Saurabh Tiwari
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Research Subjects
Research Interests
Saurabh Tiwari has a total of 99 co-authors. Here are the top co-authors by number of shared publications:
- Nokeun Park
School of Materials Science and Engineering, Yeungnam University, Gyeongsan 38541, Republic of Korea
15 shared publications - Somaditya Sen
Department of Physics Indian Institute of Technology Indore 453552 Indore India
13 shared publications - Nagireddy Gari Subba Reddy
School of Materials Science and Engineering, Engineering Research Institute, Gyeongsang National University, Jinju-si 52725, Republic of Korea
11 shared publications Sajal Biring
11 shared publications- Muhammad Ishtiaq
Department of Materials Engineering and Convergence Technology, Gyeongsang National University, Jinju, Republic of Korea
5 shared publications - Seong Jun Heo
School of Materials Science and Engineering, Yeungnam University, Gyeongsan 38541, Republic of Korea
5 shared publications
Latest Publications
- Open Access
- 9 November 2026
- Published by MDPI AG in Pharmaceutics
- Vol. 18 (11) , 1449
Multi-Task Learning of Blood–Brain Barrier Permeability and hERG Cardiotoxicity: A Random- and Scaffold-Split Evaluation on Experimental Data
Background/Objectives: Blood–brain barrier (BBB) permeability and cardiac hERG liability share physicochemical determinants, motivating the hypothesis that multi-task learning could improve BBB prediction. Methods: We tested this hypothesis using B3DB logBB regression (n = 1058) and a curated hERG dataset (n = 8960). A shared-encoder multi-task neural network was compared with a matched single-task network and Bayesian-optimized gradient boosting under ten random splits, 30 Murcko scaffold splits, and held-out classification sets. Results: Multi-task learning did not improve random-split R2 (0.485 ± 0.016 vs. 0.482 ± 0.015). Under scaffold splitting, its mean R2 was higher (−0.054 vs. −0.817) only because its failures were less severe (the same number of partitions had negative R2); medians were nearly identical (0.394 vs. 0.398). Neither loss weighting, network tuning, nor removal of 39.6% of test compounds sharing a parent structure with hERG training data changed this. Gradient boosting was more robust (scaffold R2 = 0.388 ± 0.188; better in 28–29 of 30 partitions), although not significantly so under a dependence-corrected test. Conclusions: Network failures occurred only when a scaffold group containing 18 atypical compounds (e.g., xenon) entered the test set; after removing 20 such compounds, multi-task R2 (0.409) matched gradient boosting (0.420). On 3031 held-out parent structures, AUCs were similar (0.764–0.772).
- 9 November 2026
- Published by MDPI AG in Pharmaceutics
- Vol. 18 (11) , 1449
Open AccessMulti-Task Learning of Blood–Brain Barrier Permeability and hERG Cardiotoxicity: A Random- and Scaffold-Split Evaluation on Experimental Data
Background/Objectives: Blood–brain barrier (BBB) permeability and cardiac hERG liability share physicochemical determinants, motivating the hypothesis that multi-task learning could improve BBB prediction. Methods: We tested this hypothesis using B3DB logBB regression (n = 1058) and a curated hERG dataset (n = 8960). A shared-encoder multi-task neural network was compared with a matched single-task network and Bayesian-optimized gradient boosting under ten random splits, 30 Murcko scaffold splits, and held-out classification sets. Results: Multi-task learning did not improve random-split R2 (0.485 ± 0.016 vs. 0.482 ± 0.015). Under scaffold splitting, its mean R2 was higher (−0.054 vs. −0.817) only because its failures were less severe (the same number of partitions had negative R2); medians were nearly identical (0.394 vs. 0.398). Neither loss weighting, network tuning, nor removal of 39.6% of test compounds sharing a parent structure with hERG training data changed this. Gradient boosting was more robust (scaffold R2 = 0.388 ± 0.188; better in 28–29 of 30 partitions), although not significantly so under a dependence-corrected test. Conclusions: Network failures occurred only when a scaffold group containing 18 atypical compounds (e.g., xenon) entered the test set; after removing 20 such compounds, multi-task R2 (0.409) matched gradient boosting (0.420). On 3031 held-out parent structures, AUCs were similar (0.764–0.772).
- Open Access
- 5 November 2026
- Published by MDPI AG in Pharmaceutics
- Vol. 18 (11) , 1441
Graph Neural Networks Versus Descriptor-Based Machine Learning for Blood–Brain Barrier Permeability Prediction: A Matched-Split Evaluation
Background/Objectives: Predicting blood–brain barrier (BBB) permeability from structure remains challenging. We evaluated whether graph neural networks (GNNs) improve experimental logBB prediction over descriptor-based machine learning and whether GNN attention agrees with earlier descriptor-based findings. Methods: Graph Convolutional Network (GCN) and Graph Attention Network (GAT) models were trained on 1058 B3DB compounds and compared with LightGBM using 209 RDKit 2D descriptors under identical random and Murcko scaffold splits. Analyses comprised a primary split, 10 repeated random and scaffold splits with nested hyperparameter optimization and paired tests, descriptor-augmented GNNs, classification, external evaluation on the Martins benchmark, and GAT attention analysis. Results and Discussion: In the primary split, LightGBM achieved the highest R2 (random 0.673; scaffold 0.604) versus GCN (0.514; 0.133) and GAT (0.430; 0.310). Across repeated splits with nested hyperparameter optimization, LightGBM outperformed both GNNs under both protocols (paired t-test p < 0.001). GCN and GAT did not differ significantly under scaffold splitting in either the original repeated-split evaluation (p = 0.49) or the optimized comparison (p = 0.184). Descriptor augmentation improved both GNNs (p < 0.001); under scaffold splitting, augmented GCN (R2 = 0.463 ± 0.084) was not significantly different from LightGBM (0.494 ± 0.093; p = 0.22). Using thresholded regression predictions, LightGBM had the highest mean AUC in every condition (≈0.87 versus 0.77–0.84). On 1627 compounds from the independently curated Martins benchmark, mean AUC was 0.84 (LightGBM), 0.83 (GCN) and 0.81 (GAT). GAT attention was higher on nitrogen/oxygen than carbon in 82 of 98 eligible molecules, but halogen, sulfur, and phosphorus atoms received even more attention. Conclusions: Under the evaluated conditions, 2D descriptors outperformed basic GCN and GAT models across splits; descriptor augmentation substantially narrowed the gap. GAT attention provides hypothesis-generating, not causal, insight.
- 5 November 2026
- Published by MDPI AG in Pharmaceutics
- Vol. 18 (11) , 1441
Open AccessGraph Neural Networks Versus Descriptor-Based Machine Learning for Blood–Brain Barrier Permeability Prediction: A Matched-Split Evaluation
Background/Objectives: Predicting blood–brain barrier (BBB) permeability from structure remains challenging. We evaluated whether graph neural networks (GNNs) improve experimental logBB prediction over descriptor-based machine learning and whether GNN attention agrees with earlier descriptor-based findings. Methods: Graph Convolutional Network (GCN) and Graph Attention Network (GAT) models were trained on 1058 B3DB compounds and compared with LightGBM using 209 RDKit 2D descriptors under identical random and Murcko scaffold splits. Analyses comprised a primary split, 10 repeated random and scaffold splits with nested hyperparameter optimization and paired tests, descriptor-augmented GNNs, classification, external evaluation on the Martins benchmark, and GAT attention analysis. Results and Discussion: In the primary split, LightGBM achieved the highest R2 (random 0.673; scaffold 0.604) versus GCN (0.514; 0.133) and GAT (0.430; 0.310). Across repeated splits with nested hyperparameter optimization, LightGBM outperformed both GNNs under both protocols (paired t-test p < 0.001). GCN and GAT did not differ significantly under scaffold splitting in either the original repeated-split evaluation (p = 0.49) or the optimized comparison (p = 0.184). Descriptor augmentation improved both GNNs (p < 0.001); under scaffold splitting, augmented GCN (R2 = 0.463 ± 0.084) was not significantly different from LightGBM (0.494 ± 0.093; p = 0.22). Using thresholded regression predictions, LightGBM had the highest mean AUC in every condition (≈0.87 versus 0.77–0.84). On 1627 compounds from the independently curated Martins benchmark, mean AUC was 0.84 (LightGBM), 0.83 (GCN) and 0.81 (GAT). GAT attention was higher on nitrogen/oxygen than carbon in 82 of 98 eligible molecules, but halogen, sulfur, and phosphorus atoms received even more attention. Conclusions: Under the evaluated conditions, 2D descriptors outperformed basic GCN and GAT models across splits; descriptor augmentation substantially narrowed the gap. GAT attention provides hypothesis-generating, not causal, insight.
- Open Access
Distribution-Aware Estimation of A-/B-Basis Characteristic Values for Natural-Fiber Composite Laminates from Single-Source Tensile Data
Natural-fiber composites exhibit large property scatter, and multi-batch qualification data are rarely available. Using 590 tensile specimens from ten flax, jute, and carbon-fiber laminate groups in a single experimental program, we applied established one-sided lower tolerance limits as approximations of A-basis (99% coverage, 95% confidence) and B-basis (90%, 95%) values for strength, modulus, and failure strain, and compared normal-theory, bootstrap, family-specific pivotal, and selection-robust routes. A Monte Carlo study (eight scenarios, 5000 replicates) showed that the pivotal limit reached a mean A-basis confidence of 0.949 when the family was known but only 0.896 (minimum 0.805) when the family was selected by AIC; an envelope over all three families reached at least 0.944 at the cost of lower limits. An empirical bootstrap cannot exceed a confidence of 1–0.99n (0.311 at n = 37). Two jute-twill records with a zero video-strain reading at peak load were excluded from the strain analysis by a rule based solely on the measurement record, raising the A-basis failure strain by 45%. Retention rankings partly changed when the sample size was equalized. In the five groups that contain a second file-name series, tensile strength differed significantly between the series (Welch p ≤ 0.008), so the pooled limits for these groups assume identically distributed specimens. The values are screening approximations and not design allowables.
Open AccessDistribution-Aware Estimation of A-/B-Basis Characteristic Values for Natural-Fiber Composite Laminates from Single-Source Tensile Data
Natural-fiber composites exhibit large property scatter, and multi-batch qualification data are rarely available. Using 590 tensile specimens from ten flax, jute, and carbon-fiber laminate groups in a single experimental program, we applied established one-sided lower tolerance limits as approximations of A-basis (99% coverage, 95% confidence) and B-basis (90%, 95%) values for strength, modulus, and failure strain, and compared normal-theory, bootstrap, family-specific pivotal, and selection-robust routes. A Monte Carlo study (eight scenarios, 5000 replicates) showed that the pivotal limit reached a mean A-basis confidence of 0.949 when the family was known but only 0.896 (minimum 0.805) when the family was selected by AIC; an envelope over all three families reached at least 0.944 at the cost of lower limits. An empirical bootstrap cannot exceed a confidence of 1–0.99n (0.311 at n = 37). Two jute-twill records with a zero video-strain reading at peak load were excluded from the strain analysis by a rule based solely on the measurement record, raising the A-basis failure strain by 45%. Retention rankings partly changed when the sample size was equalized. In the five groups that contain a second file-name series, tensile strength differed significantly between the series (Welch p ≤ 0.008), so the pooled limits for these groups assume identically distributed specimens. The values are screening approximations and not design allowables.
- 29 October 2026
- Published by Taylor & Francis in Journal of Adhesion Science and Technology
- p. 1-12
Unravelling the microstructural origin of strength–ductility synergy in friction stir welded dual-phase brass
1. FSW is a solid-state joining process that uses a rotating welding tool to simultaneously generate frictional heat, plastic deformation, and mechanical stirring action, thereby firmly welding two...
- 29 October 2026
- Published by Taylor & Francis in Journal of Adhesion Science and Technology
- p. 1-12
Unravelling the microstructural origin of strength–ductility synergy in friction stir welded dual-phase brass
1. FSW is a solid-state joining process that uses a rotating welding tool to simultaneously generate frictional heat, plastic deformation, and mechanical stirring action, thereby firmly welding two...
- Open Access
Investigation of Microstructural and Properties of Steels and Alloys
Advances in steel and metallic alloys continue to rely on a fundamental understanding of the relationships among chemical composition, processing, microstructure, and resulting properties
Open AccessInvestigation of Microstructural and Properties of Steels and Alloys
Advances in steel and metallic alloys continue to rely on a fundamental understanding of the relationships among chemical composition, processing, microstructure, and resulting properties
- Open Access
Interpretable Machine Learning for Mechanical Property Prediction of 5Cr-0.5Mo Steel: SHAP Explainability, Multi-Model Comparison, and Uncertainty Quantification
5Cr-0.5Mo ferritic steels are widely used in high-temperature power-plant components. Although artificial neural network (ANN) models have shown good performance in predicting tensile properties, they provide limited insight into predictions and generally do not quantify the uncertainty. In this study, three tree-based machine learning models—Random Forest (RF), XGBoost (XGB), and Gradient Boosting (GB)—were developed using 36 unique alloy grade–temperature observations from a validated NIMS 5Cr-0.5Mo tensile dataset. The model performance was evaluated using leave-one-grade-out (LOGO) cross-validation, with pooled out-of-fold (OOF) predictions used to assess the overall performance. SHapley Additive exPlanations (SHAP) were used to examine feature contributions, whereas Gaussian Process Regression (GPR) was evaluated as a proof-of-concept for uncertainty quantification of yield strength (YS). GB showed the strongest performance for ultimate tensile strength (UTS) and reduction in area (RA), achieving pooled OOF R2 values of 0.9698 and 0.9570, respectively. RF achieved corresponding R2 values of 0.9406 and 0.9488, respectively. SHAP identified the test temperature as the most influential feature across all four properties, whereas the Cr content and austenite grain size contributed significantly to the strength predictions. For YS, the GPR achieved complete empirical coverage of the 95% predictive intervals, although the relatively large mean interval width indicated conservative uncertainty estimates. Given the limited dataset and feature correlations, the SHAP results should be regarded as exploratory, rather than mechanistic. Overall, this study demonstrates the potential of interpretable, uncertainty-aware ML for small alloy datasets, while emphasizing the need for larger, compositionally diverse datasets and independent validation.
Open AccessInterpretable Machine Learning for Mechanical Property Prediction of 5Cr-0.5Mo Steel: SHAP Explainability, Multi-Model Comparison, and Uncertainty Quantification
5Cr-0.5Mo ferritic steels are widely used in high-temperature power-plant components. Although artificial neural network (ANN) models have shown good performance in predicting tensile properties, they provide limited insight into predictions and generally do not quantify the uncertainty. In this study, three tree-based machine learning models—Random Forest (RF), XGBoost (XGB), and Gradient Boosting (GB)—were developed using 36 unique alloy grade–temperature observations from a validated NIMS 5Cr-0.5Mo tensile dataset. The model performance was evaluated using leave-one-grade-out (LOGO) cross-validation, with pooled out-of-fold (OOF) predictions used to assess the overall performance. SHapley Additive exPlanations (SHAP) were used to examine feature contributions, whereas Gaussian Process Regression (GPR) was evaluated as a proof-of-concept for uncertainty quantification of yield strength (YS). GB showed the strongest performance for ultimate tensile strength (UTS) and reduction in area (RA), achieving pooled OOF R2 values of 0.9698 and 0.9570, respectively. RF achieved corresponding R2 values of 0.9406 and 0.9488, respectively. SHAP identified the test temperature as the most influential feature across all four properties, whereas the Cr content and austenite grain size contributed significantly to the strength predictions. For YS, the GPR achieved complete empirical coverage of the 95% predictive intervals, although the relatively large mean interval width indicated conservative uncertainty estimates. Given the limited dataset and feature correlations, the SHAP results should be regarded as exploratory, rather than mechanistic. Overall, this study demonstrates the potential of interpretable, uncertainty-aware ML for small alloy datasets, while emphasizing the need for larger, compositionally diverse datasets and independent validation.
- Open Access
- 2 September 2026
- Published by MDPI AG in International Journal of Molecular Sciences
- Vol. 27 (17) , 7864
Integrating Artificial Intelligence with Emerging Pharmaceutical Technologies: Current Progress, Clinical Translation, and Future Challenges
Modern scientific and technological developments are driving major advances in drug research and development. This narrative review, based on a structured search of PubMed, Scopus, and Web of Science (2018–2026), examines how artificial intelligence (AI) and machine learning (ML) are accelerating a historically prolonged and expensive process, alongside pharmacogenomics, organ-on-a-chip systems, three-dimensional (3D) bioprinting, and nanotechnology. In benchmark studies, deep learning techniques have achieved an area under the receiver operating characteristic curve (AUROC) of over 0.85 for a subset of absorption, distribution, metabolism, excretion, and toxicity (ADMET) endpoints. AI-powered models show promising, albeit platform-dependent, accuracy in predicting candidate drug properties. Pharmacogenomics enables personalized medicine by tailoring therapies according to patients’ genetic profiles, whereas organ-on-a-chip systems and 3D bioprinting provide physiologically relevant human tissue models for preclinical evaluation. In a blinded benchmark study, the Emulate Liver-Chip showed 87% sensitivity and 100% specificity for drug-induced liver injury, outperforming animal models in that specific comparison. Nanotechnology is advancing drug delivery through the use of nanoparticle systems, such as Doxil® and Onpattro®. Obstacles remain, including regulatory constraints, ethical considerations, data quality limitations, and the need for stronger validation, although ongoing funding, interdisciplinary collaboration, and evolving regulatory frameworks may support further development. Overall, these technologies show meaningful potential to shorten development time and improve treatment safety, although further prospective validation is required before realizing this potential at scale.
- 2 September 2026
- Published by MDPI AG in International Journal of Molecular Sciences
- Vol. 27 (17) , 7864
Open AccessIntegrating Artificial Intelligence with Emerging Pharmaceutical Technologies: Current Progress, Clinical Translation, and Future Challenges
Modern scientific and technological developments are driving major advances in drug research and development. This narrative review, based on a structured search of PubMed, Scopus, and Web of Science (2018–2026), examines how artificial intelligence (AI) and machine learning (ML) are accelerating a historically prolonged and expensive process, alongside pharmacogenomics, organ-on-a-chip systems, three-dimensional (3D) bioprinting, and nanotechnology. In benchmark studies, deep learning techniques have achieved an area under the receiver operating characteristic curve (AUROC) of over 0.85 for a subset of absorption, distribution, metabolism, excretion, and toxicity (ADMET) endpoints. AI-powered models show promising, albeit platform-dependent, accuracy in predicting candidate drug properties. Pharmacogenomics enables personalized medicine by tailoring therapies according to patients’ genetic profiles, whereas organ-on-a-chip systems and 3D bioprinting provide physiologically relevant human tissue models for preclinical evaluation. In a blinded benchmark study, the Emulate Liver-Chip showed 87% sensitivity and 100% specificity for drug-induced liver injury, outperforming animal models in that specific comparison. Nanotechnology is advancing drug delivery through the use of nanoparticle systems, such as Doxil® and Onpattro®. Obstacles remain, including regulatory constraints, ethical considerations, data quality limitations, and the need for stronger validation, although ongoing funding, interdisciplinary collaboration, and evolving regulatory frameworks may support further development. Overall, these technologies show meaningful potential to shorten development time and improve treatment safety, although further prospective validation is required before realizing this potential at scale.
- 1 September 2026
- Published by AIP Publishing in Applied Physics Reviews
- Vol. 13 (3)
Exciton–plasmon coupling: Mechanisms and emerging technologies
Metal–semiconductor heterostructures provide a versatile platform for tuning light–matter interactions at the nanoscale. At the heterostructure interface, excitonic resonances in semiconductors interact with plasmonic fields of metallic nanostructures giving rise to exciton-plasmon coupling (EPC). This coupling controls energy transfer, charge separation and carrier dynamics across the interface. The strength of EPC ultimately dictates the macroscopic optoelectronic response of the hybrid systems. The coupling strength is sensitive to several interdependent parameters, such as electromagnetic field confinement, oscillator strength, nanostructure geometry, and dissipative interactions, giving rise to diverse interaction regimes, from strong to weak, which show distinct physical characteristics. Despite extensive experimental and theoretical progress, a comprehensive understanding of the emergence and evolution of EPC remains elusive, as coherent hybridization, energy/charge transfer and dissipative processes often coexist within the hybrid system. The interplay between these processes manifests in complex spectral signatures like anticrossing, linewidth modification and intensity enhancement, often obscuring the underlying coupling mechanism. In this review, we present a unified perspective on EPC, treating strong, intermediate, and weak coupling within a continuous conceptual framework, ranging from hybrid eigenstates to perturbative modifications of excitonic and plasmonic resonances. This framework provides a basis for linking these ambiguous spectral signatures across coupling regimes to their underlying microscopic processes. The review further links the EPC mechanisms to device performance in functional and scalable technologies like biosensing, photodetection, photovoltaics, and quantum systems. Finally, we evaluate the fundamental and architectural bottlenecks that hinder translation beyond proof-of-concept demonstrations, thus offering a critical design roadmap for the rational design of scalable EPC-based systems. We argue that addressing these constraints requires a shift from demonstration-driven studies to custom designs that prioritize loss engineering, interface control, mechanism-resolved characterization, and scalable architectures treated as primary design parameters.
- 1 September 2026
- Published by AIP Publishing in Applied Physics Reviews
- Vol. 13 (3)
Exciton–plasmon coupling: Mechanisms and emerging technologies
Metal–semiconductor heterostructures provide a versatile platform for tuning light–matter interactions at the nanoscale. At the heterostructure interface, excitonic resonances in semiconductors interact with plasmonic fields of metallic nanostructures giving rise to exciton-plasmon coupling (EPC). This coupling controls energy transfer, charge separation and carrier dynamics across the interface. The strength of EPC ultimately dictates the macroscopic optoelectronic response of the hybrid systems. The coupling strength is sensitive to several interdependent parameters, such as electromagnetic field confinement, oscillator strength, nanostructure geometry, and dissipative interactions, giving rise to diverse interaction regimes, from strong to weak, which show distinct physical characteristics. Despite extensive experimental and theoretical progress, a comprehensive understanding of the emergence and evolution of EPC remains elusive, as coherent hybridization, energy/charge transfer and dissipative processes often coexist within the hybrid system. The interplay between these processes manifests in complex spectral signatures like anticrossing, linewidth modification and intensity enhancement, often obscuring the underlying coupling mechanism. In this review, we present a unified perspective on EPC, treating strong, intermediate, and weak coupling within a continuous conceptual framework, ranging from hybrid eigenstates to perturbative modifications of excitonic and plasmonic resonances. This framework provides a basis for linking these ambiguous spectral signatures across coupling regimes to their underlying microscopic processes. The review further links the EPC mechanisms to device performance in functional and scalable technologies like biosensing, photodetection, photovoltaics, and quantum systems. Finally, we evaluate the fundamental and architectural bottlenecks that hinder translation beyond proof-of-concept demonstrations, thus offering a critical design roadmap for the rational design of scalable EPC-based systems. We argue that addressing these constraints requires a shift from demonstration-driven studies to custom designs that prioritize loss engineering, interface control, mechanism-resolved characterization, and scalable architectures treated as primary design parameters.
- Open Access
Domain-Knowledge-Guided Feature Engineering for Small-Sample Machine Learning Prediction of Mechanical Properties in Low-Carbon Hot-Rolled Steel Strips
Industrial steel property prediction is often constrained by limited labelled data, reducing the effectiveness of conventional machine learning models. This study investigated whether metallurgy-informed feature engineering enhances predictive performance under small-data conditions. A representative set of 300 samples from an industrial low-carbon hot-rolled steel strip dataset (C: 0.02–0.06 wt%; Mn: 0.17–0.38 wt%) was used to derive five physically meaningful descriptors: carbon equivalent (CE), nitrogen-to-aluminum ratio (N/Al), microalloying efficiency index (MEI), thermal processing parameter (TPP), and solid solution strengthening index (SSSI). These descriptors were combined with the original 17 compositional and processing variables to create a 22-feature dataset. Random Forest (RF) and Extreme Gradient Boosting (XGBoost) models were evaluated on an independent 60-sample test set using 5-fold cross-validation. Feature engineering improved the prediction accuracy, with the greatest gain observed for elongation. For XGBoost, the mean percentage error decreased from 3.23% to 3.05%, whereas the test-set R2 increased from 0.4935 to 0.5444, representing a 10.3% improvement in the explained variance. For the yield strength, the Random Forest method increased the R2 from 0.4744 to 0.4861. Permutation importance and partial dependence analyses identified MEI and TPP as the six most influential predictors across all targets, confirming that the engineered descriptors provide complementary metallurgical information. Learning curve analysis showed slightly higher cross-validation R2 values at intermediate training sizes (n = 125–175), indicating modestly improved sample efficiency. These findings establish domain-informed feature engineering as an interpretable and practical strategy for improving machine learning in data-limited steel manufacturing processes.
Open AccessDomain-Knowledge-Guided Feature Engineering for Small-Sample Machine Learning Prediction of Mechanical Properties in Low-Carbon Hot-Rolled Steel Strips
Industrial steel property prediction is often constrained by limited labelled data, reducing the effectiveness of conventional machine learning models. This study investigated whether metallurgy-informed feature engineering enhances predictive performance under small-data conditions. A representative set of 300 samples from an industrial low-carbon hot-rolled steel strip dataset (C: 0.02–0.06 wt%; Mn: 0.17–0.38 wt%) was used to derive five physically meaningful descriptors: carbon equivalent (CE), nitrogen-to-aluminum ratio (N/Al), microalloying efficiency index (MEI), thermal processing parameter (TPP), and solid solution strengthening index (SSSI). These descriptors were combined with the original 17 compositional and processing variables to create a 22-feature dataset. Random Forest (RF) and Extreme Gradient Boosting (XGBoost) models were evaluated on an independent 60-sample test set using 5-fold cross-validation. Feature engineering improved the prediction accuracy, with the greatest gain observed for elongation. For XGBoost, the mean percentage error decreased from 3.23% to 3.05%, whereas the test-set R2 increased from 0.4935 to 0.5444, representing a 10.3% improvement in the explained variance. For the yield strength, the Random Forest method increased the R2 from 0.4744 to 0.4861. Permutation importance and partial dependence analyses identified MEI and TPP as the six most influential predictors across all targets, confirming that the engineered descriptors provide complementary metallurgical information. Learning curve analysis showed slightly higher cross-validation R2 values at intermediate training sizes (n = 125–175), indicating modestly improved sample efficiency. These findings establish domain-informed feature engineering as an interpretable and practical strategy for improving machine learning in data-limited steel manufacturing processes.
- Open Access
- 17 August 2026
- Published by MDPI AG in Pharmaceutics
- Vol. 18 (8) , 1018
Evaluating the Incremental Value of Three-Dimensional Conformer Descriptors for Blood–Brain Barrier Permeability Prediction Using the B3DB Benchmark Dataset
Background/Objectives: Predicting blood–brain barrier (BBB) permeability remains a major challenge in central nervous system (CNS) drug discovery. Three-dimensional (3D) conformer-derived molecular descriptors are often proposed as improvements over conventional two-dimensional (2D) topological representations; however, their incremental predictive value remains unclear. Methods: Here, we systematically evaluated six molecular feature sets comprising curated 2D Mordred descriptors, Morgan/ECFP4 fingerprints, 3D conformer-derived descriptors, and their combinations, using LightGBM regression on the B3DB benchmark dataset (1054 compounds with experimental logBB values). The model performance was assessed using 30 independent random splits and 30 Murcko scaffold-based splits. Results: The 2D descriptor model achieved mean test R2 values of 0.567 ± 0.060 and 0.418 ± 0.085 under random and scaffold splitting, respectively, whereas the 3D descriptor model performed substantially worse (0.430 ± 0.066 and 0.298 ± 0.097, respectively). Incorporating 3D descriptors into the 2D feature set yielded only a marginal improvement under random splitting (+0.008 R2; p = 0.039), which disappeared during scaffold-based validation (p = 0.599). Similarly, adding 3D descriptors to the combined 2D + fingerprint representation did not yield any significant benefits. Fingerprints alone exhibited pronounced scaffold fragility, with the mean R2 decreasing from 0.441 to 0.312 between the random and scaffold evaluations. SHAP analysis identified TopoPSA (NO) as the dominant predictor (mean |SHAP| = 0.332), highlighting the central role of desolvation in BBB permeation. Applicability domain analysis showed that 94.8% of the test compounds fell within the structural coverage of the training set. Conclusions: Overall, on the B3DB benchmark and under the descriptor implementation evaluated here, 3D conformer descriptors provided limited incremental value and no scaffold-robust advantage over well-curated 2D molecular representations for BBB permeability predictions.
- 17 August 2026
- Published by MDPI AG in Pharmaceutics
- Vol. 18 (8) , 1018
Open AccessEvaluating the Incremental Value of Three-Dimensional Conformer Descriptors for Blood–Brain Barrier Permeability Prediction Using the B3DB Benchmark Dataset
Background/Objectives: Predicting blood–brain barrier (BBB) permeability remains a major challenge in central nervous system (CNS) drug discovery. Three-dimensional (3D) conformer-derived molecular descriptors are often proposed as improvements over conventional two-dimensional (2D) topological representations; however, their incremental predictive value remains unclear. Methods: Here, we systematically evaluated six molecular feature sets comprising curated 2D Mordred descriptors, Morgan/ECFP4 fingerprints, 3D conformer-derived descriptors, and their combinations, using LightGBM regression on the B3DB benchmark dataset (1054 compounds with experimental logBB values). The model performance was assessed using 30 independent random splits and 30 Murcko scaffold-based splits. Results: The 2D descriptor model achieved mean test R2 values of 0.567 ± 0.060 and 0.418 ± 0.085 under random and scaffold splitting, respectively, whereas the 3D descriptor model performed substantially worse (0.430 ± 0.066 and 0.298 ± 0.097, respectively). Incorporating 3D descriptors into the 2D feature set yielded only a marginal improvement under random splitting (+0.008 R2; p = 0.039), which disappeared during scaffold-based validation (p = 0.599). Similarly, adding 3D descriptors to the combined 2D + fingerprint representation did not yield any significant benefits. Fingerprints alone exhibited pronounced scaffold fragility, with the mean R2 decreasing from 0.441 to 0.312 between the random and scaffold evaluations. SHAP analysis identified TopoPSA (NO) as the dominant predictor (mean |SHAP| = 0.332), highlighting the central role of desolvation in BBB permeation. Applicability domain analysis showed that 94.8% of the test compounds fell within the structural coverage of the training set. Conclusions: Overall, on the B3DB benchmark and under the descriptor implementation evaluated here, 3D conformer descriptors provided limited incremental value and no scaffold-robust advantage over well-curated 2D molecular representations for BBB permeability predictions.