Session D18: AI and ML in Monte Carlo and molecular dynamics simulations

Free Energy, Conformationa Dynamics and Simulations of Nanocrystals with Explicit Ligands

Materials whose fundamental units are nanocrystals (NC)s, instead of atoms or molecules, are emerging as major candidates to solve many of the technological challenges of our century. In this talk I will present different algorithms and methods to compute thermodynamic and dynamical quantities when NC are modeled with explicit ligands, mostly through all atom or united models, but time permitting, some discussion will be included for coarse-grained systems. I will introduce HOODLT, a software developed in my group that systematized and simplifies this type of calculations. I will discuss in some detail concrete examples of NC assembled into superlattices by two different strategies: Solvent evaporation and tuning electrostatic interactions in water as a solvent with an emphasis in free energies and ligand conformations.   

Evaluating approaches for on-the-fly machine-learning interatomic potentials for activated mechanisms sampling with ARTn

Methods, such as the Activation-Relaxation Technique Nouveau (ARTn), have proven effective at finding and identifying activated mechanisms pathways in solid-state materials.

The mechanisms discovered are physically accurate within the realm of ab initio calculations. However, due to its computational cost, this method is limited to small systems and a small set of events. However, for most problems of interest, these accurate calculations are out of reach, and empirical potentials must be used, which are notoriously unreliable resulting in qualitative results.

To overcome this limitation, in the last few years, much efforts have gone into developing general machine-learning potentials able to describe interactions for a wide range of structures and phases. Still, as attention turns to more complex materials including alloys, disordered and heterogeneous systems, the challenge of providing reliable description for all possible environment become ever more costly.

In this work, we evaluate the benefits of using specific versus general on-the-fly machine-learning potentials for the study of activated mechanisms. More specifically, we tests three fitting approaches using the moment-tensor potential to reproduce a reference potential when exploring the energy landscape around a vacancy in silicon crystal and silicon-germanium zincblende structure using ARTn.

We find that a a targeted approach specific and integrated to ARTn generates the highest precision on the energetic and geometry of activated events, while remaining cost-effective. This approach expands the type of problems that can be addressed with high-accuracy ML potentials.   

Development of MLIP to model corrosion behavior in Molten Salt Reactors

In this study, we developed and evaluated the efficacy of Machine Learning Interatomic Potentials (MLIP) designed for Molten Salt and its relevance toward the corrosion behavior. We implemented a number of methodologies and ML codes to develop the potentials, ranging the Moment Tensor Potentials (MTP), Invariant-based Deep Learning Potentials to the Equavariant-based Neural Network Potentials. We optimized the hyperparameters and incorporated the long-range interactions to account for the initial corrosion mechanisms and the cluster dynamics within the molten salts. We then compare our results with the experimental observations pertaining to, for example, de-alloying mechanisms in Ni-based alloys observed under corrosive environments.   

Machine learning assisted design of effective potentials, surface ligand patterns, and annealing protocols for colloidal self-assembly

Colloidal self-assembly provides a scalable route to creating nanomaterials with new architectures and functions. Molecular models and simulations have played an integral role in our understanding of the solvent-mediated interactions between colloidal particles, the assembly morphologies that emerge from these interactions, and the self-assembly process itself. However, detailed molecular models of the building blocks and the solvent are not ideal for exploring assembly behavior or constructing phase diagrams, as the simulations are computational expensive, and the colloidal design space is often vast. Furthermore, the relationship between assembly morphology and design is often complex, so the “inverse design” of particles targeting a given assembly morphology also requires brute-force exploration of the design space. Here, we will demonstrate how machine learning can speed up both the exploration and the targeted design efforts in colloidal assembly. First, we will discuss the development of an analytical potential based on permutationally invariant polynomials for describing the effective multibody interactions between spherical polymer-grafted nanoparticles in a polymer melt. The potential reduces the computational cost of assembly simulations by several orders of magnitude, allowing us to explore assembly behavior over large length and time scales and thereby obtain phases such as strings and hexagonal sheets that cannot be assessed using two-body potentials, and discover novel phases such as networks, clusters, and gels. Next, we will discuss the implementation of a neural adjoint framework for inverse-design of DNA-origami building blocks that can self-assemble into periodic superstructures based on patches of hydrophobic brushes introduced at specific locations on the origamis. Lastly, we will discuss how machine learning can be used for optimizing assembly protocols for improving yield. Our system consists of a binary system of ligand-grafted nanoparticles trapped at a fluid-fluid interface that self-assemble into stripe patterns and quasicrystals. Using a neuroevolutionary algorithm, we determine the optimal temperature annealing profiles that best limit the formation of defects in these assemblies.   

First-principles machine-learning quantum dynamics at 0K in SrTiO3: light-induced ultrafast ferroelectric transition

Low temperature nonequilibrium quantum dynamics in crystals is extremely challenging, and has not been possible to perform on materials of realistic complexity. In this work we develop a novel technique, the time-dependent self-consistent harmonic approximation (TDSCHA [1]), and use it to simulate the quantum dynamics in SrTiO₃ at 0K. We combine TDSCHA with state-of-the-art machine learning force fields to accelerate calculations by multiple orders of magnitude compared to first principles dynamics. 

We study the light-induced ferroelectric transition in SrTiO₃, where nuclear quantum fluctuations play a major role in stabilizing the paraelectric phase at low temperatures. 

Our approach allows for an unprecedented description of the light-induced ferroelectric transition due to the absence of empirical parameters, paving the way to the development of next-generation ultrafast nonvolatile memories.  

[1] Lorenzo Monacelli and Francesco Mauri, Time-dependent self-consistent harmonic approximation: Anharmonic nuclear quantum dynamics and time correlation functions, Phys. Rev. B 103104305 (2021)   

Diverse training data generation for machine-learning interatomic potentials

Machine-learning interatomic potentials (MLIAPs) make it feasible to aim for both accuracy and transferability, something that the earlier generations of potentials struggled to achieve. However, given their flexibility, these potentials often fail at extrapolating to properties beyond the training data, which makes the quality of the training set the determining factor in their performance. Training datasets typically consist of DFT energies and forces of relatively small systems, traditionally selected manually or randomly from subsets of the configuration space of interest. The need for human intervention in the curation of training sets makes their generation labor-intensive and time-consuming. Here, we present a generalization of a previously developed method based on the automated maximization of the information entropy of the descriptor distribution. The diversity of the entropy-optimized training dataset is compared to several other datasets from the literature. In addition, this method is applied to train MLIAPs for Be, W, and Re. W and Be are primary candidates for first wall material applications in fusion reactors, while Re is a product of transmutation of W due to neutron bombardment. The transferability of the MLIAPs trained using the entropy-optimized training data is compared to that of the traditional curation approaches, highlighting the desirable characteristics of an optimal training data, irrespective of the material chemistry.   

A machine learning interatomic potential for Ge-Te alloys

Germanium-Tellurium alloys are complex materials with a range of important applications.  Computational study of these systems commonly uses density functional theory (DFT) which severely limits the size and time scales that can be probed.  We present an interatomic potential for GeTe alloys based on the recent atomic cluster expansion (ACE) architecture, trained with representative configurations from DFT calculations spanning the entirety of the Ge_xTe_1-x system. This system contains multiple complex thermodynamic processes, most notably the fast and reversible phase-change in the 50:50 composition and negative thermal expansion in the tellurium-heavy liquid, which makes developing a stoichiometrically transferable potential challenging. By combining a refined exchange-correlation functional for the solid state with an explicit inclusion of dispersion interactions, we obtain an accurate description of the structure of elemental germanium, tellurium, and its alloys capable of simulations on the timescales of the phase change process. We report on its ability to reproduce the experimental thermodynamic properties and discuss the implications for past and future atomistic phase-change material simulations.   

Machine learned force and torque predictions for molecular dynamics of non-spherical colloids

Composite-rigid bodies of smaller spherical beads are commonly used to simulate non-spherical colloids with Molecular Dynamics (MD). To accurately represent their shape and to obtain the desired effective pair interactions between two rigid bodies, each body may need to contain hundreds of beads. Traditional MD calculate all the inter-body distances between the beads of the rigid bodies to find the net force and torque on them. These distance calculations are computationally costly and limit the number of rigid bodies that can be simulated. However, the effective interaction between the two rigid bodies depends only on the distance between their center of mass and their relative orientation. This implies the existence of a function capable of mapping the center of mass distance and orientation to the interaction energy between the two rigid bodies, which would completely bypass the individual inter-body bead distance calculations. Deriving such a function analytically for nearly any non-spherical rigid body is a significant challenge. In this study, we have trained neural nets that take the pair configuration as input and give the forces and torques between the two rigid bodies (cylinders and cubes) as output. We show that MD simulations performed with neural net predicted forces and torques can accurately reproduce the structure and kinetics of the traditional simulations with explicit distance calculations. Neural-net assisted simulations can offer speed-ups, depending on system size and hardware.   

Machine Learning Models for Partition Functions: Predicting Thermodynamic Properties and Exploring Transition Pathways

In our recent work, we have focused on developing machine learning (ML) models for predicting partition functions. To do this, we carried out computationally intensive flat histogram Monte Carlo simulations based on a Wang-Landau sampling scheme. These simulations helped us to obtain the partition function for atomic and molecular systems. The results of these simulations were then combined into datasets that allowed us to train and validate ML models. We used these models to build artificial neural networks capable of predicting partition functions for a wide range of conditions.

 

We demonstrated the accuracy and reliability of our ML models by showing their ability to predict thermodynamic properties for single-component systems, mixtures, and confined fluids. Additionally, we showed that our model could be used to create reaction coordinates for exploring phase transition pathways.