Computational Modeling and Simulation of Materials
Research Area Faculty
Research Area Overview
The research area of Computational Modeling and Simulation of Materials was established at CaSToRC in 2020, following the award of the SimEA ERA-Chair grant. The scientific team of SimEA, led by Prof. Dr. Vangelis Harmandaris, pursues a program of research excellence and innovation by applying and developing mathematical and computational methods integrated with High Performance Computing, for tackling challenging problems in engineering, materials science and biotechnological applications.
The research directions include:
- Multi-scale Modelling, Ab-initio/Atomistic Simulations of Nanostructured Materials
- Physics-based, Data-driven Machine Learning for Modelling Across Scales
- Modelling of Biomolecular Systems for Biotechnology Applications
- Bayesian Inference for Uncertainty Quantification and Model Selection
Applications feature the computer design of a broad range of materials and processes, especially complex molecular systems, such as nanocomposites, polymers, graphene-based nanostructured materials, proteins, and biomolecular systems. The team has several co-develop projects with industrial partners, centered on the in-silico design of advanced, sustainable materials for the green transition. Examples include membranes for carbon capture and storage, electrode/electrolyte optimization in energy storage, waste valorization and sustainable, bio based composite materials.
Research Highlights
Research Highlight 1
Title: A Computational Methodology for Predicting the Distribution of Mechanical Properties in Atomistic Models of Polymeric Nanostructured Materials (NANOMEC)
Related people: Hilal Reda, Anthony Chazirakis, Alireza F. Behbahani, Panayiota Katsamba, Nikos Savva, Vangelis Harmandaris
Graphical Abstract
Description: NANOMEC offers several key features: It calculates stress and strain at the atomic or particle level for nanostructured atomistic model configurations under applied deformation, averages these locally defined stress and strain values over user-defined subdomains within the nanostructured polymer nanocomposite system, and predicts the mechanical properties within these subdomains, with a particular focus on the polymer/solid interphase regions.
Overview
NANOMEC proposes a methodology for calculating the distribution of the mechanical properties in model atomistic polymer-based nanostructured systems. The use of atomistic simulations is key to unravelling the fundamental mechanical behaviour of composite materials. Most simulations involving the mechanical properties of polymer nanocomposites (PNCs) concern their global (average) properties, which are typically extracted by applying macroscopic strain on the boundaries of the simulation box and calculating the total (global) stress by invoking the Virial formalism over all atoms within the simulation box. Then, the pertinent (average) mechanical properties of the entire system can be extracted, from the corresponding stress-strain relation.
Scientific Achievement
NANOMEC refers to a computational methodology that directly computes the distribution of local stress and strain fields of atomistic model polymer nanocomposite systems under a globally applied external strain field.
Significance and impact
NANOMEC advances current understanding by providing a computational framework that bridges the gap between atomistic-level interactions and macroscopic mechanical behaviour in polymer nanocomposites. Traditional methods often struggle to capture the complexity of local stress and strain distributions, particularly in heterogeneous materials such as PNCs, where polymer/solid interphases play a critical role in determining overall properties.
Research Details
The whole methodology of NANOMEC can be summarized as follows:
- Initialization: Start from realistic model configurations obtained via long equilibrium atomistic simulations
- Stage 1: For a given strain rate, impose deformation in the equilibrium configurations up to a specific strain;
- Stage 2: Analyze the deformed configurations to compute stress and strain distributions;
- Stage 3: For a specific domain decomposition scheme, compute the average stress and strain within subdomains;
- Stage 4: Perform stages 1-3 iteratively for various deformations and strain values. Compute the mechanical properties within subdomains, via the local stress-strain relations.
References
-
H. Reda, A. Chazirakis, A. F. Behbahani, N. Savva, and V. Harmandaris, ‘Mechanical properties of glassy polymer nanocomposites via atomistic and continuum models: The role of interphases’, Computer Methods in Applied Mechanics and Engineering, vol. 395, p. 114905, May 2022, doi: 10.1016/j.cma.2022.114905.
-
Reda, A. Chazirakis, A. J. Power, and V. Harmandaris, ‘Mechanical Behavior of Polymer Nanocomposites via Atomistic Simulations: Conformational Heterogeneity and the Role of Strain Rate’, J. Phys. Chem. B, vol. 126, no. 38, pp. 7429–7444, Sep. 2022, doi: 10.1021/acs.jpcb.2c04597.
-
Reda, A. Chazirakis, A. F. Behbahani, N. Savva, and V. Harmandaris, ‘Revealing the Role of Chain Conformations on the Origin of the Mechanical Reinforcement in Glassy Polymer Nanocomposites’, Nano Lett., vol. 24, no. 1, pp. 148–155, Jan. 2024, doi: 10.1021/acs.nanolett.3c03491.
-
A set of Python scripts used to calculate and visualize the atomistic strain of a molecular trajectory can be founded: https://github.com/SimEA-ERA/NANOMEC.git
Research Highlight 2
Title: Physics-Informed Machine Learning Models for Multi-Scale Modelling of Complex Molecular Systems
Related people: Eleftherios Christofi, Nikolaos Patsalidis, Charalambos Chrysostomou, Petra Bačová, Anthony Chazirakis, Manolis Doxastakis, Wei Li, Mihalis Nicolaou, Vangelis Harmandaris
Graphical Abstract
Description: a) Schematic representation of the active learning (AL) algorithm. The algorithm generates efficient physics aware atomistic models via training on actively augmented Density Functional Theory (DFT) datasets. b) Physics-informed conditional U-net-based models were utilized for reintroducing atomistic degrees of freedom in coarse-grained molecular configurations of linear poly(lactic acid) in melt of arbitrary composition. Depicted are PLLA and PDLA monomers.
Overview
Multi-scale modelling of complex molecular systems, such as macromolecules, and molecules on solid surfaces demands bridging multiple spatiotemporal scales. It encompasses methods that link the quantum to the atomistic scale (Figure 1a), and, especially in the case of polymers, atomistic to coarse grained (CG) scale and CG back to the atomistic (Figure 1b). Therefore, information can pass through the different scales, capturing different material properties. The AL algorithm iteratively trains physics aware atomistic models on actively augmented quantum scale DFT datasets, learning structural correlations to energy. The backmapping method learns structural correlations between atomistic and coarse-grained mappings while respecting molecular features like local packing.
Scientific Achievement
We propose two innovative computational approaches: a deep learning-based backmapping method for multicomponent macromolecules and a physics-aware AL algorithm for interatomic interactions, specialized in molecule/solid interfaces. The backmapping approach uses U-net based models trained directly on probability distribution functions defined across chemical bonds and distances between atoms and CG particles, incorporating physical priors like chemical bonds and bond angles. The AL algorithm combines physics-driven and flexible parametric functions, such as Bezier-Bernstein polynomials, to model complex dependencies in interfacial systems. Both methods achieve high accuracy and computational efficiency, enabling versatile applications in molecular and interfacial modelling.
Significance and impact
These methods address key challenges in multi-scale modelling and interatomic interaction modelling by balancing accuracy, efficiency, and general applicability. The backmapping approach can generate atomistic configurations of arbitrary microstructures and perform on high molecular weight polymer melts, while the AL algorithm models strong physisorption, chemisorption, and cohesion in rough surfaces using actively augmented DFT data. These models facilitate the design of materials, such as gas nanosensors, nanoelectronics, and adhesive nanocomposites, broadening the scope of molecular and interfacial system applications.
Research Details
The methodologies are structured as follows:
- Backmapping of CG macromolecules:
- Pre-processing: Collect data on atomistic descriptors, i.e., atomistic bond vectors (target data), conditioned on CG coordinates and their corresponding types (input data).
- Training process: Train model to predict atomistic bond vectors conditioned on the given CG configurations.
- Post-processing: Generate atomistic configurations in Cartesian space.
- Active Learning for interatomic interactions:
- Modeling: Use physics-aware potential functions combined with flexible Bezier-Bernstein polynomials to capture complex dependencies.
- Data augmentation: Actively generate and select structurally dissimilar energetically favored data from molecular dynamics simulations.
- Numerical stability: Ensure smooth and controllable modeling, reducing overfitting behavior and improving accuracy.
References
N. Patsalidis et al., “Generic active learning algorithm for atomic clusters and their interaction with gases.”, Nat. Comput. Sci., submitted
E. Christofi et al., ‘Deep convolutional neural networks for generating atomistic configurations of multi-component macromolecules from coarse-grained models’, The Journal of Chemical Physics, vol. 157, no. 18, p. 184903, Nov. 2022, doi: 10.1063/5.0110322.
E. Christofi, P. Bačová, and V. A. Harmandaris, ‘Physics-Informed Deep Learning Approach for Reintroducing Atomic Detail in Coarse-Grained Configurations of Multiple Poly(lactic acid) Stereoisomers’, J. Chem. Inf. Model., vol. 64, no. 6, pp. 1853–1867, Mar. 2024, doi: 10.1021/acs.jcim.3c01870.
N. Patsalidis, G. Papamokos, G. Floudas, and V. Harmandaris, ‘Understanding the Interaction between Polybutadiene and Alumina via Density Functional Theory Calculations and Machine-Learned Atomistic Simulations’, J. Phys. Chem. C, vol. 126, no. 39, pp. 16792–16803, Oct. 2022, doi: 10.1021/acs.jpcc.2c03630.
Backmapping of CG macromolecules:
A set of Python scripts used to train the deep learning model and examine the performance of the trained model by investigating the predicted atomistic structures, can be found in the following GitHub repositories:
Active learning for interatomic interactions:
A set of Bash and Python scripts used for the implementation of the method, can be found in the following GitHub repository:
Selected Publications
- P. Bačová et al., ‘Coupling between Polymer Conformations and Dynamics Near Amorphous Silica Surfaces: A Direct Insight from Atomistic Simulations’, Nanomaterials, vol. 11, no. 8, p. 2075, Aug. 2021, doi: 10.3390/nano11082075.
- G. Baxevani, V. Harmandaris, E. Kalligiannaki, and I. Tsantili, ‘Bottom-Up Transient Time Models in Coarse-Graining Molecular Systems’, Multiscale Model. Simul., vol. 21, no. 4, pp. 1746–1774, Dec. 2023, doi: 10.1137/23M1548451.
- D. Demou and N. Savva, ‘AI-assisted modeling of capillary-driven droplet dynamics’, DCE, vol. 4, p. e24, 2023, doi: 10.1017/dce.2023.19.
- N. Patsalidis, G. Papamokos, G. Floudas, and V. Harmandaris, ‘Temperature dependence of the dynamics and interfacial width in nanoconfined polymers via atomistic simulations’, The Journal of Chemical Physics, vol. 160, no. 10, p. 104904, Mar. 2024, doi: 10.1063/5.0189652.
- H. Reda, A. Chazirakis, A. F. Behbahani, N. Savva, and V. Harmandaris, ‘Revealing the Role of Chain Conformations on the Origin of the Mechanical Reinforcement in Glassy Polymer Nanocomposites’, Nano Lett., vol. 24, no. 1, pp. 148–155, Jan. 2024, doi: 10.1021/acs.nanolett.3c03491.
