A physics-regularized machine learning approach for predicting time-temperature-transformation curves in alloys: application to uranium-based alloys Authors: Sunidhi Garg, Jishnu Bhattacharyya, Vineet V Joshi, Sean R Agnew, Prasanna V Balachandran Published: 2026 ABSTRACT AbstractA physics-regularized machine learning (ML) approach is developed for predicting time-temperature-transformation (TTT) diagrams from alloy composition in uranium (U)-based systems. To ensure physically realistic C-curve shape while maintaining predictive accuracy, four loss function formulations are systematically evaluated: (1) baseline mean squared error (MSE), (2) MSE regularized with a semi-empirical nucleation-based model, (3) MSE regularized with shape constraints that enforce characteristic C-curve convexity, and (4) a hybrid approach that combines both semi-empirical regularization and shape constraints with MSE. Models that incorporate shape-constrained loss functions generate TTT predictions exhibiting a physically consistent C-curve form and demonstrate good agreement with experimental isothermal transformation data. Post hoc model explanation using global feature importance and partial dependence plots reveals systematic changes in the learned composition-TTT relationships as a function of the loss function formulation. The optimized ML model, which used a hybrid loss function, was applied to previously unexplored U-Mo-X ternary systems. This analysis identified U-Mo-Pt, U-Mo-Au, U-Mo-Ta, U-Mo-W, and U-Mo-Fe as candidate alloys that exhibit enhanced kinetic stability against γ-to-α phase decomposition during isothermal holds compared to the U-Mo binary baseline. Furthermore, TTT nose time predictions from the optimized ML model are compared with those from a Gaussian Process Regression (GPR) model using MSE loss. The developed physics-regularized ML framework offers potential for accelerating alloy discovery in U-based systems, where traditional TTT experimental determination is time and resource-intensive.Subject terms: Engineering, Materials science, Mathematics and computing, Physics Introduction Time-temperature-transformation (TTT) diagrams capture the kinetics of isothermal solid-state phase transformations through characteristic C-shaped curves. These diagrams show the time needed to achieve a specific volume fraction of transformation at various isothermal temperatures. The nose of the TTT diagram marks the temperature with the fastest transformation rate. TTT diagrams are valuable in physical metallurgy because they link processing parameters, such as temperature, time, and cooling rate, to microstructure evolution, which determines mechanical properties and alloy performance. However, experimentally generating TTT diagrams requires extensive isothermal heat treatment studies across multiple temperatures and hold times, making the process time-consuming and resource-intensive. Computational methods have been developed to predict time-temperature-transformation (TTT) diagrams prior to synthesis and characterization. Bhadeshia introduced a physics-inspired approach grounded in the Russell’s nucleation theory1, integrating fundamental thermodynamic principles, such as nucleation barrier, activation energy, driving force, with empirically fitted parameters to model TTT curve shapes2. Although this method effectively captures the characteristic C-curve topology, its applicability is restricted by the availability of experimental data and the necessity of extensive parametrization for each alloy system. Moreover, the validity of this technique has been extensively tested in ferrous alloys, such as steels2,3, and its application to non-ferrous alloys is limited or largely unexplored. Machine learning (ML) offers an alternative approach that addresses some of these limitations by learning composition-transformation relationships directly from experimental databases. In contrast to physics-based models, which require explicit parametrization of nucleation kinetics for each alloy family via model calibration, ML algorithms can model complex, nonlinear relationships between alloy composition and TTT curve characteristics without imposing restrictive functional forms. Malinov et al.4 developed an artificial neural network (ANN)-based ML model for predicting TTT diagrams for titanium alloys. Although the authors do not explicitly discuss their choice of loss function, their model description suggests the use of a method that minimizes the difference between experimentally measured and ANN-predicted targets during training. Kundu et al.5 developed an ANN model for predicting the TTT diagrams of steels. Rather than utilizing the entire TTT diagram, the authors divided the curve into upper and lower halves and selected two temperatures as outputs. Mean squared error (MSE) was employed as the loss function to train the ANN models. Recently, Huang et al.6–8 introduced a methodology that integrates multiple ML techniques to predict the TTT diagram of stainless steels, high-alloy steels, and low-alloy steels. The authors used three statistical metrics, namely mean absolute error (MAE), root mean squared error (RMSE), and correlation coefficient, to evaluate the prediction accuracy of the trained ML models. Hernández-Flores et al.9 developed an ANN-based approach to predict the TTT diagrams of Ni-Al alloys. The authors used RMSE as the loss function to train the ANN model. Although these studies demonstrate the potential of ML for predicting TTT diagrams, model training has relied solely on standard statistical loss functions, such as MSE, MAE, or RMSE. These functions minimize prediction error but fail to incorporate physical constraints on TTT curve morphology or the thermodynamic principles governing phase transformation kinetics. This study develops physics-regularized ML models to predict complete TTT curves for binary and ternary uranium (U) alloys by employing custom loss functions. The design of U-based nuclear fuels requires consideration of multiple factors, including phase stability, burnup characteristics, and mechanical properties, which vary depending on reactor design and application. Many research and test reactors are transitioning from highly enriched uranium (HEU) fuels to low-enriched uranium (LEU) fuels to mitigate nuclear proliferation risks while maintaining performance comparable to HEU fuels10. HEU fuels are those with a greater than 20% proportion of the naturally occurring fissile isotope, U-235, which is also the least abundant uranium isotope. One method to convert HEU to LEU fuel involves increasing the uranium density within the fuel. This is necessary because, as previously reported, each atom of U-235 in LEU must be accompanied by nearly four atoms of U-23811. Of the three allotropic forms of U, the body-centered cubic γ-phase (stable above 1440 K) exhibits favorable irradiation characteristics and thermal stability, which are critical during both processing and in-reactor use12. In contrast, the orthorhombic α-phase (stable up to 939 K) and the tetragonal β-phase (stable between 939 K and 1440 K) are unsuitable as fuel phases due to inadequate irradiation properties. Among U-based alloys, the U-molybdenum (U-Mo) binary alloy system is of particular interest because it stabilizes the desired metastable γ-phase13,14. This metastable γ-phase can undergo eutectoid decomposition to orthorhombic α-U and intermetallic -phases (UMo) upon heat treatment. Studies have shown that introducing small amounts of a ternary element to this alloy (U-Mo-X, where X is the ternary element) can further stabilize the γ-phase and extend the nucleation time for the α-phase14. Certain 4d and 5d transition metals form solid solutions with the γ-U phase, and narrowing the selection of metals and their compositions can expedite the development of optimal LEU fuels15. Predictive modeling of TTT diagrams in the U-Mo-X composition space has the promise to enable systematic identification of alloying additions that extend the incubation time of the γ- to α-phase transformation and slow eutectoid decomposition, thereby minimizing the need for exhaustive experimental characterization of each candidate ternary system. Johns and Burkes investigated the effect of incorporating a ternary element into U-Mo alloys by predicting the nose of the TTT curves using ANN16. Their models used the weight percentages of U, Mo, and the ternary element, along with the isothermal temperature from TTT curves, as input variables. The only output was the time corresponding to the nose of the TTT curve. Forty-two experimentally determined TTT curves from various literature sources were used to develop the ANN models. Consistent with previous studies, MSE was used as the loss function for model training. Although this work demonstrated the feasibility of ML for U-based alloy systems, it was limited to predicting a single point (the nose) rather than the entire TTT curve and did not include mechanisms to enforce physically realistic TTT curve topology. The present study addresses these limitations through three methodological advancements: (1) physics-regularized loss functions that integrate Bhadeshia’s semi-empirical model2 and geometric constraints such as C-curve convexity to ensure physically plausible predictions, a capability not present in previous ML-based TTT models; (2) complete TTT curve prediction for compositionally unexplored U-Mo-X ternary systems that await experimental validation, enabling computational screening of candidate alloying additions; and (3) quantitative uncertainty estimates for all predictions, addressing a limitation explicitly acknowledged by Johns and Burkes in their ANN approach. Predictions are compared with a Gaussian Process Regression (GPR) model that predicts the TTT nose time using MSE loss, following the approach of Johns and Burkes16. This model serves as a baseline for evaluating the physics-regularized TTT predictions. Methods The computational workflow used to build a quantitative relationship between the alloy composition and their respective TTT curves is given in Fig. 1. The details of this workflow are explained below. A schematic illustration of the computational workflow developed in this study. The approach begins with the collection of a dataset comprising experimentally determined TTT curves for U-based alloys. To establish a relationship between alloy composition and TTT curves, temperatures from 40 TTT curves and 10 composition-weighted elemental descriptors are used as input data. Feature engineering is conducted using principal component analysis (PCA) on these composition-weighted elemental descriptors to remove redundant or strongly statistically correlated descriptors. The first three principal components or PCs (PC1, PC2, PC3), the temperature from the TTT curves, and the number of components in the alloy () are selected as final descriptors and divided into training and independent test sets for the ensemble support vector regression (eSVR) ML model. Hyperparameters are optimized by minimizing the loss function to identify the optimal eSVR model. Four types of loss functions are evaluated: mean squared error (), +, ++, and +, to generate TTT curve predictions. The two main regularization loss function terms are as follows: (1) The term integrates time predictions from the calibrated analytical Bhadeshia model2 through Bayesian inference, and (2) the term enforces the characteristic convexity of the C-curve. The optimized eSVR models are subsequently employed to predict TTT curves for U-Mo-based ternary alloys not included in the training and independent test sets. Finally, post hoc model explanation analysis techniques are applied to interpret the learning patterns of these physics-regularized eSVR models. Analytical TTT Bhadeshia model H. K. D. H. Bhadeshia developed a closed-form mathematical model by adapting the nucleation framework of Russell to predict the C-shaped TTT curves for the initiation of phase transformations in various steel compositions 1–3. This model defines the relationship between time (τ ) and temperature (T) as shown in Eq. (1). Here, R denotes the universal gas constant, represents the maximum free energy change during nucleation, Q is the activation enthalpy for diffusion, and C, p, and z are empirical fitting constants. The variable (τ ) refers to the incubation period, defined as the time required for the onset of phase transformation. The model can be further simplified as follows. The parameters in Eq. (2), specifically Q, A, and z, are treated as fitting parameters. This formulation is applied to predict TTT curves for the initiation of transformation from the desired γ-phase in U-Mo-X alloys to other thermodynamically stable phases, such as the α-phase or γ '-phase (UMo). In this context, , where is the temperature at which the incubation time is minimized, commonly referred to as the nose of the TTT curve. This relationship for Q is obtained by differentiating Equation (1) with respect to temperature (T) and setting to zero. The equation yields a single solution, , which defines the nose. The value of z is fixed and selected based on Bayesian inference, as discussed below. The constant A is given by . Bayesian inference The Bayesian inference approach applies Bayes’ theorem17,18 to estimate parameter distributions in the analytical TTT model from experimental data. Bayes’ theorem is defined as follows: In this context, θ denotes the parameters of the analytical TTT model, and X refers to the collected data. P(θ ) represents prior knowledge of the model parameters as a probability distribution before observing the data (X). P(X|θ ) is the likelihood function, which expresses the joint probability of the observed data as a function of the model parameters. P(X) serves as a normalizing constant. P(θ |X) denotes the posterior probability distribution of the model parameters after incorporating the observed data (X). This method is employed to calculate the posterior distributions for the parameters of the simplified analytical model described in Eq. (2). The model is formulated as a single-parameter (A) model and evaluated across a range of z values, incorporating noise in the form of , where denotes normal distribution. The parameter A is assigned a uniform prior distribution, P(θ ). For each curve, the minimum value for the uniform prior of A is given by , and the maximum is calculated as . In these expressions, and denote the temperatures at minimum () and maximum () times from the experimentally measured TTT curves, respectively. The adaptive Metropolis-Hastings Markov Chain Monte Carlo (MCMC) algorithm is utilized to compute the posterior distributions for A. This approach is applied to multiple TTT curve datasets taken from the literature16 to estimate the constant A for each curve. The distributions corresponding to z=54 are selected as the final result based on the comparison between predicted and experimental TTT curve data. The predicted A values are subsequently used as input for Eq. (2) to generate the predicted time values from the analytical model. These predicted time values (Bayesian TTT curves) are later used to optimize ML model performance, as discussed in subsequent subsections. All Bayesian calculations are performed using the open-source BayesianTools19 package implemented in the R language. Dataset curation The TTT dataset for multiple U-Mo-X alloys was curated from the work of Johns and Burkes16, who compiled 42 U-Mo-X-based TTT curves from four primary sources, each employing different methods to measure these curves20–23. In this context, X represents chromium (Cr), niobium (Nb), rhenium (Re), ruthenium (Ru), or zirconium (Zr), with varying weight percentages. The U-10Mo and U-12Mo curves measured by Peterson et al. were excluded from further analysis due to previously reported discrepancies in the behavior of U-12Mo compared to U-10Mo24. Among the remaining 40 curves, the U-8Mo TTT curve was measured by three separate groups, resulting in 38 distinct alloys included in this study. The U-10Mo TTT curve measured by Repas et al. was retained, as recent literature demonstrates agreement between their reported nose time and temperature and those reported in other studies25. The complete list of these 38 alloys is provided in Table S1 in the Supplementary Information (SI). Some discrepancies are anticipated in the reported TTT curves, as different experimental techniques and conditions were employed to measure the TTT curves and synthesize the U-Mo-X alloys. The temperature range for all TTT curves is 440 C to 560 C. To establish relationships between alloy composition and TTT curves, ten elemental properties for each element (U, Mo, Cr, Nb, Re, Ru, Zr) were used as descriptors: atomic radius, number of valence electrons, electronegativity, melting point, thermal conductivity, bulk modulus, density, polarizability, heat of fusion, and the ratio of volume to number of atoms in a unit cell (V/Z)26,27. Each elemental descriptor was multiplied by its respective weight percentage in the alloy, excluding uranium, and averaged over the total composition. An additional descriptor, the number of alloy elements (), was included to indicate whether the alloy is binary (two elements, U-Mo) or ternary (three elements, U-Mo-X). For each of the 40 TTT curves, a vector of 11 input descriptors was constructed to represent the alloy composition and the corresponding time-temperature data for the TTT curve. Feature extraction A curated dataset comprising 38 distinct alloys and 11 input descriptors was analyzed using the Pearson correlation coefficient (PCC)28. To remove redundancy, principal component analysis (PCA), a linear dimensionality reduction technique, was applied to the elemental descriptors29–31. The resulting principal components (PCs) are linear combinations of the elemental descriptors, each with distinct weights (loadings). PC1 accounts for the largest variance in the dataset, with subsequent PCs capturing progressively less variance. It is important to note that the ten elemental descriptors remain constant across all temperature-time pairs within a given TTT curve because composition does not change with temperature. Consequently, temperature is included as an additional independent input descriptor, along with PC1, PC2, PC3, and , to capture temperature-dependent transformation kinetics within each TTT curve. Future research could benefit from incorporating thermodynamic descriptors that depend on both composition and temperature, such as the Gibbs energy difference between the α- and γ-phases. Machine learning with custom loss function A quantitative relationship between alloy composition and the corresponding TTT curve is established using ML models. A bagging ensemble learning approach is implemented32 with two ML models: (1) least absolute shrinkage and selection operator (LASSO)33 and (2) support vector regression (SVR)34,35, referred to as eLASSO and eSVR, respectively. The bootstrap resampling method generates multiple replicas of the input data through sampling with replacement36. The non-linear eSVR model with a radial basis function kernel demonstrates superior performance compared to the linear eLASSO model and is therefore selected for subsequent analysis. The eLASSO performance plot for the training and test datasets is provided in Fig. S3 in the SI. The final prediction from the eSVR method is a probability distribution characterized by the ensemble mean and standard deviation. The ensemble comprises 25 bootstrap samples. Bayesian optimization is employed to optimize the hyperparameters, (C, ε ) for the SVR method and (σ ) for the RBF kernel37, by minimizing the ten-fold cross-validation error. The final predictions from the eSVR approach yield a probability distribution for time, given the three principal components (PCs), , and the temperature from the TTT curves. The eSVR code utilizes mean squared error () as the default loss function to assess model performance during hyperparameter optimization. For this study, the loss function is modified as described below to ensure conformity with the characteristic C-curve topology of the TTT curve. The code for the eSVR model is implemented in R. The generalization capability of the trained eSVR models was subsequently assessed using an independent test set and predictions on previously unseen compositions (see “ML model training and performance” and “Testing ML prediction of TTT curves on known alloys”). Loss function In ML models, loss functions quantify the difference between predicted and experimentally obtained time values during training and guide the iterative optimization of model parameters. A custom loss function approach is used in this work, where this function is modified. Standard statistical loss functions, such as MSE, RMSE, and MAE, prioritize prediction accuracy but do not account for physical plausibility. To address this limitation, a custom loss function is developed that integrates physics-based regularization terms with prediction error. This approach ensures that predicted TTT curves display physically realistic features, including C-curve morphology and alignment with nucleation theory. This concept is inspired by physics-informed ML approaches, which incorporate domain knowledge from governing physical laws into the learning process using custom loss functions, network architectures, or training constraints38. The loss function for training SVR models comprises three terms: (1) mean squared error , which quantifies prediction accuracy on experimental data; (2) a semi-empirical model regularization term that penalizes deviations from Bhadeshia nucleation-based predictions; and (3) a shape constraint term that enforces the characteristic C-curve topology. The custom loss function is mathematically expressed as follows: where , , and denote the weights assigned to the , , and loss terms, respectively. In this study, all λ values are set to one. Each component of this custom loss function is described in detail in the following section. Mean squared error The mean squared error term penalizes large differences between the predicted and experimental values. It is mathematically defined as, Here, n is the total number of data points, and and are the true (experimental) and predicted time values for the data point from the SVR models. Semi-empirical model regularization The semi-empirical model regularization loss term incorporates the predicted time values from the SVR model and the calculated time values obtained from Eq. (2). The analytical time values are derived from the posterior distribution of parameter A under the Bayesian inference approach. This loss term is mathematically expressed as follows: In this context, n denotes the total number of data points. Both and correspond to values from the training set. Curvature loss The curvature loss term () penalizes the model when the predicted TTT curves deviate from the C-shape. The loss function term incorporates a calibration set to monitor the curvature of the predicted TTT curve during model training. The U-10Mo alloy was selected as the calibration alloy due to its proposed use as a metal fuel composition with sufficient U-235 density to replace HEU in nuclear research and test reactors39. Therefore, accurately predicting its phase transformation kinetics is essential for fuel fabrication and performance. The SVR model, for a given set of hyperparameters, predicts time values for the training and calibration sets (U-10Mo TTT curve). The calibration set used for this work comprises PC1, PC2, PC3, , and experimentally collected time and temperature data for the U-10Mo alloy, consisting of 30 experimentally measured time-temperature pairs. Training set predictions are used to compute and , while calibration set predictions are used to compute . The total loss is the weighted sum of all three terms. The curvature penalty is computed through a two-step process applied to the calibration set predictions: Step 1: Detecting invalid curve topology. The predicted curve is assessed for violations of C-curve shape by examining sign changes in the slope (). A valid C-curve is characterized by exactly one local minimum (the nose) and the absence of local maxima. Specifically, multiple sign changes from negative to positive slope indicate the presence of multiple local optima, while any sign change from positive to negative slope indicates a local maximum. If either condition is detected, a hard penalty is assigned, effectively rejecting the hyperparameter configuration. Step 2: Penalizing deviations from C-curve convexity. If the curve demonstrates valid shape, characterized by a single minimum at the nose, quantifies deviations from the ideal C-curve shape by summing slope contributions before and after the nose: Each term is calculated as follows: In this context, k denotes the number of calibration points before or after the nose, and represent the experimental nose time and temperature, respectively. is the ML predicted time and correspond to the temperature for the calibration point. This step penalizes predicted times that fall on the incorrect side of the nose. For temperatures above the nose (), predictions are expected to exceed (negative slope), whereas for temperatures below the nose (), predictions should also exceed (positive slope). To enforce this criterion, if the values are negative or the are positive then, and are set to zero, as these indicate conformity with the expected C-curve behavior. Post hoc model explanation This study employs two types of model-agnostic post hoc explanation techniques to provide insights into the learning patterns of the “black box” eSVR models: (i) global feature importance (GFI) and (ii) partial dependence plots (PDPs). Global feature importance (GFI) The goal of GFI is to calculate the importance of individual descriptors relative to others in predicting the output of the trained ML models. GFI is quantified by randomly permuting the values of a specific descriptor across all samples in the dataset, while keeping the values of other descriptors unchanged. This process intentionally breaks the association between the selected descriptor and the target variable40. The trained ML model is then applied to the permuted dataset, and the resulting increase in prediction error indicates the importance of the descriptor. In general, descriptors critical to accurate prediction exhibit large performance drops when permuted, while uninformative descriptors show minimal change. In this study, GFI values are calculated for each input descriptor (PC1, PC2, PC3, , and temperature) to identify which descriptor most strongly influences TTT curve predictions. Mathematically, the GFI [V(f, i)] is defined as the difference between the loss function values of the ML model evaluated on the original dataset and on the permuted dataset, as follows: where and R(f) denote the loss function values for permuted and original data, respectively. The error bars for the GFI values of each descriptor are derived from the standard deviation of multiple bootstrap resamples employed in the ensemble ML approach. Partial dependence plots (PDP) Partial dependence plots (PDPs) characterize the marginal effect of individual descriptors on model predictions by averaging predictions over the distribution of all other descriptors41. To construct PDPs, each of the five input descriptors (PC1, PC2, PC3, , and temperature) is selected and varied across a grid of 50 evenly spaced values spanning their observed ranges in the training data. For each grid value, a simulated dataset is created by setting the target feature to that value for all training samples, while keeping the other features at their original values. The ensemble of trained SVR models then generates predictions for each simulated dataset. The PDP at each grid point represents the mean prediction averaged first across all training samples (marginalizing over other features) and then across all ensemble members. This procedure reveals how the average predicted transformation time varies with each descriptor, independent of the effects of the other descriptors42. Gaussian process regression (GPR) The GPR model uses Bayesian inference to learn the distribution of functions, f, that is most likely to have generated the observed data43. This involves computing the posterior distribution of f given the data, using Bayes’ theorem17. The GPR model assumes that the function f is chosen from a Gaussian process i.e. the prior distribution of f is multivariate Gaussian with mean function and covariance function . The covariance function measures the correlation between two points in the input space , and is defined using kernel functions. Once the posterior distribution of f has been learned, the model can make predictions at new test points by computing the posterior predictive distribution. The MATLAB function fitrgp with the Automatic Relevance Determination (ARD) squared exponential kernel function, ardsquaredexponential44, with a separate length scale for each predictor, , and signal standard deviation was employed: . The default constant basis function was employed and the kernel parameters were estimated by maximizing the log-marginal likelihood (maximum likelihood estimation) using quasi-Newton gradient-based optimization method. Results and discussion Relationship between U-Mo-X alloys and principal components (PCs) from principal component analysis (PCA) The Pearson correlation coefficient (PCC) analysis of the 11 descriptors revealed strong correlations among the 10 elemental descriptors, except for . All 10 elemental descriptors exhibited absolute pairwise PCC values greater than 0.8, indicating a high degree of correlation. The pairwise PCC correlation plot for these descriptors is provided in Fig. S1 in the SI. To remove redundant or strongly statistically correlated descriptors, PCA was applied. The 10 elemental descriptors were transformed into principal components (PCs), and three PCs (PC1, PC2, PC3) were selected as input descriptors for training the ML models. These PC vectors are linear combinations of the elemental descriptors, each weighted by specific loadings. Three PCs with negligible pairwise PCCs collectively capture 99.9% of the dataset’s variance. The loadings for PC1, PC2, and PC3 are given in Table S2 in the SI. The scree plot describing the cumulative variance captured by the PCs is given in Fig. S2 in the SI. PC1 accounts for the largest proportion of variance (94.8%), with similar absolute loading values across all elemental descriptors. PC2 explains 3.6% of the variance and is primarily influenced by thermal conductivity, polarizability, and the ratio of unit-cell volume to the number of atoms in a unit cell (V/Z), each with absolute loadings greater than 0.4. PC3 captures 1.5% of the variance and is dominated by thermal conductivity and density. These three PCs, along with and temperature, serve as the five input features for the ML models discussed below, with time as the output variable. The final dataset contains 906 data points derived from 40 total experimentally determined TTT curves representing 38 alloys. K-means clustering using the silhouette method was performed on this dataset, followed by an 80:20 split of the clusters to generate training and test sets for the ML models comprising 724 and 182 points, respectively. Prior to ML model training, all input descriptors are min-max normalized between zero and one. In Fig. 2, the projection of the original dataset in the transformed PC space is shown. While no distinct clustering is evident in these plots (Fig. 2), certain trends related to PC values and the U-Mo-X alloys are discerned. In Fig. 2a, a linear trend is observed for the binary U-Mo alloy, where decreasing Mo content corresponds to increasing PC1 and PC2 values, as indicated by blue points. Literature indicates that increasing Mo content extends the α-U phase nucleation time. Segregation of alloys based on composition is more pronounced along the PC2 axis than along the PC1. Most U-Mo-X alloys with X = Zr or Nb (black and red points, respectively) are located in the PC2 > 0 range, whereas alloys with X = Re or Ru (magenta and orange points, respectively) are found in the lower PC2 range, as shown in Fig. 2a,c. For PC3 in Fig. 2b,c, alloys with X = Re (magenta points) occupy higher PC3 values compared to other ternary and binary alloys. Thus, PCA transforms ten correlated elemental descriptors into three PCs, retaining 99.9% of the variance and revealing composition-based trend in U-Mo-X alloys, particularly along PC2 and PC3. The three PCs are utilized to represent the compositions of U-Mo-X alloys for ML model development. Scores plots are presented in which each PC (PC1, PC2, PC3) is plotted against the others: (a) PC1 versus PC2, (b) PC1 versus PC3, and (c) PC2 versus PC3. PC1 accounts for the largest proportion of total variance in the dataset (94.8%), followed by PC2 (3.6%) and PC3 (1.5%). A linear trend is observed in binary U-Mo alloys (blue points), with decreasing Mo content as PC values increase. No clear segregation based on alloying elements is observed along the PC1 axis. Along the PC2 axis, Zr- and Nb-alloyed compositions, represented by black and red points, respectively, tend to cluster near higher PC2 values in (a) and (c). Re-based alloys, represented by magenta points, also group near higher PC3 values in (b) and (c). ML model training and performance The input descriptors are PC1, PC2, PC3, and temperature. The output or target variable is time. The goal of eSVR is to establish a quantitative relationship between the inputs and the output, along with the associated uncertainties. Four independent eSVR models are trained using four variations of the loss function: (1) , (2) , (3) , and (4) . The first two configurations ( and ) utilize the training and test sets, while the third () and fourth () configurations incorporate the training, test, and calibration sets. The values for these four models are presented in Table 1. Table 1The values for training and test data were evaluated across four eSVR models utilizing the following loss functions: , +, +, and +. For both training and test datasets, the value decreased with the addition of loss function terms beyond MSE. The reduction was most pronounced when the curvature loss function term () was included.Loss functionTrain Test 0.9860.9530.9740.9580.8870.8860.8840.881 The values for training and test data were evaluated across four eSVR models utilizing the following loss functions: , +, +, and +. For both training and test datasets, the value decreased with the addition of loss function terms beyond MSE. The reduction was most pronounced when the curvature loss function term () was included. The four loss functions evaluated in this study affect model predictions in distinct ways. Among these, the loss produces the highest value, indicating the best overall performance, as shown in Table 1. The loss function yields the next highest value, while the addition of leads to a noticeable reduction in . This trend is also evident in the parity plots in Fig. 3, where predicted time values increasingly deviate from the y=x blue dashed line for both training (black circles) and test (red triangles) sets. Figure 3a,b, corresponding to the default and loss functions, respectively, demonstrate high values, with training and test set points closely aligned with the blue dashed line, except for some deviation at higher values in Figure 3(b). In contrast, Fig. 3c,d, where the loss term is included, show a more pronounced decrease in , as indicated by the greater deviation of predicted data points from the blue dashed line. The uncertainties associated with the predictions, as measured by the standard deviations from the ensemble learning approach, also increase with the addition of loss function terms beyond . Overall, model performance, as reflected by values, declines with the inclusion of additional loss function terms, with a significant decrease observed upon adding the curvature loss term (). The reduction in prediction accuracy constitutes a deliberate trade-off. By sacrificing fit to individual data points, physics-regularized models enforce a characteristic C-curve topology, thereby enhancing physical plausibility and improving generalization to previously unexplored compositions. Future work should explore the use of multiple calibration alloys across a broader range of transformation times. This approach has the potential to mitigate the persistent underestimation observed in alloys characterized by longer TTT times. The eSVR predicted time values versus the experimental time values for the four different loss functions (a) , (b) +, (c) + and (d) +. The black filled points represent the training data and the red triangular points represent the independent test data. The points lying on the blue dashed line represent complete agreement between the eSVR predicted and experimental time. The eSVR predicted points tend to deviate from this blue line as the loss function terms are added from (a) to (d). eSVR model with only as the loss function shows the best performance on both training and independent test data in (a). The uncertainties associated with all the predictions are calculated from the ensemble approach used in this work. Ensemble XGBoost (eXGBoost) models were also trained using the and loss functions45,46. Despite achieving train and test values greater than 0.9, these models did not preserve the expected continuity of the characteristic C-curve shape. The predicted time values exhibited abrupt deviations and flattening in both higher and lower temperature regions. Consequently, this method was excluded from further analysis. Post hoc model explanation After analyzing model performance and predictions, the trends learned by the four trained eSVR models are examined to assess the impact of loss function modifications. Two model-agnostic post hoc explanation techniques are employed: GFI and PDP. GFI analysis indicates that all five input features (PC1, PC2, PC3, , and temperature) are important across all loss function variants, although their relative rankings vary with different regularization terms (detailed results are shown in Fig. S4 in the SI). Because GFI does not reveal clear trends distinguishing the loss function formulations, the focus shifts to PDPs, which offer more interpretable insights into how each feature influences predicted transformation kinetics and how these relationships change under physics-based constraints. In Fig. 4, the PDPs for predicted time (hr) versus normalized temperature (C) across all four eSVR models are shown. The overall trend exhibits a U-shaped curve, consistent with expectations for a TTT curve. In Fig. 4a,b, the curves display a U-shape within the intermediate temperature range but deviate at both lower and higher temperature values. Incorporating the term into the loss function corrects this deviation, as demonstrated in Fig. 4c,d. PDP for predicted time versus normalized temperature (C) is shown for four eSVR models with different loss function terms: (a) , (b) +, (c) +, and (d) +. The curves generally exhibit the expected C-shaped trend characteristic of TTT curves. However, deviations occur at low and high temperature ranges in (a) and (b), where the loss function term is not included. Incorporating in (c) and (d) corrects these deviations. Prediction uncertainties remain elevated, particularly at the extremes of the temperature range. PDPs for normalized PC1, which captures 94.8% of the variance, versus predicted time are shown in Fig. 5. The absolute loading values across all elemental descriptors are similar for PC1 (see Table S2 in SI). This result indicates that PC1 captures variance related to correlated changes across all elemental properties. Consequently, PC1 does not distinguish the contributions of individual elemental properties. The PDPs shown in Fig. 5a,b reveal that models without the loss term are characterized by a dominant maximum at PC1≈0.5. In contrast, incorporating the loss term alters how the model interprets the role of PC1. The PDPs shown in Fig. 5c,d exhibit a peak at PC1 ≈ 0.35, followed by a valley near 0.75, and then an increase at higher PC1 values. PDP for predicted time versus normalized PC1 (capturing 94.8% variance) are shown for four eSVR models, each employing different loss function terms: (a) , (b) +, (c) +, and (d) +. In models lacking the loss function term, the curves exhibit a bell-shaped profile with substantial uncertainties in the intermediate PC1 region. The addition of the term in (c) and (d) alters the overall trend throughout the normalized PC1 range. Similarly, the PDPs for predicted time versus normalized PC2, which accounts for 3.6% of the variance, are shown in Fig. 6. The curvature in the lower PC2 value range changes when the term is added to the loss function, as shown in Fig. 6c,d. PC2 is primarily influenced by thermal conductivity, polarizability, and the ratio of unit-cell volume to the number of atoms in a unit cell (V/Z), each exhibiting absolute loadings greater than 0.4. Negative loadings for thermal conductivity and polarizability indicate that alloying elements with high values in these properties reduce the overall contribution to PC2. This reduction is favorable, as the PDP plots in Fig. 6c,d demonstrate a delayed nucleation time for the α-phase, thereby enhancing γ-phase stability. Conversely, the V/Z descriptor shows a positive loading, indicating that higher V/Z values contribute more to PC2. Therefore, the competition between thermal conductivity, polarizability and V/Z determines the value of PC2. The PDP plots in Fig. 6c,d further show that elevated PC2 values decrease the nucleation time of the α-phase, consequently reducing γ-phase stability. Further, as seen in Fig. 2a,c, the Mo content decreases in U-Mo alloy series as the PC2 values increases indicating decrease in predicted time and destabilization of γ phase with decreasing Mo content. Additionally, Zr alloyed U-Mo-X alloys also lie in the elevated PC2 range as shown by black points in Fig. 2a,c, indicating lower γ-phase stability as reported in the experimental literature21,47. PDP for predicted time versus normalized PC2 (capturing 3.6% variance) are shown for four eSVR models with different loss function terms: (a) , (b) +, (c) +, and (d) +. Predicted time generally decreases as PC2 increases, with a slight increase observed in the low temperature region for (a) and (b), where the loss function term is absent. This increase is mitigated in (c) and (d), where the term is included. Uncertainties are lower in the high temperature region than in the low temperature region across all plots. Additional PDPs showing predicted time against normalized PC3 and predicted time against normalized are provided in Figs. S5 and S6, respectively, in the SI. PC3 captures 1.5% of the variance and is primarily influenced by thermal conductivity and density with loading values being greater than 0.4 for each. Predicted time has a linearly decreasing trend with increasing PC3 as shown in Fig. S5 in the SI. Negative loadings for thermal conductivity and positive loadings for density indicate that they are competing descriptors in determining the final PC3 value. A higher thermal conductivity and lower density will reduce the PC3 value. All the four models show a similar trend for the PDPs describing the variation of i.e. the predicted time increases as a ternary element is alloyed. These PDPs are provided in the SI because their marginal effects on predicted transformation time are smaller than those of PC1 and PC2. The key takeaway from the PDP analysis is that enforcing C-curve shape constraints through the term fundamentally alters the learned marginal effects of composition (as represented by PC1, PC2, and PC3) and temperature on transformation kinetics. Testing ML prediction of TTT curves on known alloys Following the analysis of loss function modifications in the PDP, the next step is to evaluate which loss function formulation produces the most physically realistic TTT predictions for compositions outside the training set. Four eSVR models generate complete TTT curves for three U-Mo-X ternary alloys: U-9.8Mo-0.2Ni, U-9.8Mo-0.2Cr, and U-9.8Mo-0.2Co. These compositions are excluded from both training and test datasets but have been experimentally characterized in the literature25,48, enabling direct validation of model performance and assessment of the effectiveness of different loss function terms. Peterson et al.25 performed ex-situ neutron diffraction measurements on samples heat-treated at three temperatures (450 C, 500 C, and 525 C) for two durations (20 and 100 h), yielding discrete phase fraction data instead of continuous TTT curves. These results demonstrate that Ni and Co additions accelerate γ-phase decomposition at all tested temperatures, whereas Cr shows transformation kinetics similar to or slightly slower than those of binary U-10Mo. Therefore, model validation assesses whether predicted TTT curves exhibit physically realistic C-curve topology and whether the relative transformation kinetics among the alloys are consistent with these experimental findings. The predicted TTT curves for the three U-9.8Mo-0.2X alloys (where X = Ni, Cr, and Co), along with the U-10Mo baseline, are shown in Fig. 7. The uncertainties in the time predictions arise from the ensemble learning approach used in this study. Loss functions that exclude the C-curve shape, such as in Fig. 7a and in Fig. 7b, fail to reproduce the characteristic C-shape of the TTT curves. Conversely, Fig. 7c,d, which incorporate the C-curve shape constraint through the term, largely preserve the characteristic shape of the TTT-curves, despite some local irregularities. This comparison indicates that the term is crucial for producing physically plausible TTT predictions on previously unseen compositions. Based on these results, all subsequent analyses use the eSVR model trained with the full loss function ( + + ). Although produces only modest differences compared to + alone, its inclusion ensures that predictions remain consistent with semi-empirical nucleation theory and enforce geometric constraints. The eSVR model trained using the hybrid ++ loss function effectively capture the relative ordering in U-9.8Mo-0.2X system, where X = Ni, Co, and Cr (Fig. 7c). The predicted curves exhibit close similarity, with the primary difference observed in the 500–520 C range. These results indicate that small ternary additions (0.2 wt%) induce modest shifts in TTT behavior rather than substantial changes. Error bars increase at lower temperatures because the training dataset has fewer experimental measurements in this temperature range. The predicted TTT curves for U-9.8Mo-0.2X (where X= Ni, Cr, Co) from the eSVR models employing (a) , (b) +, (c) +, and (d) + loss functions. Each plot also includes predicted TTT curve for U-10Mo alloy as the baseline for comparison. The predicted TTT curves for U-10Mo and U-8Mo exhibit inflection points near both the lower and upper bounds of the temperature range (440–560C). This trend can be attributed to two primary factors. (1) Limited data coverage at these temperature extremes in the training dataset reduces model accuracy outside the nose region. (2) The curvature loss term is applied using only a single calibration composition (U-10Mo), which may not sufficiently constrain the C-curve shape for alloys with markedly different transformation kinetics. The U-Mo-Ti ternary alloy system has been experimentally investigated and compared with the baseline U-10Mo alloy. Experimental results demonstrate that the U-8Mo-Ti alloy reduces the γ-phase stability relative to U-10Mo binary alloys22. The predicted TTT curve for U-8Mo-Ti, obtained using the loss function, is shown in Fig. S8b in the SI. Consistent with experimental observations, the predicted TTT curve shifts leftward, indicating that Ti addition decreases γ-phase stability compared to U-10Mo (see Fig. S7 in the SI) and U-9Mo (see Fig. S8b in the SI) binary alloys. Furthermore, Fig. S8b in the SI shows good agreement between predicted and experimentally measured TTT curves at higher temperatures, while discrepancies increase at lower temperatures. The error bars for the predicted TTT curves are also larger at lower temperatures than at higher temperatures. Due to the inherent experimental variability in TTT measurements across different research groups, the primary value of the current physics-regularized eSVR framework lies in its ability to accurately capture compositional trends in γ-phase stability within the U-Mo-X alloy system, rather than achieving precise quantitative agreement with individual experimental measurements. Figure 8 shows a direct comparison of nose temperature versus nose time for six distinct alloy compositions: U-9.8Mo-0.2Ni, U-9.8Mo-0.2Co, U-9.8Mo-0.2Cr, U-8Mo-Ti, U-9Mo, and U-10Mo. The predicted nose times for U-9.8Mo-0.2Ni, U-9.8Mo-0.2Co, and U-9.8Mo-0.2Cr (orange, magenta, and blue squares, respectively) are nearly identical. These three alloys are also predicted to exhibit higher nose temperatures compared to the U-10Mo baseline. In contrast, the predicted nose time for U-8Mo-Ti (green squares) is substantially lower than those predicted for U-9Mo and U-10Mo, as well as the experimental values. The experimental nose time for U-8Mo-Ti is also displayed in Fig. 8. There is excellent qualitative agreement between the experimental and predicted data. Additionally, the eSVR model trained with the loss function consistently overestimates the nose temperature of these alloy systems relative to the available experimental data. The predicted nose time and corresponding nose temperature for four U-Mo-X (X = Ni, Cr, Co and Ti) alloys, U-10Mo and U-9Mo alloys from the model employing loss function. Minor addition (0.2 wt%) of ternary element (Ni, Co and Cr represented by orange, magenta and blue squares respectively) to U-10Mo alloys stabilizes the γ-phase with higher corresponding nose temperature compared to baseline U-10Mo predicted (black circle) and experimental values (black star). Ni and Co predictions nearly overlap each other with Cr nose time being slightly lower and temperature being higher than Ni and Co alloyed alloys. The predicted and experimental nose time and temperature for U-10Mo are also comparable. The predicted (green square) and experimental (green star) nose time for U-8Mo-Ti alloys is shifted to the left indicating a lower γ-phase stability compared to the baseline U-10Mo and U-9Mo alloys aligning with the experimental results. The predicted and experimental nose time values for U-8Mo-Ti are comparable with the corresponding nose temperature being higher than the experimentally reported value. Predictions on previously unexplored U-Mo-X alloys The trained eSVR model with hybrid loss function is employed to predict the TTT curves for 24 U-Mo-X alloys not previously reported in the literature. The investigated alloys have general compositions U-9Mo-X and U-8Mo-2X, where X represents the ternary element. Out of these 24 alloys, 12 of these compositions have X = Ta, Pt, Hf, Ge, Ga and Au and were selected based on their solubility in the γ-phase of uranium, as derived from equilibrium phase diagrams49. Predictions for these 12 alloy compositions, are compared with those for baseline alloys U-8Mo, U-9Mo, and U-10Mo. A representative set of ternary elements (X) in the remaining 12 U-Mo-X alloys was chosen due to their frequent presence as elemental impurities in uranium ore concentrates50. These alloys also have the same general compositions U-9Mo-X and U-8Mo-2X, where X = W, V, Fe, Mg, Cu, and Al. Among these six ternary systems, V-, Fe- and Al-alloyed U-Mo-X compositions have been studied in the literature previously but none of these studies report the TTT curves for the alloys.47,51,52. Storhok et al.47 discussed the effect of ternary and quaternary alloying of Cr, Mo, Nb, Ru, V, and Zr on γ-phase stability of U-Zr, U-Nb, and U-Mo-based alloys and their effect on properties such as the transformation kinetics, transformation temperature, hot hardness, and corrosion resistance; however, no TTT curves were reported. More specifically, addition of V was reported to be detrimental to the γ-phase stability of U-Mo base alloys. Nicholson et al.51 investigated the influence of Fe, Al, Ru, Cr, and Nb alloying on the microstructure of dilute U-Mo-based alloys and estimated the phase ratio for these ternary alloys. Although no TTT curves were provided, the phase ratio was observed to decrease in Fe- and Al-containing U-3.6 (at%)Mo alloys. Slattery and Miller52 measured and compared the creep strength of uranium alloys alloyed with Mo, V, Zr, Cr, Al, and Ti. They evaluated properties including creep moduli, maximum strain rate, and deflection rate for these alloys; however, no TTT curves were reported. Comparisons between alloys containing X = Ta, Pt, Hf, and Au and the baseline alloys are shown in Fig. 9a,b. Figure 9a shows the predicted U-8Mo-2X (X = Ta, Pt, Hf, and Au) alongside U-8Mo (black triangles) and U-10Mo (red triangles) TTT diagrams. The predicted nose temperature from the TTT curves for all four U-8Mo-2X alloys exceed that of U-10Mo, consistent with the previously identified systematic overestimation in the trained eSVR model. The predicted nose TTT time increases for X = Pt (blue circles) and Au (magenta circles), indicating enhanced γ-phase stability, whereas X = Ta (green circles) shows only a marginal increase relative to U-10Mo. The U-8Mo-2Au TTT curve lacks a well-defined C-shaped profile and exhibits greater uncertainties than the Pt- and Ta-alloyed compositions. For U-9Mo-X alloy compositions shown in Fig. 9b, the predicted nose temperature from the TTT curves for all alloys exceeds that of the baseline U-10Mo curve again. The C-curvature is more pronounced for the U-9Mo-Au curve, although it is associated with high uncertainties. The predicted TTT nose times for X = Pt and Au are higher than those for U-10Mo, while U-9Mo-Ta remains marginally close. These results further support the predicted γ-phase stabilizing effect of Pt, Au, and Ta in U-10Mo alloys. In both Fig. 9a,b, Hf-containing compositions are predicted to exhibit a detrimental effect on γ-phase stability compared to U-10Mo, as indicated by lower predicted nose time values from the TTT curves; however, these values remain higher than those for U-8Mo and U-9Mo, respectively. Interestingly, the predicted effect of Hf addition on the γ-phase stability is less detrimental than that of Ti addition. For X = Ge and Ga, the TTT curves indicate a destabilizing effect on the γ-phase stability, reflected by lower predicted TTT nose times and a leftward shift relative to the U-10Mo curve for both U-8Mo-2X and U-9Mo-X. Comparisons for X = Ge and Ga are provided in Fig. S9a,b in the SI. Predicted TTT curves for previously unexplored (a) U-8Mo-2X (X= Ta, Pt, Au, Hf), (b) U-9Mo-X alloys (X= Ta, Pt, Au, Hf), (c) U-8Mo-2X (X = W, V, Fe) and (d) U-9Mo-X alloys (X = W, V, Fe). These alloys are compared with the predicted TTT curves of baseline alloys: U-8Mo (black triangles) and U-10Mo (red triangles) in (a) and (c), and U-9Mo (black triangles) and U-10Mo (red triangles) in (b) and (d). All predicted curves for the unexplored alloys exhibit higher TTT temperatures than the predicted U-10Mo curve. However, this increase is more evident in (a) and (b) than for the alloys in (c) and (d). In both (a) and (b), Pt- and Au-alloyed compositions demonstrate γ-phase stabilizing effect, as indicated by higher predicted TTT time values compared to the U-10Mo curve, while Ta-alloyed compositions are only marginally higher. The model does not predict a well-defined C-shaped curvature for U-8Mo-2Au in (a), and this is accompanied by high uncertainty values. Similarly, in (c) and (d) W alloying indicate γ-phase stabilizing effect, as shown by the shift in the predicted TTT curve to the right as compared to the U-10Mo curve, while Fe-alloyed compositions are only marginally higher. For Hf-alloyed compositions in (a) and (b), the curves shift to lower time values compared to U-10Mo, indicating a detrimental effect on γ-phase stability, which can also be seen for V-alloyed U-Mo compositions in (c) and (d). However, the ternary Hf-alloyed and V-alloyed systems exhibit a higher α-phase nucleation time compared to the binary U-8Mo and U-9Mo alloys. For the set of 12 U-Mo-X alloys chosen based on their frequent presence as elemental impurities in uranium ore concentrates, the TTT curves for U-Mo-X alloys with X = W, V, and Fe are compared against baseline alloys and are shown in Fig. 9c,d. Figure 9c shows the predicted U-8Mo-2X (X = W, V and Fe) alongside U-8Mo (black triangles) and U-10Mo (red triangles) predicted TTT diagrams. There is a slight increase in the predicted nose temperature corresponding to the predicted time for all the three U-8Mo-2X curves compared to U-10Mo. The predicted nose TTT time increases for X = W (blue circles) indicating enhanced γ-phase stability, whereas X = Fe (magenta circles) shows only a marginal increase relative to U-10Mo. For U-9Mo-X alloy compositions shown in Fig. 9d, the nose temperature corresponding to the predicted nose time for all alloys exceeds that of the baseline U-10Mo curve again but the increase is marginal. The predicted TTT nose time for X = W is higher than those for U-10Mo, while U-9Mo-Fe remains marginally close. These results further support the predicted γ-phase stabilizing effect of W-alloying in U-10Mo alloys. V-alloyed U-Mo compositions are predicted to exhibit a detrimental effect on γ-phase stability compared to U-10Mo, as indicated by lower predicted nose time values shown in both Fig. 9c,d; however, these values remain higher than those for the predicted TTT curve for U-8Mo and U-9Mo alloys, respectively. For X = Mg, Cu, and Al, the TTT curves indicate a destabilizing effect on the uranium γ-phase stability, reflected by lower predicted TTT nose times and a leftward shift of the predictive TTT curves relative to the U-10Mo curve for both U-8Mo-2X and U-9Mo-X. Comparisons for X = Mg, Cu, and Al are provided in Fig. S9c,d in the SI. Although TTT curves provide insight into γ-phase stability when alloyed with ternary elements, additional factors, such as their absorption cross-sections and any activation products generated as well as potential interactions with fission products and daughter products during reactor operation, are also critical in determining an optimal alloy composition for a given application. Benchmarking the performance of the physics-regularized eSVR model against single-point predictions A Gaussian Process Regression (GPR) model is developed to benchmark the physics-regularized eSVR framework against traditional single-point prediction methods by predicting TTT nose features, particularly the time at the lowest transformation rate. The input to this model is PC1, PC2, PC3, and corresponding to each alloy and the output is the nose TTT time. Temperature is excluded as an input descriptor in this model, unlike in the physics-regularized eSVR model above, where the TTT temperature serves as an input variable. The TTT curve data for 42 alloys presented in the Johns and Burkes paper16 are also utilized in this model. A 90:10 split is used to generate the training and test split comprising of 37 and 5 points, respectively. The automatic relevance determination (ARD) kernel is used for the GPR model and the optimized hyperparameter values are given in Table S3 in the SI. The GPR model performs well on both training and test sets with an value of 0.97 and 0.95, respectively. The parity plot of experimentally measured nose and GPR predicted nose is given in Fig. 10a. The blue dashed line represents the points where experimentally measured and predicted values align. The GPR-predicted nose time closely matches experimentally reported values along the blue dashed line for both the training and test sets. Unlike the previous method by Johns and Burkes, the GPR approach provides associated uncertainty estimates for the predicted TTT nose time, as noted in the paper. To further analyze the relationship between predicted nose time and alloy composition, and to compare these with experimental trends, both experimental and predicted nose times are plotted against varying Mo and Zr content in U-Mo and U-Mo-X alloys in Fig. S10 in the SI. The GPR predictions accurately reflect experimentally observed trends: increasing Mo content systematically extends nose time in binary alloys, whereas Zr additions reduce nose time, indicating accelerated transformation kinetics. This observation aligns with the reported destabilizing effect of Zr on the metastable γ-phase. TTT nose time predictions from the GPR model: (a) predictions for the training set (black circles) and test set (red triangles) compared with experimental values; (b) predictions for unexplored U-8Mo-2X ternary alloys compared with physics-regularized eSVR model predictions; (c) predictions for unexplored U-9Mo-X ternary alloys compared with physics-regularized eSVR model predictions; (d) experimental TTT nose times for all 42 alloys (combined training and test sets) plotted as a function of PC2. The single-point TTT nose time predictions from the GPR model are plotted against those from the physics-regularized eSVR model in Fig. 10b,c. Alloys with baseline composition U-8Mo-2X, where X = Hf, Pt, Ta, Au, Fe, V, and W, are shown in Fig. 10b. Pt- and Fe-alloyed compositions lie near the blue dashed line, indicating strong agreement between the GPR and the eSVR model with a hybrid loss function. However, GPR predictions for Hf-, V-, and Ta-alloyed compositions are lower than the eSVR predictions, whereas for X = Au and W, the GPR predictions exceed those of the eSVR model. The difference between the GPR and eSVR model predictions is significant for the U-8Mo-2Au alloy, with the GPR model exhibiting very high uncertainty values. Figure 10c shows that predictions for U-9Mo-X baseline alloys exhibit closer agreement between the Gaussian Process Regression (GPR) and eSVR models, with most points near the blue dashed line. However, for the Au-alloyed U-9Mo-Au composition, the GPR prediction remains significantly higher than the eSVR prediction. Additional GPR model predictions for ternary alloys with X = Ge, Ga, Mg, Cu, and Al are provided in Fig. S11 in the SI. Two post hoc model explanation techniques, namely GFI and PDP, were employed to analyze the learning patterns of the GPR model. The GFI plot shows that the PC2 descriptor is more important than the other descriptors, as shown in Fig. S12 in the SI. The PDP reveals an inverse correlation between PC2 values and the logarithmic predicted nose time values, as presented in Fig. S13c in the SI. This finding aligns with the trend observed in Fig. 10d, where the experimentally reported TTT nose time for 42 U-Mo-X alloys decreases as PC2 values increase. Furthermore, this result corroborates the previous discussion, where the PDP for PC2 in the eSVR model, which incorporates a loss function combining semi-empirical regularization and shape constraints in Fig. 6c, exhibited a similar trend of decreasing predicted time values with increasing normalized PC2 values. Extending the GPR model to predict complete TTT curves using multi-output GPR53 produced non-physical predictions. The resulting predictions were unable to capture the alloy dependence of the TTT curves as explained in Section S2.5 in the SI. Conclusion This study establishes a quantitative relationship between U-based alloy compositions and their TTT diagrams using physics-regularized ML models. Within a defined temperature range and by representing alloy composition with elemental descriptors, these models achieve strong predictive accuracy for TTT times. After initial dataset curation, PCA is applied to eliminate descriptor redundancy and reduce data dimensionality. To improve generalizability, four variations of the eSVR model loss function are evaluated: (1) baseline mean squared error (MSE), (2) MSE regularized with a semi-empirical nucleation-based model, (3) MSE with shape constraints enforcing characteristic C-curve convexity, and (4) a hybrid approach combining both semi-empirical regularization and shape constraints with MSE. The coefficient of determination () values for training and test datasets decrease as additional terms are added to the baseline MSE loss function, with a notable reduction upon inclusion of the curvature term that enforces the C-shape of the TTT curves. Post hoc model explanations based on PDPs reveal trends learned by the trained models and map elemental descriptors to the predicted time-value trends via PCA loadings. Ultimately, the eSVR model integrating both semi-empirical regularization and shape constraints is used to predict TTT curves for U-9.8Mo-0.2Ni, U-9.8Mo-0.2Cr, U-9.8Mo-0.2Co, and U-8Mo-Ti, none of which are present in the training or test datasets. Models incorporating semi-empirical regularization and curvature-loss function terms, despite lower values, produce C-shaped predicted TTT curves that are compared with experimentally reported trends. The models accurately capture the qualitative trend associated with nose time but systematically overestimate nose temperature. This model is also applied to predict TTT curves for 24 previously unexplored alloys, out of which the ternary elements for 12 of these alloys are selected based on the solubility of the ternary element with the γ-phase uranium, identifying Pt-, Ta-, and Au-alloyed U-Mo compositions as candidates for stabilizing the γ-phase. The ternary elements for the other 12 U-Mo-X alloys are selected based on their frequent presence as elemental impurities in uranium ore concentrates. These predictions indicated that W- and Fe-alloying have a stabilizing effect on the γ-phase. The predicted nose time values for these new alloys obtained with the physics-regularized hybrid model are benchmarked against a GPR model that predicts nose time solely using the MSE loss function. The findings from this study demonstrate the importance of a constraint-based ML framework for preserving the C-shape of predicted TTT curves. Incorporating these constraints improves alignment with experimental results and enables accelerated alloy design through TTT curve predictions. The predictions await experimental validation. Supplementary Information Supplementary Information. Author contributions S.G. and J.B. developed the machine learning models, with S.G. leading the eSVR and physics-regularized eSVR models and post-hoc model explanation, and J.B. leading the GPR models. S.R.A. and P.V.B. supervised the model development. The study was conceived and designed by S.G., J.B., V.V.J., S.R.A., and P.V.B. The manuscript was written by S.G., J.B., V.V.J., S.R.A., and P.V.B. Funding was acquired by P.V.B. and S.R.A. All authors reviewed and approved the manuscript. Funding All authors gratefully acknowledge support from the Battelle Memorial Institute, Pacific Northwest Division under Contract No. DE-AC05-76RL01830. Data availability The data that support the findings of this study are available in the UVA Dataverse (https://doi.org/10.18130/V3/EXJA3W). 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/s41598-026-56755-4. References Associated Data Supplementary Materials Supplementary Information. Data Availability Statement The data that support the findings of this study are available in the UVA Dataverse (https://doi.org/10.18130/V3/EXJA3W).