Transfer (machine) learning approaches coupled with target data augmentation to predict the mechanical properties of concrete

Ford, Emily; Maneparambil, Kailasnath; Kumar, Aditya; Sant, Gaurav; Neithalath, Narayanan

doi:10.1016/j.mlwa.2022.100271

Cited by 10 publications

(5 citation statements)

References 35 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…This implies that the proposed knowledge integration methods may not be effective in terms of accuracy enhancement for ML models on the evaluated dataset. On the other hand, it is worthwhile to note that the estimated accuracy of the baseline NN model (mean RMSE =5.12 MPa) is within the range of the lowest RMSE values obtained by the existing studies on Yeh's dataset 11 (the source of the laboratory dataset used in this work), i.e., 4.5–6 MPa 5,12,41,57–59 . This suggests that the NN model may have achieved a prediction accuracy similar to the data accuracy (i.e., representativeness to the ground‐true observations 4 ), even without the use of prior knowledge.…”

Section: Resultssupporting

confidence: 72%

“…On the other hand, it is worthwhile to note that the estimated accuracy of the baseline NN model (mean RMSE = 5.12 MPa) is within the range of the lowest RMSE values obtained by the existing studies on Yeh's dataset 11 (the source of the laboratory dataset used in this work), i.e., 4.5-6 MPa. 5,12,41,[57][58][59] This suggests that the NN model may have achieved a prediction accuracy similar to the data accuracy (i.e., representativeness to the ground-true observations 4 ), even without the use of prior knowledge. Further accuracy improvement in the model may result in overfitting to data noise introduced by experiments (due to variation of raw materials, operators, and measuring equipment).…”

Section: Prediction Accuracymentioning

confidence: 94%

“…Transfer learning has demonstrated its potential to speed up the learning process and improve model generalizability even with limited training data 38–40 . In the context of compressive strength prediction for a certain type of concrete, a model pre‐trained on data for other properties (e.g., flexural or tensile strengths) of the same type of concrete or same property (i.e., compressive strength) of similar concrete type can be of significant use 41,42 . Nevertheless, existing transfer learning approaches rely on large datasets in the pre‐training process, which has been a challenge in the domain of concrete science (cf.…”

Section: Knowledge‐informed Machine Learningmentioning

confidence: 99%

See 2 more Smart Citations

Can domain knowledge benefit machine learning for concrete property prediction?

Li,

Pei,

Ying

et al. 2023

J Am Ceram Soc.

View full text Add to dashboard Cite

Understanding and predicting process–structure–property–performance relationships for concrete materials is key to designing resilient and sustainable infrastructure. While machine learning has emerged as a powerful tool to supplement empirical analysis and physical modeling, its capabilities are yet to be fully realized due to the massive data requirements and generalizability challenges. To address these limitations, we propose a knowledge‐informed machine learning framework for concrete property prediction that aggregates the wealth of domain knowledge condensed in empirical formulas and physics‐based models. By integrating the knowledge through data augmentation, feature enhancement, and model pre‐training, we demonstrate that this framework has the potential to (i) accelerate model convergence, (ii) improve model performance with limited training data, and (iii) increase generalizability to real‐world scenarios (including extrapolation capability to other datasets and robustness against data outliers). The overall improvement of machine learning models by knowledge integration is particularly critical when these models are scaled up to tackle the increasing complexity of modern concrete and deployed in practical applications. While demonstrated for predicting concrete strength, this versatile framework is applicable to a wide range of properties of concrete and other composite materials, paving the way for accelerated materials design and discovery.

show abstract

Section: Resultssupporting

confidence: 72%

Section: Prediction Accuracymentioning

confidence: 94%

Section: Knowledge‐informed Machine Learningmentioning

confidence: 99%

See 1 more Smart Citation

Can domain knowledge benefit machine learning for concrete property prediction?

Li,

Pei,

Ying

et al. 2023

J Am Ceram Soc.

View full text Add to dashboard Cite

show abstract

“…36,37 In recent years, transfer learning and multi-delity learning have also arrived in materials science. 7,[38][39][40][41][42][43][44][45][46][47][48][49] The published works deal with rather small datasets for both pre-training and transfer learning, usually (10 4 data points for the transfer dataset and (10 5 data points for the pre-training dataset. In this context, band gaps 7,38,46,48 and formation energies 39,48 are the most popular features for transfer learning, since there is abundance of multi-delity theoretical and experimental data (∼10 3 measurements).…”

Section: Introductionmentioning

confidence: 99%

Transfer learning on large datasets for the accurate prediction of material properties

Hoffmann,

Schmidt,

Botti

et al. 2023

Digital Discovery

View full text Add to dashboard Cite

show abstract

“…In recent years, transfer learning and multi-fidelity learning have also arrived in materials science [7,[34][35][36][37][38][39][40][41][42][43][44][45]. The published works deal with rather small datasets for both pre-training and transfer learning, usually < ∼ 10 4 data points for the transfer dataset and < ∼ 10 5 data points for the pre-training dataset.…”

Section: Introductionmentioning

confidence: 99%