Physics

An unconventional magnet may be mapped onto a simple ferromagnet by the existence of a high-symmetry point. Knowledge of conventional ferromagnetic systems may then be carried over to provide insight into more complex orders. Here we demonstrate how an unsupervised and interpretable machine-learning approach can be used to search for potential high-symmetry points in unconventional magnets without any prior knowledge of the system. The method is applied to the Heisenberg-Kitaev model on a honeycomb lattice, where our machine learns the transformations that manifest its hidden O(3) symmetry, without using data of these high-symmetry points. Moreover, we clarify that, in contrast to the stripy and zigzag orders, a set of D 2 and D 2h ordering matrices provides a more complete description of the magnetization in the Heisenberg-Kitaev model. In addition, our machine also learns the local constraints at the phase boundaries, which manifest a subdimensional symmetry. This work highlights the importance of explicit order parameters to many-body spin systems and the property of interpretability for the physical application of machine-learning techniques.

Cellular solids are known to exhibit size effects, i.e., differences in the apparent effective elastic moduli, when the specimen size becomes comparable to the cell size. The present contribution employs direct numerical simulations (DNS) of the mesostructure to investigate the influences of porosity, shape of pores, and thus material distribution along the struts, and orientation of loading on the size effects and effective moduli of regular honeycomb structures. Beam models are compared to continuum models for simple shear, uniaxial loading and pure bending of strips of finite width. It is found that the honeycomb structure exhibits a considerable anisotropy of the size effects and that honeycomb structures with circular pores exhibit considerably stronger size effects than those with hexagonal pores (and thus straight struts). Positive (stiffening) size effects are observed under simple shear and negative (softening) size effects under bending and uniaxial loading. The negative size effects are interpreted in terms of the stress-gradient theory.

Machine learning is poised as a very powerful tool that can drastically improve our ability to carry out scientific research. However, many issues need to be addressed before this becomes a reality. This article focuses on one particular issue of broad interest: How can we integrate machine learning with physics-based modeling to develop new interpretable and truly reliable physical models? After introducing the general guidelines, we discuss the two most important issues for developing machine learning-based physical models: Imposing physical constraints and obtaining optimal datasets. We also provide a simple and intuitive explanation for the fundamental reasons behind the success of modern machine learning, as well as an introduction to the concurrent machine learning framework needed for integrating machine learning with physics-based modeling. Molecular dynamics and moment closure of kinetic equations are used as examples to illustrate the main issues discussed. We end with a general discussion on where this integration will lead us to, and where the new frontier will be after machine learning is successfully integrated into scientific modeling.

Many complex multiphysics systems in fluid dynamics involve using solvers with varied levels of approximations in different regions of the computational domain to resolve multiple spatiotemporal scales present in the flow. The accuracy of the solution is governed by how the information is exchanged between these solvers at the interface and several methods have been devised for such coupling problems. In this article, we construct a data-driven model by spatially coupling a microscale lattice Boltzmann method (LBM) solver and macroscale finite difference method (FDM) solver for reaction-diffusion systems. The coupling between the micro-macro solvers has one to many mapping at the interface leading to the interface closure problem, and we propose a statistical inference method based on neural networks to learn this closure relation. The performance of the proposed framework in a bifidelity setting partitioned between the FDM and LBM domain shows its promise for complex systems where analytical relations between micro-macro solvers are not available.

Complex natural or engineered systems comprise multiple characteristic scales, multiple spatiotemporal domains, and even multiple physical closure laws. To address such challenges, we introduce an interface learning paradigm and put forth a data-driven closure approach based on memory embedding to provide physically correct boundary conditions at the interface. To enable the interface learning for hyperbolic systems by considering the domain of influence and wave structures into account, we put forth the concept of upwind learning towards a physics-informed domain decomposition. The promise of the proposed approach is shown for a set of canonical illustrative problems. We highlight that high-performance computing environments can benefit from this methodology to reduce communication costs among processing units in emerging machine learning ready heterogeneous platforms toward exascale era.

Volume of fluid(VOF) method is a sharp interface method employed for simulations of two phase flows. Interface in VOF is usually represented using piecewise linear line segments in each computational grid based on the volume fraction field. While VOF for cartesian coordinates conserve mass exactly, existing algorithms do not show machine-precision mass conservation for axisymmetric coordinate systems. In this work, we propose analytic formulae for interface reconstruction in axisymmetric coordinates, similar to those proposed by Scardovelli and Zaleski (J. Comput. Phys. 2000) for cartesian coordinates. We also propose modifications to the existing advection schemes in VOF for axisymmetric coordinates to obtain higher accuracy in mass conservation

This is an introductory machine learning course specifically developed with STEM students in mind. We discuss supervised, unsupervised, and reinforcement learning. The notes start with an exposition of machine learning methods without neural networks, such as principle component analysis, t-SNE, and linear regression. We continue with an introduction to both basic and advanced neural network structures such as conventional neural networks, (variational) autoencoders, generative adversarial networks, restricted Boltzmann machines, and recurrent neural networks. Questions of interpretability are discussed using the examples of dreaming and adversarial attacks.

Inverse design has greatly expanded nanophotonic devices and brought optimized performance. However, the use of inverse design for plasmonic structures has been challenging due to local field concentrations that can lead to errors in gradient calculation when the continuum adjoint method is used. On the other hand, with the discrete adjoint method one can achieve the exact gradient. Historically the discrete version is exclusively used with a Finite Element model, and applying the Finite-Difference Time-Domain (FDTD) method in inverse design of plasmonic structures is rarely attempted. Due to the popularity of using FDTD in simulating plasmonic structures, we integrate the discrete adjoint method with FDTD and present a framework to carry out inverse design of plasmonic structures using density-based topology optimization. We demonstrate the exactness of the gradient calculation for a plasmonic block structure with varying permittivity. Another challenge that is unique with plasmonic structures is that non-physical amplification caused by poorly chosen material interpolation can destroy a stable convergence of the optimization. To avoid this, we adopt a non-linear material interpolation scheme in the FDTD solver. In addition, filtering-and-projection regularization is incorporated to ensure manufacturability of the designed structures. As an example of this framework, successful reconstruction of electric fields of a plasmonic bowtie aperture is presented.

We present two algorithms by which a set of short, unbiased trajectories can be iteratively reweighted to obtain various observables. The first algorithm estimates the stationary (steady state) distribution of a system by iteratively reweighting the trajectories based on the average probability in each state. The algorithm applies to equilibrium or non-equilibrium steady states, exploiting the `left' stationarity of the distribution under dynamics -- i.e., in a discrete setting, when the column vector of probabilities is multiplied by the transition matrix expressed as a left stochastic matrix. The second procedure relies on the `right' stationarity of the committor (splitting probability) expressed as a row vector. The algorithms are unbiased, do not rely on computing transition matrices, and make no Markov assumption about discretized states. Here, we apply the procedures to a one-dimensional double-well potential, and to a 208 μ s atomistic Trp-cage folding trajectory from D.E. Shaw Research.

A Kohn-Sham (KS) inversion determines a KS potential and orbitals corresponding to a given electron density, a procedure that has applications in developing and evaluating functionals used in density functional theory. Despite the utility of KS inversions, application of these methods among the research community is disproportionately small. We implement the KS inversion methods of Zhao-Morrison-Parr and Wu-Yang in a framework that simplifies analysis and conversion of the resulting potential in real-space. Fully documented Python scripts integrate with PySCF, a popular electronic structure prediction software, and Fortran alternatives are provided for computational hot spots.