic positions and lattice parameters were relaxed using density functional theory (DFT) calculations to obtain equilibrium volumes, from which elemental misfit volumes () and the δ-parameter are derived. Subsequently, an active learning workflow trained an ensemble of support vector regression models (eSVR) to establish the input-output relationship, , where and ε represents model uncertainty. Information condensing active learning (ICAL) was employed as the acquisition function to recommend new compositions from the virtual set, S-T. These new compositions were then processed through SQS generation and DFT geometry optimization to calculate δ. The updated δ values were incorporated into the training set T for the next iteration. Methods First principles calculations for crystal structure relaxation Density functional theory (DFT) calculations were performed using the open-source plane-wave pseudopotential code Quantum ESPRESSO42,43. The generalized gradient approximation (GGA) with the revised Perdew-Burke-Ernzerhof (PBE) for solids exchange-correlation functional (PBEsol) and PSLibrary ultrasoft pseudopotentials were employed for all computations44–46. The crystal structures and their corresponding Quantum ESPRESSO input files are managed by Atomic Simulation Environment (ASE) python package47. A kinetic energy cutoff of 60 Ry was applied for the wavefunctions. The Brillouin zone was sampled using a Γ-centered Monkhorst-Pack (MP) mesh48 with a minimum k-point spacing of 0.045 Å in reciprocal space. Structural optimization proceeded until forces on all atoms converged below 10 Ry/Bohr and the absolute cell pressure converged below 0.1 Kbar. Misfit volume calculations The special quasirandom structures (SQS) concept was employed to simulate five-component, or quinary, RHEAs49. Crystal structures were generated using the Alloy Theoretic Automated Toolkit (ATAT)50. The quinary system was initially defined with an equiatomic composition comprising five constituent elements arranged in a BCC lattice. This composition was optimized within a 120-atom supercell, with each element occupying 24 lattice sites. This initial, equiatomic composition is denoted as “ABCDE”. The DFT approach for calculating the misfit volume of a constituent element n, denoted as , was described by Yin and Curtin33, assuming a pseudo-binary alloy environment. Beyond the equiatomic configuration, the atomic ratio of each element in the system was perturbed to enable fitting of the atomic misfit for individual elements. Two additional configurations were generated for each element type by perturbing the atom counts. In the first perturbation, one atom each from elements B, C, D, and E was randomly selected and reassigned, resulting in the configuration “ABCDE”. Subsequently, an SQS was regenerated based on this composition. Similarly, another configuration was created by randomly selecting four A atoms and assigning one atom to each of the other element types, yielding “ABCDE”. In total, 11 configurations were generated, consisting of one equiatomic and ten perturbed near-equiatomic configurations. The elemental misfit volumes () are obtained from constrained linear regression (Figure S1 in the Supplementary Information, SI), where each slope yields the misfit volume for the corresponding element and the constraint is enforced. The δ-descriptor is calculated based on the fitted elemental misfit volume value as shown in Equation 1: where and denote the fully relaxed supercell volume of the alloy and the atomic ratio of the composition, respectively. In addition to the δ-descriptor, the RMSAD was calculated for each equiatomic composition using fully relaxed DFT calculations. The RMSAD is expressed as follows in Equation 2: where is the position of atom i in the fully relaxed structure (with local lattice distortions), is the corresponding ideal lattice position, N is the total number of atoms, and the angular brackets denote averaging over multiple random configurations27,51–53. Active learning workflow An active learning (AL) workflow, as described by Biswas et al. 54, was employed to establish a quantitative relationship between alloy composition and the δ-descriptor. Information condensing active learning (ICAL) served as the acquisition method in this study55. Each RHEA composition was represented by elemental features generated using the Magpie framework56. Three variance-based features were used as ML inputs: the standard deviation of atomic volume (sdAtomicVolume), electronegativity (sdElectronegativity), and the number of unfilled d-orbital electrons (sdNdUnfilled). Because Magpie does not natively provide standard deviation features, these descriptors were calculated directly from elemental-level Magpie values, assuming equiatomic composition. Specifically, for each alloy, the standard deviation was computed across the values of its constituent elements. For instance, the sdAtomicVolume of the quinary composition ABCDE corresponds to the standard deviation of the Magpie AtomicVolume values of elements {A, B, C, D, E}. These features capture compositional variations known to influence phase stability and solid-solution strengthening in compositionally complex alloys57–61. The alloy composition design space includes nine transition metal elements: Ti, V, Cu, Zr, Nb, Mo, Hf, Ta, and W. Although other elements are commonly reported in the RHEA literature, Cu (face-centered cubic as a pure element) was intentionally introduced to induce significant atomic distortions during DFT relaxation. The total RHEA design space of quinary alloys comprises equiatomic alloy compositions. Ten compositions were randomly sampled as the initial training set. For each selected composition, the δ-descriptor was calculated via DFT. An ensemble of non-linear support vector regression (eSVR) models with radial basis function (RBF) kernels62–64 was employed to model the relationship between the selected compositional features and the δ-descriptor, thereby establishing a mapping between RHEA composition and atomic misfit. Hyperparameters (gamma and cost) for individual SVR estimators within the ensemble were independently optimized via grid search. The optimal ensemble size was selected from {5, 10, 20, 25, 50, 100} estimators based on the maximum on the training set. New RHEA compositions were recommended based on the ICAL score, which provides recommendations without distributional assumptions. Each new composition underwent the same DFT-based δ-descriptor calculation and was subsequently added to the training set for the next iteration. The adaptive learning loop terminated when (1) the tenth iteration was reached or (2) the maximum training of the model did not improve for three consecutive iterations. Post-Hoc model explanation Two post-hoc model explanation methods are used to uncover the insights learned by the trained eSVR models: (1) permutation-based global feature importance (GFI) and (2) partial dependence plots (PDPs)54,65–68. While GFI ranks input feature importance based on the loss increase due to losing the feature information, PDP analysis reveals the marginal functional relationship of each input feature’s contribution to the predicted δ as learned by the eSVR models. The insights gained from PDPs are further used to inform the symbolic regression model. Error bars in the GFI and PDP analyses represent the standard deviation across the ensemble of SVR models. Symbolic regression Symbolic regression using Grammatical Evolution (GE) was performed to identify interpretable, closed-form mathematical relationships between the three input features (sdAtomicVolume, sdElectronegativity, and sdNdUnfilled) and the eSVR-predicted δ-descriptor across the entire composition space of 126 RHEAs54,69–71. A context-free grammar was defined with allowed operations including addition (+), subtraction (−), and multiplication (×), and functions including hyperbolic tangent (tanh) and quadratic or squaring (). The choice of these functions was informed by the PDP analysis. The cost function minimized , where is the coefficient of determination. GE was configured with 5000 iterations using an evolution strategy optimizer, with recursion depth controlled through wrappings = 40, sequence length = 60, and maximum depth = 30 to prevent stack overflow while maintaining sufficient expression complexity. The symbolic regression analysis was performed independently 100 times using unique random seeds (424 + i, where i = 1, … , 100), yielding 100 candidate expressions with their associated cost values. To propagate eSVR prediction uncertainty while preserving physical constraints, the eSVR-predicted δ-descriptor was sampled from a truncated normal distribution bounded at zero to prevent non-physical negative values. To find a simple yet accurate mathematical model, the 100 discovered expressions were examined for how often variables were used, how often certain functions were retained, and how often parts were repeated. Based on these evaluations, a final model was selected for Bayesian analysis72 and further interpretation. Results Training surrogate models via active learning Figure 2 shows the predictive performance of the optimized eSVR model following nine AL iterations. The AL workflow concluded after nine iterations because the eSVR performance, as quantified by the score, saturated after the 6 iteration (Figure S2 in the SI). The final model, selected from the ninth iteration, exhibited high stability, with values consistently ranging from 0.961 to 0.973 during the final refinement phases. To further quantify the prediction error during AL, the root-mean-square error (RMSE) and mean absolute error (MAE) of the eSVR model were also evaluated as a function of AL iteration (Figure S3). Both RMSE and MAE decrease during the early iterations, consistent with the trend observed in the score. This consistency indicates good predictive capability across the compositional space. The final eSVR model attains scores of 0.962 on the training set and 0.81 on an independent test set comprising 14 randomly selected compositions after nine AL iterations. The lower in the test set primarily results from four test points at the upper boundary of the predicted δ range. The performance of the eSVR model after nine iterations is shown for both the training set (filled black circles) and the independent test set (filled red triangles). The x-axis represents the δ-descriptor from the DFT calculation, while the y-axis shows the eSVR model predictions with error bars indicating the standard deviation of the predicted values from the ensemble of SVR models. The coefficient of determination () is 0.962 and 0.812 for the training and independent test set, respectively. Principal component analysis (PCA) visualization (Figure S4 in the SI) shows that these compositions occupy regions of latent space that were not sampled during model training, leading to out-of-distribution predictions. Despite this, the agreement between DFT-calculated and eSVR-predicted δ remains good. The performance of eSVR in these undersampled regions demonstrates its capacity for generalization beyond the immediate vicinity of the training data. The PCA visualization also shed light into the sampling strategy and exploration behavior of the AL workflow (Figure S4 in the SI). The initial training data are concentrated in the central-left region of the PC1-PC2 latent space, whereas the points selected by the ICAL acquisition function are scattered across the periphery or the outer regions of the composition space, extending well beyond the initial training points. This distribution demonstrates that the ICAL algorithm systematically identifies and prioritizes underrepresented regions, thereby exploring the boundaries of the compositional space by broadening the compositional coverage and enhancing model robustness. Table S1 provides a direct comparison of computational costs for exhaustive DFT, Vegard’s law, and the AL workflow. The results indicate that substantial cost reductions can be achieved without compromising predictive accuracy. The GFI (Figure S5 in the SI) and PDPs (Fig. 3) provide insights about the trained eSVR models predicting the δ-descriptor. The GFI analysis identifies sdAtomicVolume as the primary predictor of the δ-descriptor, followed by sdNdUnfilled and sdElectronegativity. The PDP for sdAtomicVolume [Fig. 3(a)] shows a nearly linear positive correlation with the predicted δ, consistent with the established understanding that atomic-size mismatch induces local lattice distortions in these alloy systems. Conversely, sdElectronegativity [Fig. 3(b)] exhibits a non-linear inverse relationship with predicted δ, characterized by a steep initial decline followed by saturation at higher values, resembling a negative hyperbolic tangent (tanh). This pattern indicates that electronegativity affects δ via charge-transfer effects that saturate beyond a critical threshold. The sdNdUnfilled input feature [Fig. 3(c)] shows a weak parabolic dependence with substantial prediction uncertainty and a limited predicted δ range, suggesting a minor contribution compared to sdAtomicVolume and sdElectronegativity. The contrasting trends of sdAtomicVolume (positive correlation) and sdElectronegativity (negative correlation) emphasize the complex interplay between geometric size mismatch and electronic charge-transfer effects in governing δ, a phenomenon extensively discussed in the HEA literature31,32. Partial dependence plots (PDPs) of the trained eSVR model after the 9 AL iteration. The marginal effects of individual input features: (a) sdAtomicVolume, (b) sdElectronegativity, and (c) sdNdUnfilled on δ-descriptor as learned by the trained eSVR models are captured by the plots. Descriptors sdElectronegativity and sdNdUnfilled show a non-linear relationship with δ, whereas sdAtomicVolume shows a linear relationship with δ. Error bars indicate the standard deviation of the predicted values from the eSVR models. The DFT-simulated δ-descriptor was compared with Vegard’s Law (linear rule-of-mixtures) and RMSAD derived from DFT calculations. Although Vegard’s Law estimated δ correlates with the DFT-simulated δ-descriptor, the coefficient of determination () is 0.81 (see Figure S6 in the SI), which is substantially lower than that of the ML prediction (see Figure 2). This discrepancy arises from the non-linear interactions introduced by sdElectronegativity, which is not accounted for by Vegard’s law. Conversely, RMSAD also correlates with the DFT-simulated δ; however, the Pearson correlation coefficient (PCC) was used instead of due to the differing physical meanings of RMSAD and δ. The PCC between RMSAD and the DFT-simulated δ is 0.55 (see Figure S7 in SI), indicating a weaker linear correlation. Although numerous studies employ Vegard’s Law or RMSAD as proxies for δ-descriptor, these results suggest that the relationship is relatively weak. Yield strength prediction The Maresca-Curtin edge dislocation model for BCC RHEA systems23,73 was used to compare predicted with experimental data at room temperature. The model requires three composition-dependent inputs: (1) the elastic constants (C, C, C), (2) the lattice constant, and (3) the δ-descriptor. For inputs (1) and (2), the ML models previously developed by Liu et al. were used to predict elastic constants74 and lattice constants75 from chemical composition. The δ-descriptor was obtained from the AL workflow developed in this work. With all three predicted inputs determined solely from composition, was predicted using the Maresca-Curtin model without empirical fitting or parameter calibration. Figure 4 shows the predicted vs experimentally measured data. Prediction performance was evaluated by fitting a linear relationship of the form , where the slope k quantifies systematic bias in the mechanistic predictions and measures the strength of the linear correlation. The data used to generate Figure 4 are provided in Table S2 of the SI. Room temperature predicted yield strength () from the Maresca-Curtin mechanistic model () is compared with experimental for BCC RHEAs: (a) Zr, Hf-free alloys (see Table S2a) and (b) Zr, Hf-containing alloys (see Table S2b). The gray dashed line represents the x=y trend. The blue dotted line shows the linear fit , with k=1.042 and for (a) Zr, Hf-free alloys, and k=1.553 and for (b) Zr, Hf-containing alloys. Data points are categorized by alloy components from ternary to senary. Since the ML model for δ was trained on quinary alloys (blue diamonds), ternary alloys (black circles), quaternary alloys (red squares), and senary alloys (violet triangles), these are considered extrapolations. The comparison between predicted and experimental reveals two distinct behaviors based on alloy composition (Figure 4). Zr and Hf-free refractory RHEAs exhibit near-ideal predictions with k=1.042 and [Figure 4(a)]. In contrast, Zr- and Hf-containing RHEAs show systematic overestimation with k = 1.553 but maintain excellent correlation [, Figure 4(b)]. The high values in both cases indicate that the mechanistic model, incorporating the three ML inputs, accurately captures trends by correctly ranking compositions by relative strength, despite differences in absolute magnitudes between Zr- and Hf-containing systems. For Zr- and Hf-free RHEAs, the mechanistic model achieves high accuracy with k approximately equal to 1, indicating minimal systematic bias. This is shown in Figure 4(a). The ML-predicted elastic constants yield Voigt shear moduli of 85–100 GPa. Although these values exceed first-principles estimates of 60 to 75 GPa76,77, they remain within uncertainty bounds. This discrepancy arises because Liu et al.’s elastic constant models were primarily trained on face-centered cubic (FCC) alloys and are applied here in an extrapolation regime for BCC RHEAs. Nevertheless, the modest error in shear modulus results in only minor deviations in predicted . A notable outlier is the ternary alloy NbTaTi, for which the model predicts 1203± 916 MPa compared to the experimental value of 620 MPa78. This discrepancy likely reflects limitations in the ML-predicted δ-descriptor for ternary compositions, where extrapolation beyond the quinary training data may fail to adequately capture atomic misfit behavior. For Zr- and Hf-containing RHEAs, the systematic overestimation (k = 1.553) results directly from inaccuracies in the predicted elastic constants, as shown in Figure 4(b). The ML models estimate shear moduli between 70 and 80 GPa, which is approximately two to three times higher than the literature-reported values of 20 to 50 GPa for Zr- and Hf-based BCC RHEAs79,80. According to the Maresca-Curtin model23,27, the yield strength scales with shear modulus () and the δ-descriptor as shown in Equation 3: where f(δ ) is a non-linear function of δ capturing the collective effect of misfit volume. Since the term contributes linearly, the approximate twofold overestimation in directly explains the observed scaling factor of about 1.5. Notably, the near-perfect demonstrates that the AL-predicted δ-descriptor accurately captures the physical contributions governing lattice distortion and relative strengthening trends. Consequently, the systematic offset originates entirely from the elastic constant predictions rather than from the δ-descriptor. To further evaluate the current workflow, predicted values were compared with Vegard’s law-based predictions reported in Yin et al.73, using the same Maresca-Curtin mechanistic framework. The results are shown in Figure 5. In this comparison, the Vegard’s law baseline uses lattice constants, elastic constants, and δ-descriptors, all estimated from Vegard’s law. In contrast, the present workflow incorporates ML-predicted lattice constants, elastic constants obtained from previously published studies74,75, and δ-descriptors predicted through AL in this study. Three primary trends emerge from this comparison. First, for the Zr/Hf-containing alloys HfNbTiZr and NbTiZr, the present workflow overestimates the experimental , consistent with the previously discussed overestimation of ML-predicted shear moduli. Second, for MoNbTaVW and MoNbTaW, the present workflow produces predictions comparable to Vegard’s law baseline. Third, for MoNbTaTiW and MoNbTaTiVW, the present workflow yields predictions closer to experimental values than those from Vegard’s law baseline. These findings indicate that the principal limitation of the present workflow is the prediction of elastic constants for Zr/Hf-containing BCC RHEAs, whereas for the Zr/Hf-free alloys examined, the workflow performs comparably to or better than Vegard’s law baseline. Room temperature yield strength () predictions from the present workflow are compared with Vegard’s law baseline from Yin et al.73 using the Maresca-Curtin mechanistic model. Experimental is plotted on the x-axis and predicted on the y-axis. The dashed line indicates the ideal y = x relationship. The present study, represented by black dots, employs ML-predicted lattice constants, elastic constants from previous studies74,75, and AL-predicted δ-descriptors from this study. The error bars represent the prediction uncertainty propagated from the various ML models, including those developed in this study. The Vegard’s law baseline, shown as red triangles, utilizes lattice constants, elastic constants, and δ-descriptors, all estimated from Vegard’s law. Symbolic regression for closed-form analytical expression The trends identified through PDPs, specifically the linear relationship with sdAtomicVolume and the negative tanh behavior of sdElectronegativity, informed the symbolic regression search for a closed-form mathematical expression. An analysis of 100 GE-generated expressions revealed the following patterns: sdAtomicVolume appeared in all expressions, sdElectronegativity in 99, sdNdUnfilled in 34, tanh transformations in 89, and quadratic terms in 67. Based on this frequency analysis, the following function given in Equation 4 was selected for Bayesian linear regression: The fitted coefficients follow normal distributions []: , , and . The symbolic regression model attains an value of 0.99, demonstrating strong agreement with the eSVR predictions. Residual analysis indicates that errors are randomly distributed and free of systematic bias, confirming model adequacy. This model offers further insights into the input variables governing the misfit volume of quinary RHEAs analyzed in this study, thereby complementing the PDP analysis. The positive coefficient for sdAtomicVolume () indicates that greater variations in atomic size lead to increased misfit volume. The negative coefficient for () indicates a compensation mechanism whereby increasing electronegativity variation reduces δ. The negative tanh form suggests that this compensation decreases monotonically and saturates at high values of sdElectronegativity, potentially due to electronic charge-transfer effects. The GE-based symbolic regression results quantitatively confirm the trends identified through PDP analysis, demonstrating consistency between the data-driven model interpretation and the analytically derived closed-form expression. Discussion Analysis of the eSVR-predictions across 126 quinary RHEA compositions shows that the choice of constituent elements strongly dictates the magnitude of δ. Compositions containing Zr and Cu exhibit significantly larger δ values compared to those lacking these elements. The presence of V and Hf also produces a positive effect on the δ-descriptor. In contrast, refractory elements W, Ti, Nb, Ta, and Mo reduce δ reflecting their closer atomic size compatibility within the BCC matrix. Pairwise elemental analysis also supports this observation. Compositions with Cu-Zr co-substitution show the highest mean δ, and the five top-ranked alloys share a Cu-Hf-V-Zr backbone. Conversely, the alloys with the lowest δ values are exclusively from the {Mo, Nb, Ta, Ti, W} subspace, confirming that deliberate inclusion of Zr and Cu is the most effective compositional strategy to maximize misfit volume in this design space. However, this approach requires careful control of Cu and Zr additions to maintain the single-phase BCC solid solution due to phase stability constraints. These compositional trends are consistent with prior reports in BCC RHEAs73,81–84. The differing predictive behaviors between Zr- and Hf-free and Zr- and Hf-containing RHEAs allow decomposition of model accuracy, revealing elastic constant models for BCC systems as the primary source of uncertainty. This decomposition confirms that the two components, (1) the AL-predicted δ-descriptor and (2) the elastic constant models, independently contribute to the overall prediction error. Further refinement is required mainly for the elastic constant model. The high correlation observed for Zr- and Hf-containing RHEAs indicates that the AL workflow effectively captures the physical mechanisms governing solid-solution strengthening, even within compositional regimes with limited elastic constant data. More broadly, the mechanistic model incorporating ML inputs accurately ranks RHEAs by relative strengthening potential across both alloy families, demonstrating its suitability for guiding compositional design in its current form. Quantitatively accurate predictions are achievable once BCC-specific elastic constant models are developed. Although the PDP for sdAtomicVolume reveals a nearly linear relationship with the δ-descriptor [Figure 3(a)], sdAtomicVolume alone is not sufficient. Linear regression analysis of the eSVR-predicted misfit volumes across 126 RHEAs demonstrates that the contribution of tanh is statistically significant (t-statistic = -21.02, p-value ), confirming that sdElectronegativity is a necessary predictor. This finding aligns with established physical principles26,40,41: differences in electronegativity among constituent elements drive charge redistribution, which partially offsets local volume perturbations resulting from atomic-size mismatch. Further support is provided by symbolic regression, which identified sdElectronegativity in 99 out of 100 GE runs, thereby reinforcing its importance in predicting the δ-descriptor. Conclusion An AL workflow was developed to rapidly predict the δ-descriptor in equiatomic quinary BCC RHEAs. By integrating DFT calculations with iterative compositional sampling, a surrogate model was trained to capture the non-linear compositional dependencies of misfit volume. Post-hoc model explanation using PDPs and symbolic regression revealed two key findings: (1) linear atomic size effects influence misfit volume without saturation, and (2) variations in electronegativity provide substantial electronic compensation. The DFT-informed δ predictions provide a more physically representative descriptor of misfit volume than Vegard’s law approximations, which neglect local atomic relaxation effects. The integration of the AL-predicted δ-descriptor into the parameter-free Maresca-Curtin mechanistic model enables predictions of . For Zr- and Hf-free RHEAs, the model achieves good accuracy. However, for systems containing Zr- and Hf-based RHEAs, systematic overestimation arises from inaccuracies in elastic constant predictions rather than the δ-descriptor. These findings indicate that the AL-predicted δ-descriptor reliably captures the local atomic distortion trends across the BCC RHEA design space and underscores the need to develop improved elastic constant models specifically calibrated for BCC refractory systems to achieve quantitative accuracy. The current ML model was trained exclusively on equiatomic quinary BCC RHEAs. Expanding the training dataset to include binary, ternary, quaternary, and senary compositions would better represent short-range chemical interactions and improve model transferability within the BCC RHEA design space. Supplementary Information Supplementary Information. Acknowledgements The authors acknowledge Research Computing at The University of Virginia for providing computational resources and technical support that have contributed to the results reported within this publication. The computational work also used PSC Bridges-2 GPU resource through allocation #MAT220039 from the Advanced Cyberinfrastructure Coordination Ecosystem: Services & Support (ACCESS) program, which is supported by U.S. National Science Foundation grants #2138259, #2138286, #2138307, #2137603, and #2138296. Author contributions The study was designed by S.L. and P.V.B. The manuscript was prepared by S.L. and P.V.B. S.L. ran the DFT calculations and active learning, under the supervision of P.V.B. Both authors participated in discussing the results, writing, and providing feedback on the manuscript. Data availability The relaxed crystal structure of all compositions are archived in the UVA Dataverse 85(https://doi.org/10.18130/V3/AD6H08). Declarations Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Supplementary Information The online version contains supplementary material available at 10.1038/s41598-026-57006-2. References Associated Data Supplementary Materials Supplementary Information. Data Availability Statement The relaxed crystal structure of all compositions are archived in the UVA Dataverse 85(https://doi.org/10.18130/V3/AD6H08).