Surrogate Models for the Compressive Stress–Strain Response of Nanosilica/PVA-Modified Engineered Cementitious Composites: A Leakage-Controlled, Blind-Mixture Benchmark of Machine-Learning, Physics-Regularized, Spline, and Operator Architectures

Constitutive surrogates for nanosilica- and polyvinyl alcohol (PVA)-modified engineered cementitious composites (ECCs) may reduce the cost of evaluating compressive stress–strain responses across mix designs. This study benchmarks data-driven, spline-based, operator-learning, and physics-regularized architectures using 24 digitized experimental curves, one per mixture, resampled to 737 within-curve collocation points; the composition-level sample size is therefore 24, and no specimen-level replicates were available. Random pointwise and blind-mixture validation are compared to distinguish within-curve interpolation from prediction for unseen mixtures. Across repeated five-mixture hold-outs, XGBoost achieved the lowest mean RMSE: 2.56±0.21 MPa over six splits and 2.77±0.61 MPa over fifteen splits compared with 3.41±0.88 MPa for the automatically selected (zero-physics-weight) neural network over the same fifteen splits (exact two-sided Wilcoxon W=9, unadjusted p=0.002; the splits overlap and are not independent replicates). On the principal blind split, this automatically selected neural-network configuration achieved R2=0.9930 and RMSE =1.81 MPa; however, its selected physics-loss weights were zero, so this is a data-only result and does not establish an improvement due to physical regularization. Adding damage and smoothness penalties to the perceptron increased its blind RMSE (to 4.96 MPa or 5.72 MPa, depending on the weights). Under non-matched physics objectives, a 576-parameter spline network with an elastic-modulus penalty achieved R2=0.9810 and RMSE =2.98 MPa compared with R2=0.9302 and RMSE =5.72 MPa for a 6368-weight perceptron trained with all three penalties; this contrast cannot be attributed to architecture alone. The soft damage-boundedness and smoothness penalties did not guarantee monotonic damage evolution or complete constitutive admissibility; the damage measure used is an apparent stiffness-degradation indicator rather than a validated damage law for fiber-bridged ECC, and Gaussian-process intervals covered 86.2% (89.9% after recalibration) of blind points against a nominal 95%. Differences between digitized and companion-study tabulated peak properties (up to 27.3% in peak stress) remain unresolved, so all performance statements are conditional on the digitized benchmark. These findings support mixture-level validation, the careful attribution of physics-regularization effects, and the verification of source data before any deployment in nonlinear structural simulations.

Authors

Institutions

Publication Details

Journal
Polymers
Published
2026-11-16
DOI
https://doi.org/10.3390/polym18222877
Primary Topic
Innovative concrete reinforcement materials
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Surrogate Models for the Compressive Stress–Strain Response of Nanosilica/PVA-Modified Engineered Cementitious Composites: A Leakage-Controlled, Blind-Mixture Benchmark of Machine-Learning, Physics-Regularized, Spline, and Operator Architectures

Haifa F. Alhasson, Shuaa S. Alharbi, Mohammed Aqeel Albadrani
2 links
Polymers
Innovative concrete reinforcement materials
article

Surrogate Models for the Compressive Stress–Strain Response of Nanosilica/PVA-Modified Engineered Cementitious Composites: A Leakage-Controlled, Blind-Mixture Benchmark of Machine-Learning, Physics-Regularized, Spline, and Operator Architectures

Haifa F. Alhasson, Shuaa S. Alharbi, Mohammed Aqeel Albadrani
article en

Abstract

Constitutive surrogates for nanosilica- and polyvinyl alcohol (PVA)-modified engineered cementitious composites (ECCs) may reduce the cost of evaluating compressive stress–strain responses across mix designs. This study benchmarks data-driven, spline-based, operator-learning, and physics-regularized architectures using 24 digitized experimental curves, one per mixture, resampled to 737 within-curve collocation points; the composition-level sample size is therefore 24, and no specimen-level replicates were available. Random pointwise and blind-mixture validation are compared to distinguish within-curve interpolation from prediction for unseen mixtures. Across repeated five-mixture hold-outs, XGBoost achieved the lowest mean RMSE: 2.56±0.21 MPa over six splits and 2.77±0.61 MPa over fifteen splits compared with 3.41±0.88 MPa for the automatically selected (zero-physics-weight) neural network over the same fifteen splits (exact two-sided Wilcoxon W=9, unadjusted p=0.002; the splits overlap and are not independent replicates). On the principal blind split, this automatically selected neural-network configuration achieved R2=0.9930 and RMSE =1.81 MPa; however, its selected physics-loss weights were zero, so this is a data-only result and does not establish an improvement due to physical regularization. Adding damage and smoothness penalties to the perceptron increased its blind RMSE (to 4.96 MPa or 5.72 MPa, depending on the weights). Under non-matched physics objectives, a 576-parameter spline network with an elastic-modulus penalty achieved R2=0.9810 and RMSE =2.98 MPa compared with R2=0.9302 and RMSE =5.72 MPa for a 6368-weight perceptron trained with all three penalties; this contrast cannot be attributed to architecture alone. The soft damage-boundedness and smoothness penalties did not guarantee monotonic damage evolution or complete constitutive admissibility; the damage measure used is an apparent stiffness-degradation indicator rather than a validated damage law for fiber-bridged ECC, and Gaussian-process intervals covered 86.2% (89.9% after recalibration) of blind points against a nominal 95%. Differences between digitized and companion-study tabulated peak properties (up to 27.3% in peak stress) remain unresolved, so all performance statements are conditional on the digitized benchmark. These findings support mixture-level validation, the careful attribution of physics-regularization effects, and the verification of source data before any deployment in nonlinear structural simulations.

PolymersVol. 18(22)
Qassim University (SA)
Industry, innovation and infrastructure
Innovative concrete reinforcement materials
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon