Search

The accuracy of atomistic simulations depends on the precision of force fields. Traditional numerical methods often struggle to optimize the empirical force field parameters for reproducing target properties. Recent approaches rely on training these force fields based on forces and energies from first-principle simulations. However, it is unclear whether these approaches will enable capturing complex material responses such as vibrational, or elastic properties. To this extent, we introduce a framework, employing inner loop simulations and outer loop optimization, that exploits automatic differentiation for both property prediction and force-field optimization by computing gradients of the simulation analytically. We demonstrate the approach by optimizing classical Stillinger-Weber and EDIP potentials for silicon systems to reproduce the elastic constants, vibrational density of states, and phonon dispersion. We also demonstrate how a machine-learned potential can be fine-tuned using automatic differentiation to reproduce any target property such as radial distribution functions. Interestingly, the resulting force field exhibits improved accuracy and generalizability to unseen temperatures than those fine-tuned on energies and forces. Finally, we demonstrate the extension of the approach to optimize the force fields towards multiple target properties. Altogether, differentiable simulations, through the analytical computation of their gradients, offer a powerful tool for both theoretical exploration and practical applications toward understanding physical systems and materials., Comment: 15 pages, 5 figures

The time evolution of physical systems is described by differential equations, which depend on abstract quantities like energy and force. Traditionally, these quantities are derived as functionals based on observables such as positions and velocities. Discovering these governing symbolic laws is the key to comprehending the interactions in nature. Here, we present a Hamiltonian graph neural network (HGNN), a physics-enforced GNN that learns the dynamics of systems directly from their trajectory. We demonstrate the performance of HGNN on n-springs, n-pendulums, gravitational systems, and binary Lennard Jones systems; HGNN learns the dynamics in excellent agreement with the ground truth from small amounts of data. We also evaluate the ability of HGNN to generalize to larger system sizes, and to hybrid spring-pendulum system that is a combination of two original systems (spring and pendulum) on which the models are trained independently. Finally, employing symbolic regression on the learned HGNN, we infer the underlying equations relating the energy functionals, even for complex systems such as the binary Lennard-Jones liquid. Our framework facilitates the interpretable discovery of interaction laws directly from physical system trajectories. Furthermore, this approach can be extended to other systems with topology-dependent dynamics, such as cells, polydisperse gels, or deformable bodies.

Neural networks (NNs) that exploit strong inductive biases based on physical laws and symmetries have shown remarkable success in learning the dynamics of physical systems directly from their trajectory. However, these works focus only on the systems that follow deterministic dynamics, for instance, Newtonian or Hamiltonian dynamics. Here, we propose a framework, namely Brownian graph neural networks (BROGNET), combining stochastic differential equations (SDEs) and GNNs to learn Brownian dynamics directly from the trajectory. We theoretically show that BROGNET conserves the linear momentum of the system, which in turn, provides superior performance on learning dynamics as revealed empirically. We demonstrate this approach on several systems, namely, linear spring, linear spring with binary particle types, and non-linear spring systems, all following Brownian dynamics at finite temperatures. We show that BROGNET significantly outperforms proposed baselines across all the benchmarked Brownian systems. In addition, we demonstrate zero-shot generalizability of BROGNET to simulate unseen system sizes that are two orders of magnitude larger and to different temperatures than those used during training. Altogether, our study contributes to advancing the understanding of the intricate dynamics of Brownian motion and demonstrates the effectiveness of graph neural networks in modeling such complex systems.

Optimization of atomic structures presents a challenging problem, due to their highly rough and non-convex energy landscape, with wide applications in the fields of drug design, materials discovery, and mechanics. Here, we present a graph reinforcement learning approach, StriderNET, that learns a policy to displace the atoms towards low energy configurations. We evaluate the performance of StriderNET on three complex atomic systems, namely, binary Lennard-Jones particles, calcium silicate hydrates gel, and disordered silicon. We show that StriderNET outperforms all classical optimization algorithms and enables the discovery of a lower energy minimum. In addition, StriderNET exhibits a higher rate of reaching minima with energies, as confirmed by the average over multiple realizations. Finally, we show that StriderNET exhibits inductivity to unseen system sizes that are an order of magnitude different from the training system.

Physical systems are commonly represented as a combination of particles, the individual dynamics of which govern the system dynamics. However, traditional approaches require the knowledge of several abstract quantities such as the energy or force to infer the dynamics of these particles. Here, we present a framework, namely, Lagrangian graph neural network (LGnn), that provides a strong inductive bias to learn the Lagrangian of a particle-based system directly from the trajectory. We test our approach on challenging systems with constraints and drag -- LGnn outperforms baselines such as feed-forward Lagrangian neural network (Lnn) with improved performance. We also show the zero-shot generalizability of the system by simulating systems two orders of magnitude larger than the trained one and also hybrid systems that are unseen by the model, a unique feature. The graph architecture of LGnn significantly simplifies the learning in comparison to Lnn with ~25 times better performance on ~20 times smaller amounts of data. Finally, we show the interpretability of LGnn, which directly provides physical insights on drag and constraint forces learned by the model. LGnn can thus provide a fillip toward understanding the dynamics of physical systems purely from observable quantities., Comment: 16 pages, 5 figures, Supplementary Material

Learning the structure--dynamics correlation in disordered systems is a long-standing problem. Here, we use unsupervised machine learning employing graph neural networks (GNN) to investigate the local structures in disordered systems. We test our approach on 2D binary A65B35 LJ glasses and extract structures corresponding to liquid, supercooled and glassy states at different cooling rates. The neighborhood representation of atoms learned by a GNN in an unsupervised fashion, when clustered, reveal local structures with varying potential energies. These clusters exhibit dynamical heterogeneity in the structure in congruence with their local energy landscape. Altogether, the present study shows that unsupervised graph embedding can reveal the structure--dynamics correlation in disordered structures., Comment: 13 pages, 11 figures

Vapor deposition can yield glasses that are more stable than those obtained by the traditional melt-quenching route. However, it remains unclear whether vapor-deposited glasses are "allowable" or "forbidden," that is, if they are equivalent to glasses formed by cooling extremely slowly a liquid or if they differ in nature from melt-quenched glasses. Here, based on reactive molecular dynamics simulation (MD) of silica glasses, we demonstrate that the allowable or forbidden nature of vapor-deposited glasses depends on the temperature of the substrate and, in turn, is found to be encoded in their medium-range order structure.

Despite the extensive usage of oxide glasses for a few millennia, the composition-property relationships in these materials still remain poorly understood. While empirical and physics-based models have been used to predict properties, these remain limited to a few select compositions or a series of glasses. Designing new glasses requires a priori knowledge of how the composition of a glass dictates its properties such as stiffness, density, or processability. Thus, accelerated design of glasses for targeted applications remain impeded due to the lack of universal composition-property models. Herein, using deep learning, we present a methodology for the rational design of oxide glasses. Exploiting a large dataset of glasses comprising of up to 37 oxide components and more than 100,000 glass compositions, we develop high-fidelity deep neural networks for the prediction of eight properties that enable the design of glasses, namely, density, Young's modulus, shear modulus, hardness, glass transition temperature, thermal expansion coefficient, liquidus temperature, and refractive index. These models are by far the most extensive models developed as they cover the entire range of human-made glass compositions. We demonstrate that the models developed here exhibit excellent predictability, ensuring close agreement with experimental observations. Using these models, we develop a series of new design charts, termed as glass selection charts. These charts enable the rational design of functional glasses for targeted applications by identifying unique compositions that satisfy two or more constraints, on both compositions and properties, simultaneously. The generic design approach presented herein could catalyze machine-learning assisted materials design and discovery for a large class of materials including metals, ceramics, and proteins., Comment: 15 pages, 5 figures

Due to its ability to bond with living tissues upon dissolution, 45S5 bioglass and related compositions are promising materials for the replacement, regeneration, and repair of hard tissues in the human body. However, the details of their atomic structure remain unclear. This is partially due to the non-equilibrium nature of glasses, as their non-crystalline structure is highly dependent on their thermal history, namely, the cooling rate used during quenching. Herein, using molecular dynamics (MD) simulations and magic angle spinning nuclear magnetic resonance (MAS-NMR) spectroscopy experiments, we investigate the structure of the nominal 45S5 bioglass composition prepared using cooling rates ranging over several orders of magnitude. We show that the simulations results are in very good agreement with experimental data, provided that they are extrapolated toward lower cooling rates achieved in experiments. These results highlight that previously reported inconsistencies between simulations and experiments stem from the difference in cooling rate, thereby addressing one of the longstanding questions on the structure of bioglass. Based on these results, we demonstrate the existence of a Si--P avoidance behavior, which may be key in controlling the bioactivity of 45S5 bioglass., Comment: 22 pages, 13 figures

Borosilicate glasses are traditionally challenging to model using atomic scale simulations due to the composition and thermal history dependence of the coordination state of B atoms. Here, we report a new empirical interatomic potential that shows a good transferability over a wide range of borosilicate glasses--ranging from pure silicate to pure borate end members--while relying on a simple formulation and a constant set of energy parameters. In particular, we show that our new potential accurately predicts the compositional dependence of the average coordination number of boron atoms, glass density, overall short-range and medium-range order structure, and shear viscosity values for several borosilicate glasses and liquids. This suggests that our new potential could be used to gain new insights into the structure of a variety of advanced borosilicate glasses to help elucidate composition-structure-property relationships--including in complex nuclear waste immobilization glasses.

Predicting the dissolution rates of silicate glasses in aqueous conditions is a complex task as the underlying mechanism(s) remain poorly understood and the dissolution kinetics can depend on a large number of intrinsic and extrinsic factors. Here, we assess the potential of data-driven models based on machine learning to predict the dissolution rates of various aluminosilicate glasses exposed to a wide range of solution pH values, from acidic to caustic conditions. Four classes of machine learning methods are investigated, namely, linear regression, support vector machine regression, random forest, and artificial neural network. We observe that, although linear methods all fail to describe the dissolution kinetics, the artificial neural network approach offers excellent predictions, thanks to its inherent ability to handle non-linear data. Overall, we suggest that a more extensive use of machine learning approaches could significantly accelerate the design of novel glasses with tailored properties.

Although molecular dynamics (MD) simulations are commonly used to predict the structure and properties of glasses, they are intrinsically limited to short time scales, necessitating the use of fast cooling rates. It is therefore challenging to compare results from MD simulations to experimental results for glasses cooled on typical laboratory time scales. Based on MD simulations of a sodium silicate glass with varying cooling rate (from 0.01 to 100 K/ps), here we show that thermal history primarily affects the medium-range order structure, while the short-range order is largely unaffected over the range of cooling rates simulated. This results in a decoupling between the enthalpy and volume relaxation functions, where the enthalpy quickly plateaus as the cooling rate decreases, whereas density exhibits a slower relaxation. Finally, we demonstrate that the outcomes of MD simulations can be meaningfully compared to experimental values if properly extrapolated to slower cooling rates.

Upon loading, atomic networks can feature delayed viscoplastic relaxation. However, the effect of composition and structure on such a relaxation remains poorly understood. Herein, relying on accelerated molecular dynamics simulations and topological constraint theory, we investigate the relationship between atomic topology and stress-induced relaxation, by taking the example of creep deformations in calcium--silicate--hydrates, the binding phase of concrete. Under constant shear stress, C--S--H is found to feature delayed logarithmic shear deformations. We demonstrate that the propensity for relaxation is minimum for isostatic atomic networks, which are characterized by the simultaneous absence of floppy internal modes of relaxation and eigen stress. This suggests that topological nano-engineering could lead to the discovery of non-aging materials.