Deep learning for automated classification and characterization of amorphous materials

Machine learning for structure-property mapping of Ising models: Scalability and limitations

We present a scalable machine learning (ML) framework for predicting intensive properties and particularly classifying phases of Ising models. Scalability and transferability are central to the unprecedented computational efficiency of ML methods. In general, linear-scaling computation can be achieved through the divide-and-conquer approach, and the locality of physical properties is key to partitioning the system into subdomains that can be solved separately. Based on the locality assumption, ML model is developed for the prediction of intensive properties of a finite-size block. Predictions of large-scale systems can then be obtained by averaging results of the ML model from randomly sampled blocks of the system. We show that the applicability of this approach depends on whether the block-size of the ML model is greater than the characteristic length scale of the system. In particular, in the case of phase identification across a critical point, the accuracy of the ML prediction is limited by the diverging correlation length. We obtain an intriguing scaling relation between the prediction accuracy and the ratio of ML block size over the spin-spin correlation length. Implications for practical applications are also discussed. While the two-dimensional Ising model is used to demonstrate the proposed approach, the ML framework can be generalized to other many-body or condensed-matter systems.

Machine learning predicts the glass transition of two-dimensional colloids besides medium-range crystalline order

We employ only the positions of colloidal particles and construct machine learning (ML) models to test the presence of structural order in glass transition for two kinds of two-dimensional (2D) colloids: 2D polydisperse colloids (PC) with medium-range crystalline order (MRCO) and 2D binary colloids (BC) without MRCO. ML models predict the glass transition of 2D colloids successfully without any information on MRCO. Even certain ML models trained with BC predict the glass transition of PC successfully, thus suggesting that universal structural characteristics would exist besides MRCO.

Quantum Graph Neural Network Models for Materials Search

Inspired by classical graph neural networks, we discuss a novel quantum graph neural network (QGNN) model to predict the chemical and physical properties of molecules and materials. QGNNs were investigated to predict the energy gap between the highest occupied and lowest unoccupied molecular orbitals of small organic molecules. The models utilize the equivariantly diagonalizable unitary quantum graph circuit (EDU-QGC) framework to allow discrete link features and minimize quantum circuit embedding. The results show QGNNs can achieve lower test loss compared to classical models if a similar number of trainable variables are used, and converge faster in training. This paper also provides a review of classical graph neural network models for materials research and various QGNNs.

Correlations between defect propensity and dynamical heterogeneities in supercooled water

A salient feature of supercooled liquids consists in the dramatic dynamical slowdown they undergo as temperature decreases while no significant structural change is evident. These systems also present dynamical heterogeneities (DH): certain molecules, spatially arranged in clusters, relax various orders of magnitude faster than the others. However, again, no static quantity (such as structural or energetic measures) shows strong direct correlations with such fast-moving molecules. In turn, the dynamic propensity approach, an indirect measure that quantifies the tendency of the molecules to move in a given structural configuration, has revealed that dynamical constraints, indeed, originate from the initial structure. Nevertheless, this approach is not able to elicit which structural quantity is, in fact, responsible for such a behavior. In an effort to remove dynamics from its definition in favor of a static quantity, an energy-based propensity has also been developed for supercooled water, but it could only find positive correlations between the lowest-energy and the least-mobile molecules, while no correlations could be found for those more relevant mobile molecules involved in the DH clusters responsible for the system's structural relaxation. Thus, in this work, we shall define a defect propensity measure based on a recently introduced structural index that accurately characterizes water structural defects. We shall show that this defect propensity measure provides positive correlations with dynamic propensity, being also able to account for the fast-moving molecules responsible for the structural relaxation. Moreover, time dependent correlations will show that defect propensity represents an appropriate early-time predictor of the long-time dynamical heterogeneity.

Which molecular properties determine the impact sensitivity of an explosive? A machine learning quantitative investigation of nitroaromatic explosives

We decomposed density functional theory charge densities of 53 nitroaromatic molecules into atom-centered electric multipoles using the distributed multipole analysis that provides a detailed picture of the molecular electronic structure. Three electric multipoles, (the charge of the nitro groups), (the total dipole, i.e., polarization, of the nitro groups), (the total electron delocalization of the C ring atoms), and the number of explosophore groups (#NO₂) were selected as features for a comprehensive machine learning (ML) investigation. The target property was the impact sensitivity h₅₀ (cm) values quantified by drop-weight measurements, with a large h₅₀ (e.g., 150 cm) indicating that an explosive is insensitive and vice versa. After a preliminary screening of 42 ML algorithms, four were selected based on the lowest root mean square errors: Extra Trees, Random Forests, Gradient Boosting, and AdaBoost. Compared to experimental data, the predicted h₅₀ values of molecules having very different sensitivities for the four algorithms have differences in the range 19-28%. The most important properties for predicting h₅₀ are the electron delocalization in the ring atoms and the polarization of the nitro groups with averaged weights of 39% and 35%, followed by the charge (16%) and number (10%) of nitro groups. A significant result is how the contribution of these properties to h₅₀ depends on their actual sensitivities: for the most sensitive explosives (h₅₀ up to ∼50 cm), the four properties contribute to reducing h₅₀, and for intermediate ones (∼50 cm ≲ h₅₀ ≲ 100 cm) #NO₂ and contribute to increasing it and the other two properties to reducing it. For highly insensitive explosives (h₅₀ ≳ 200 cm), all four properties essentially contribute to increasing it. These results furnish a consistent molecular basis of the sensitivities of known explosives that also can be used for developing safer new ones.

Building Chemical Property Models for Energetic Materials from Small Datasets Using a Transfer Learning Approach

For many experimentally measured chemical properties that cannot be directly computed from first-principles, the existing physics-based models do not extrapolate well to out-of-sample molecules, and experimental datasets themselves are too small for traditional machine learning (ML) approaches. To overcome these limitations, we apply a transfer learning approach, whereby we simultaneously train a multi-target regression model on a small number of molecules with experimentally measured values and a large number of molecules with related computed properties. We demonstrate this methodology on predicting the experimentally measured impact sensitivity of energetic crystals, finding that both characteristics of the computed dataset and model architecture are important to prediction accuracy of the small experimental dataset. Our directed-message passing neural network (D-MPNN) ML model using transfer learning outperforms direct-ML and physics-based models on a diverse test set, and the new methods described here are widely applicable to modeling many other structure-property relationships.

Graph neural networks for materials science and chemistry

Machine learning plays an increasingly important role in many areas of chemistry and materials science, being used to predict materials properties, accelerate simulations, design new structures, and predict synthesis routes of new materials. Graph neural networks (GNNs) are one of the fastest growing classes of machine learning models. They are of particular relevance for chemistry and materials science, as they directly work on a graph or structural representation of molecules and materials and therefore have full access to all relevant information required to characterize materials. In this Review, we provide an overview of the basic principles of GNNs, widely used datasets, and state-of-the-art architectures, followed by a discussion of a wide range of recent applications of GNNs in chemistry and materials science, and concluding with a road-map for the further development and application of GNNs.

Machine Learning of Implicit Combinatorial Rules in Mechanical Metamaterials

Combinatorial problems arising in puzzles, origami, and (meta)material design have rare sets of solutions, which define complex and sharply delineated boundaries in configuration space. These boundaries are difficult to capture with conventional statistical and numerical methods. Here we show that convolutional neural networks can learn to recognize these boundaries for combinatorial mechanical metamaterials, down to finest detail, despite using heavily undersampled training sets, and can successfully generalize. This suggests that the network infers the underlying combinatorial rules from the sparse training set, opening up new possibilities for complex design of (meta)materials.

Predicting the failure of two-dimensional silica glasses

Being able to predict the failure of materials based on structural information is a fundamental issue with enormous practical and industrial relevance for the monitoring of devices and components. Thanks to recent advances in deep learning, accurate failure predictions are becoming possible even for strongly disordered solids, but the sheer number of parameters used in the process renders a physical interpretation of the results impossible. Here we address this issue and use machine learning methods to predict the failure of simulated two dimensional silica glasses from their initial undeformed structure. We then exploit Gradient-weighted Class Activation Mapping (Grad-CAM) to build attention maps associated with the predictions, and we demonstrate that these maps are amenable to physical interpretation in terms of topological defects and local potential energies. We show that our predictions can be transferred to samples with different shape or size than those used in training, as well as to experimental images. Our strategy illustrates how artificial neural networks trained with numerical simulation results can provide interpretable predictions of the behavior of experimentally measured structures.

The sheer number of parameters in deep learning makes the physical interpretation of failure predictions in glasses challenging. Here the authors use Grad-CAM to reveal the role of topological defects and local potential energies in failure predictions.

Collective Variables for Free Energy Surface Tailoring: Understanding and Modifying Functionality in Systems Dominated by Rare Events

We introduce a method for elucidating and modifying the functionality of systems dominated by rare events that relies on the semiautomated tuning of their underlying free energy surface. The proposed approach seeks to construct collective variables (CVs) that encode the essential information regarding the rare events of the system of interest. The appropriate CVs are identified using harmonic linear discriminant analysis (HLDA), a machine-learning-based method that is trained solely on data collected from short ordinary simulations in the relevant metastable states of the system. Utilizing the interpretable form of the resulting CVs, the critical interaction potentials that determine the system's rare transitions are identified and purposely modified to tailor the free energy surface in a manner that alters functionality as desired. The applicability of the method is illustrated in the context of three different systems, thereby demonstrating that thermodynamic and kinetic properties can be tractably modified with little to no prior knowledge or intuition.

Phase classification using neural networks: application to supercooled, polymorphic core-softened mixtures

Characterization of phases of soft matter systems is a challenge faced in many physical chemical problems. For polymorphic fluids it is an even greater challenge. Specifically, glass forming fluids, as water, can have, besides solid polymorphism, more than one liquid and glassy phases, and even a liquid-liquid critical point. In this sense, we apply a neural network algorithm to analyze the phase behavior of a mixture of core-softened fluids that interact through the continuous-shouldered well (CSW) potential, which have liquid polymorphism and liquid-liquid critical points, similar to water. We also apply the neural network to mixtures of CSW fluids and core-softened alcohols models. We combine and expand methods based on bond-orientational order parameters to study mixtures, applied to mixtures of hardcore fluids and to supercooled water, to include longer range coordination shells. With this, the trained neural network was able to properly predict the crystalline solid phases, the fluid phases and the amorphous phase for the pure CSW and CSW-alcohols mixtures with high efficiency. More than this, information about the phase populations, obtained from the network approach, can help verify if the phase transition is continuous or discontinuous, and also to interpret how the metastable amorphous region spreads along the stable high density fluid phase. These findings help to understand the behavior of supercooled polymorphic fluids and extend the comprehension of how amphiphilic solutes affect the phases behavior.

Machine learning for phase behavior in active matter systems

We demonstrate that deep learning techniques can be used to predict motility-induced phase separation (MIPS) in suspensions of active Brownian particles (ABPs) by creating a notion of phase at the particle level. Using a fully connected network in conjunction with a graph neural network we use individual particle features to predict to which phase a particle belongs. From this, we are able to compute the fraction of dilute particles to determine if the system is in the homogeneous dilute, dense, or coexistence region. Our predictions are compared against the MIPS binodal computed from simulation. The strong agreement between the two suggests that machine learning provides an effective way to determine the phase behavior of ABPs and could prove useful for determining more complex phase diagrams.

Characterising soft matter using machine learning

Machine learning is making a major impact in materials research. I review current progress across a selection of areas of ubiquitous soft matter. When applied to particle tracking, machine learning using convolution neural networks is providing impressive performance but there remain some significant problems to solve. Characterising ordered arrangements of particles is a huge challenge and machine learning has been deployed to create the description, perform the classification and tease out an interpretation using a wide array of techniques often with good success. In glass research, machine learning has proved decisive in quantifying very subtle correlations between the local structure around a site and the susceptibility towards a rearrangement event at that site. There are also beginning to be some impressive attempts to deploy machine learning in the design of composite soft materials. The discovery aspect of this new materials design meets the current interest in teaching algorithms to learn to extrapolate beyond the training data.

A neural network-based algorithm for high-throughput characterisation of viscoelastic properties of flowing microcapsules

Microcapsules, consisting of a liquid droplet enclosed by a viscoelastic membrane, have a wide range of biomedical and pharmaceutical applications and also serve as a popular mechanical model for biological cells. In this study, we develop a novel high throughput approach, by combining a machine learning method with a high-fidelity mechanistic capsule model, to accurately predict the membrane elasticity and viscosity of microcapsules from their dynamic deformation when flowing in a branched microchannel. The machine learning method consists of a deep convolutional neural network (DCNN) connected by a long short-term memory (LSTM) network. We demonstrate that with a superior prediction accuracy the present hybrid DCNN-LSTM network can still be faster than a conventional inverse method by five orders of magnitude, and can process thousands of capsules per second. We also show that the hybrid network has fewer restrictions compared with a simple DCNN.