Elucidating the constitutive relationship of calcium–silicate–hydrate gel using high throughput reactive molecular simulations and machine learning

Gideon A. Lyngdoh, Hewenxuan Li, Mohd Zaki, N. M. Anoop Krishnan, Sumanta DasScientific 2020 · DOI

TL;DR

This paper tackles the challenge of predicting the mechanical properties of C-S-H gel, a crucial component of concrete, by combining high-throughput molecular dynamics simulations with machine learning. They generated a comprehensive dataset of C-S-H elastic constants and found that neural networks (NNs) excel at predicting these properties, with SHAP analysis revealing that the silicate network significantly influences all elastic constants, while water content plays a key role in shear properties.

What Problem Does This Paper Address?

Concrete, the most widely used construction material, relies on calcium-silicate-hydrate (C-S-H) gel as its primary binding phase. Optimizing C-S-H properties is key to developing high-performance and durable concrete. However, the complex, heterogeneous, and poorly crystalline nature of C-S-H makes experimentally probing its composition-property relationships exceedingly difficult. Machine learning (ML) offers a promising avenue for accelerating materials design, but its application in cementitious materials is hampered by the lack of consistent, representative, and large experimental datasets. This paper addresses this gap by creating such a dataset using molecular dynamics simulations and then applying ML to elucidate the constitutive relationships of C-S-H.

Methods

The core of this research lies in the synergistic integration of two powerful techniques:

  • High-Throughput Reactive Molecular Dynamics (MD) Simulations: The authors employed the ReaxFF reactive force field to simulate over 300 different compositions of C-S-H gel. This approach allows for the simulation of chemical reactions and bond breaking/formation, crucial for accurately modeling C-S-H. They systematically varied the proportions of CaO, SiO2, and water to generate a comprehensive dataset of nine elastic constants for each composition.
  • Machine Learning (ML) Algorithms: A suite of ML algorithms was used to build predictive models for the elastic constants. This included:
  • Gaussian Process (GP) Regression: Used with both RBF and Matern kernels, GP provides not only predictions but also an estimate of the uncertainty associated with those predictions.
  • Neural Networks (NN): A multi-layer perceptron with two hidden layers and nine neurons per layer was employed. NNs were found to be particularly adept at handling the multi-target nature of predicting multiple elastic constants simultaneously.
  • Other ML techniques: Polynomial Regression (PR), Random Forest (RF), Support Vector Machine (SVM), k-Nearest Neighbors (k-NN), and Decision Trees (DT) were also evaluated for comparative analysis.
  • SHapley Additive exPlanations (SHAP): To interpret the "black box" nature of the NN predictions, SHAP was utilized. This method quantifies the contribution of each input feature (CaO, SiO2, water content) to the model's output (elastic constants), providing crucial insights into the underlying physics.

Datasets

The primary dataset used in this study was generated de novo through the high-throughput MD simulations. This dataset comprises the nine elastic constants for over 300 distinct compositions of C-S-H gel, with variations in CaO, SiO2, and water content. This dataset is novel in its consistency, comprehensiveness, and coverage of experimentally relevant compositional ranges, addressing a key limitation in previous ML applications for cementitious materials.

Key Insights and Ideas

  • ML as a Bridge for Complex Materials: The study demonstrates that ML, when coupled with physics-based simulations, can effectively bridge the gap between fundamental material structure and macroscopic properties, especially for complex materials like C-S-H where experimental data is scarce.
  • NNs Excel at Multi-Target Prediction: Neural networks, by implicitly considering the interdependencies between multiple outputs, outperformed other ML models in predicting the nine elastic constants of C-S-H, highlighting their suitability for multivariate regression problems.
  • Silicate Network Dominance: SHAP analysis revealed that the silicate network (SiO2 content) has a significantly higher influence on all elastic constants of C-S-H compared to water and CaO content. This underscores the fundamental role of the silicate structure in determining the material's mechanical response.
  • Water's Role in In-Plane vs. Transverse Response: Interestingly, water content was found to have a more pronounced influence on the shear components (in-plane response) than on the normal components (transverse response along interlayer spaces). This suggests that water molecules primarily govern the material's response to shear forces, while the silicate network dictates its behavior under normal stress.

Key Results

  • High Predictive Accuracy: Neural networks achieved excellent prediction accuracy for the elastic constants, with R² values generally above 0.80 for the test set across most constants. Gaussian Process regression also performed well, though slightly less accurate than NNs.
  • Dataset Validation: The generated MD dataset showed good agreement with existing experimental data for bulk modulus and density, lending confidence to the simulation results.
  • Feature Importance Revealed: SHAP analysis provided quantitative insights into feature importance. For instance, SiO2 was consistently the most influential feature across all elastic constants. Water content showed a higher impact on shear moduli (C44, C55, C66) compared to normal moduli (C11, C33).
  • Non-Linear Relationships: The study confirmed that the relationship between C-S-H composition and its elastic constants is non-linear and coupled, making simple linear models insufficient for accurate prediction.

Limitations and Future Work

  • Computational Cost of NNs: The paper notes that while NNs offer superior accuracy, they require significant computational resources and time for training.
  • Interpretability of Complex Models: While SHAP provides valuable insights, interpreting the complex interactions within NN models remains an ongoing challenge.
  • Experimental Validation: The study relies on MD simulations. Future work could focus on experimentally validating these ML-driven predictions to further solidify the findings.
  • Broader Material Properties: The current study focuses on elastic constants. Extending this approach to predict other mechanical properties (e.g., Young's modulus, hardness) or transport properties would be a valuable next step.
  • Accelerated Design: The authors envision this work as a starting point for developing integrated experiment-multiscale simulation-ML-based design strategies for optimizing C-S-H nanostructures to create more efficient and durable cementitious binders.

Summary updated

Abstract

Prediction of material behavior using machine learning (ML) requires consistent, accurate, and, representative large data for training. However, such consistent and reliable experimental datasets are not always available for materials. To address this challenge, we synergistically integrate ML with high-throughput reactive molecular dynamics (MD) simulations to elucidate the constitutive relationship of calcium-silicate-hydrate (C-S-H) gel-the primary binding phase in concrete formed via the hydration of ordinary portland cement. Specifically, a highly consistent dataset on the nine elastic constants of more than 300 compositions of C-S-H gel is developed using high-throughput reactive simulations. From a comparative analysis of various ML algorithms including neural networks (NN) and Gaussian process (GP), we observe that NN provides excellent predictions. To interpret the predicted results from NN, we employ SHapley Additive exPlanations (SHAP), which reveals that the influence of silicate network on all the elastic constants of C-S-H is significantly higher than that of water and CaO content. Additionally, the water content is found to have a more prominent influence on the shear components than the normal components along the direction of the interlayer spaces within C-S-H. This result suggests that the in-plane elastic response is controlled by water molecules whereas the transverse response is mainly governed by the silicate network. Overall, by seamlessly integrating MD simulations with ML, this paper can be used as a starting point toward accelerated optimization of C-S-H nanostructures to design efficient cementitious binders with targeted properties.
BibTeX
@inproceedings{lyngdoh2020elucidating,
  author = {Gideon A. Lyngdoh and Hewenxuan Li and Mohd Zaki and N. M. Anoop Krishnan and Sumanta Das},
  title = {{Elucidating the constitutive relationship of calcium–silicate–hydrate gel using high throughput reactive molecular simulations and machine learning}},
  booktitle = {Scientific},
  year = {2020},
  doi = {10.1038/s41598-020-78368-1},
}
Figure