Structural Prediction and Inverse Design by a Strongly Correlated Neural Network

Template Design for Complex Block Copolymer Patterns Using a Machine Learning Method

This study represents the first attempt to address the inverse design problem of the guiding template for directed self-assembly (DSA) patterns using solely machine learning methods. By formulating the problem as a multi-label classification task, the study shows that it is possible to predict templates without requiring any forward simulations. A series of neural network (NN) models, ranging from the basic two-layer convolutional neural network (CNN) to the large NN models (32-layer CNN with 8 residual blocks), have been trained using simulated pattern samples generated by thousands of self-consistent field theory (SCFT) calculations; a number of augmentation techniques, especially suitable for predicting morphologies, have been also proposed to enhance the performance of the NN model. The exact match accuracy of the model in predicting the template of simulated patterns was significantly improved from 59.8% for the baseline model to 97.1% for the best model of this study. The best model also demonstrates an excellent generalization ability in predicting the template for human-designed DSA patterns, while the simplest baseline model is ineffective in this task.

Variational methods and deep Ritz method for active elastic solids

Variational methods have been widely used in soft matter physics for both static and dynamic problems. These methods are mostly based on two variational principles: the variational principle of minimum free energy (MFEVP) and Onsager's variational principle (OVP). Our interests lie in the applications of these variational methods to active matter physics. In our former work [H. Wang, T. Qian and X. Xu, Soft Matter, 2021, 17, 3634-3653], we have explored the applications of OVP-based variational methods for the modeling of active matter dynamics. In the present work, we explore variational (or energy) methods that are based on MFEVP for static problems in active elastic solids. We show that MFEVP can be used not only to derive equilibrium equations, but also to develop approximate solution methods, such as the Ritz method, for active solid statics. Moreover, the power of the Ritz-type method can be further enhanced using deep learning methods if we use deep neural networks to construct the trial functions of the variational problems. We then apply these variational methods and the deep Ritz method to study the spontaneous bending and contraction of a thin active circular plate that is induced by internal asymmetric active contraction. The circular plate is found to be bent towards its contracting side. The study of such a simple toy system gives implications for understanding the morphogenesis of solid-like confluent cell monolayers. In addition, we introduce a so-called activogravity length to characterize the importance of gravitational forces relative to internal active contraction in driving the bending of the active plate. When the lateral plate dimension is larger than the activogravity length (about 100 micron), gravitational forces become important. Such gravitaxis behaviors at multicellular scales may play significant roles in the morphogenesis and in the up-down symmetry broken during tissue development.

Autonomous Construction of Phase Diagrams of Block Copolymers by Theory-Assisted Active Machine Learning

Equilibrium phase diagrams serve as blueprints for rational design of nanostructured materials of block copolymers, but their construction is time-consuming and requires profound expertise. Herein, by virtue of the knowledge of self-consistent field theory (SCFT), the active-learning method is developed to autonomously construct the phase diagrams of block copolymers. Without human intervention, the SCFT-assisted active-learning method can rapidly search the undetected phases and efficiently reproduce the complicated phase diagrams of diblock copolymers and multiblock terpolymers via decreasing the number of sampling points to about 20%. It is clearly demonstrated that the combined uncertainty sampling/random selection scheme in the active-learning method shows the outperformance in spite of a small amount of initial data set. This work highlights the promising integration of theoretical modeling with machine learning and represents a crucial step toward rational design of nanostructured materials.

Artificial neural network discovery of a switchable metasurface reflector

Optical materials engineered to dynamically and selectively manipulate electromagnetic waves are essential to the future of modern optical systems. In this paper, we simulate various metasurface configurations consisting of periodic 1D bars or 2D pillars made of the ternary phase change material Ge₂Sb₂Te₅ (GST). Dynamic switching behavior in reflectance is exploited due to a drastic refractive index change between the crystalline and amorphous states of GST. Selectivity in the reflection and transmission spectra is manipulated by tailoring the geometrical parameters of the metasurface. Due to the immense number of possible metasurface configurations, we train deep neural networks capable of exploring all possible designs within the working parameter space. The data requirements, predictive accuracy, and robustness of these neural networks are benchmarked against a ground truth by varying quality and quantity of training data. After ensuring trustworthy neural network advisory, we identify and validate optimal GST metasurface configurations best suited as dynamic switchable mirrors depending on selected light and manufacturing constraints.