Inside the Black Box: A Physical Basis for the Effectiveness of Deep Generative Models of Amorphous Materials

2021 ◽  
pp. 110885
Author(s):  
Michael Kilgour ◽  
Lena Simine
Keyword(s):  
Amorphous Materials ◽  
Generative Models ◽  
Black Box ◽  
Physical Basis

scholarly journals Inside the Black Box: A Physical Basis for the Effectiveness of Deep Generative Models of Amorphous Materials

2021 ◽  
Author(s):  
Michael Kilgour ◽  
Lena Simine
Keyword(s):  
Machine Learning ◽  
Amorphous Materials ◽  
Physical Simulation ◽  
Generative Models ◽  
Black Box ◽  
Small Scale ◽  
Learning Approach ◽  
Computational Costs ◽  
Training Samples ◽  
Machine Learning Approach

<p>We have recently demonstrated an effective protocol for the simulation of amorphous molecular configurations using the PixelCNN generative model (J. Phys. Chem. Lett. 2020, 11, 20, 8532). The morphological sampling of amorphous materials via such an autoregressive generation protocol sidesteps the high computational costs associated with simulating amorphous materials at scale, enabling practically unlimited structural sampling based on only small-scale experimental or computational training samples. An important question raised but not rigorously addressed in that report was whether this machine learning approach could be considered a physical simulation in the conventional sense. Here we answer this question by detailing the inner workings of the underlying algorithm that we refer to as the Morphological Autoregression Protocol or MAP. <br></p>

scholarly journals Inside the Black Box: A Physical Basis for the Effectiveness of Deep Generative Models of Amorphous Materials

2021 ◽  
Author(s):  
Michael Kilgour ◽  
Lena Simine
Keyword(s):  
Machine Learning ◽  
Amorphous Materials ◽  
Physical Simulation ◽  
Generative Models ◽  
Black Box ◽  
Small Scale ◽  
Learning Approach ◽  
Computational Costs ◽  
Training Samples ◽  
Machine Learning Approach

<p>We have recently demonstrated an effective protocol for the simulation of amorphous molecular configurations using the PixelCNN generative model (J. Phys. Chem. Lett. 2020, 11, 20, 8532). The morphological sampling of amorphous materials via such an autoregressive generation protocol sidesteps the high computational costs associated with simulating amorphous materials at scale, enabling practically unlimited structural sampling based on only small-scale experimental or computational training samples. An important question raised but not rigorously addressed in that report was whether this machine learning approach could be considered a physical simulation in the conventional sense. Here we answer this question by detailing the inner workings of the underlying algorithm that we refer to as the Morphological Autoregression Protocol or MAP. <br></p>

scholarly journals LOGAN: Membership Inference Attacks Against Generative Models

2019 ◽  
Vol 2019 (1) ◽  
pp. 133-152 ◽  
Author(s):  
Jamie Hayes ◽  
Luca Melis ◽  
George Danezis ◽  
Emiliano De Cristofaro
Keyword(s):  
Diabetic Retinopathy ◽  
State Of The Art ◽  
Generative Models ◽  
Black Box ◽  
Mitigation Strategies ◽  
Generative Adversarial Networks ◽  
Underlying Distribution ◽  
Target Model ◽  
Adversarial Networks ◽  
Inference Attacks

Abstract Generative models estimate the underlying distribution of a dataset to generate realistic samples according to that distribution. In this paper, we present the first membership inference attacks against generative models: given a data point, the adversary determines whether or not it was used to train the model. Our attacks leverage Generative Adversarial Networks (GANs), which combine a discriminative and a generative model, to detect overfitting and recognize inputs that were part of training datasets, using the discriminator’s capacity to learn statistical differences in distributions. We present attacks based on both white-box and black-box access to the target model, against several state-of-the-art generative models, over datasets of complex representations of faces (LFW), objects (CIFAR-10), and medical images (Diabetic Retinopathy). We also discuss the sensitivity of the attacks to different training parameters, and their robustness against mitigation strategies, finding that defenses are either ineffective or lead to significantly worse performances of the generative models in terms of training stability and/or sample quality.

AEM analysis of crystallization in Ti-based amorphous alloys

1983 ◽  
Vol 41 ◽  
pp. 270-271
Author(s):  
A.R. Pelton ◽  
A.F. Marshall ◽  
Y.S. Lee
Keyword(s):  
Magnetic Properties ◽  
Amorphous Alloys ◽  
Amorphous Materials ◽  
Isothermal Annealing ◽  
Amorphous Matrix ◽  
Nucleation And Growth ◽  
Current Interest ◽  
Amorphous Films ◽  
Electrical And Magnetic Properties

Amorphous materials are of current interest due to their desirable mechanical, electrical and magnetic properties. Furthermore, crystallizing amorphous alloys provides an avenue for discerning sequential and competitive phases thus allowing access to otherwise inaccessible crystalline structures. Previous studies have shown the benefits of using AEM to determine crystal structures and compositions of partially crystallized alloys. The present paper will discuss the AEM characterization of crystallized Cu-Ti and Ni-Ti amorphous films.Cu60Ti40: The amorphous alloy Cu60Ti40, when continuously heated, forms a simple intermediate, macrocrystalline phase which then transforms to the ordered, equilibrium Cu3Ti2 phase. However, contrary to what one would expect from kinetic considerations, isothermal annealing below the isochronal crystallization temperature results in direct nucleation and growth of Cu3Ti2 from the amorphous matrix.

TEM studies of amorphization processes and amorphous structures

1988 ◽  
Vol 46 ◽  
pp. 448-449
Author(s):  
T. E. Mitchell ◽  
R. B. Schwarz
Keyword(s):  
Amorphous Materials ◽  
Fission Products ◽  
Rapid Quenching ◽  
Surface Films ◽  
List Type ◽  
Oxide Glasses ◽  
Crystalline Materials ◽  
Amorphous Structures ◽  
Amorphous Surface ◽  
Interfacial Films

Traditional oxide glasses occur naturally as obsidian and can be made easily by suitable cooling histories. In the past 30 years, a variety of techniques have been discovered which amorphize normally crystalline materials such as metals. These include [1-3]:Rapid quenching from the vapor phase.Rapid quenching from the liquid phase.Electrodeposition of certain alloys, e.g. Fe-P.Oxidation of crystals to produce amorphous surface oxide layers.Interdiffusion of two pure crystalline metals.Hydrogen-induced vitrification of an intermetal1ic.Mechanical alloying and ball-milling of intermetal lie compounds.Irradiation processes of all kinds using ions, electrons, neutrons, and fission products.We offer here some general comments on the use of TEM to study these materials and give some particular examples of such studies.Thin specimens can be prepared from bulk homogeneous materials in the usual way. Most often, however, amorphous materials are in the form of surface films or interfacial films with different chemistry from the substrates.

Analysis of Dark-Field Images of Disordered Materials

1972 ◽  
Vol 30 ◽  
pp. 560-561
Author(s):  
J. M. Cowley
Keyword(s):  
Amorphous Materials ◽  
Careful Analysis ◽  
Dark Field ◽  
Minimum Diameter ◽  
Statistical Fluctuations ◽  
Bright Spots ◽  
Sufficient Detail ◽  
Random Arrangement ◽  
Diffraction Patterns ◽  
Imaging Conditions

Recently a number of authors have reported detail in dark-field images obtained from diffuse-scattering regions of electron diffraction patterns. Bright spots in images from short-range order diffuse peaks of disordered binary alloys have been interpreted as evidence for the existence of microdomains of ordered lattice or of segragated clusters of one component. Spotty contrast in dark field images of near-amorphous materials has been interpreted as evidence for the existense of microcrystals. Without a careful analysis of the imaging conditions such conclusions may be invalid. Usually the conditions of the experiment have not been specified in sufficient detail to allow evaluation of the conclusions.Elementary considerations show that even for a completely random arrangement of atoms the statistical fluctuations of density will give a spotty contrast with spots of minimum diameter determined by the dark field aperture size and other factors influencing the minimum resolvable distance under darkfield imaging conditions, including fluctuations and drift over long exposure times (resolution usually 10Å or more).

Simulations of high-resolution images of amorphous silicon films

1986 ◽  
Vol 44 ◽  
pp. 544-545
Author(s):  
G. Y. Fan ◽  
J. M. Cowley
Keyword(s):  
Electron Microscope ◽  
High Frequency ◽  
Fourier Transformation ◽  
Amorphous Materials ◽  
Low Frequency ◽  
Chromatic Aberration ◽  
Structure Information ◽  
Frequency Components ◽  
High Resolution Images ◽  
Spatial Frequency Spectrum

It is well known that the structure information on the specimen is not always faithfully transferred through the electron microscope. Firstly, the spatial frequency spectrum is modulated by the transfer function (TF) at the focal plane. Secondly, the spectrum suffers high frequency cut-off by the aperture (or effectively damping terms such as chromatic aberration). While these do not have essential effect on imaging crystal periodicity as long as the low order Bragg spots are inside the aperture, although the contrast may be reversed, they may change the appearance of images of amorphous materials completely. Because the spectrum of amorphous materials is continuous, modulation of it emphasizes some components while weakening others. Especially the cut-off of high frequency components, which contribute to amorphous image just as strongly as low frequency components can have a fundamental effect. This can be illustrated through computer simulation. Imaging of a whitenoise object with an electron microscope without TF limitation gives Fig. 1a, which is obtained by Fourier transformation of a constant amplitude combined with random phases generated by computer.

Multislice calculations for quasi-periodic and non-periodic objects

1990 ◽  
Vol 48 (4) ◽  
pp. 138-139
Author(s):  
R. Herrera ◽  
A. Gómez
Keyword(s):  
Electron Diffraction ◽  
Computer Simulations ◽  
Periodic Table ◽  
Amorphous Materials ◽  
Space Group ◽  
Essential Step ◽  
Shape Effects ◽  
Atomic Scattering ◽  
Diffraction Patterns ◽  
Periodic Continuation

Computer simulations of electron diffraction patterns and images are an essential step in the process of structure and/or defect elucidation. So far most programs are designed to deal specifically with crystals, requiring frequently the space group as imput parameter. In such programs the deviations from perfect periodicity are dealt with by means of “periodic continuation”.However, for many applications involving amorphous materials, quasiperiodic materials or simply crystals with defects (including finite shape effects) it is convenient to have an algorithm capable of handling non-periodicity. Our program “HeGo” is an implementation of the well known multislice equations in which no periodicity assumption is made whatsoever. The salient features of our implementation are: 1) We made Gaussian fits to the atomic scattering factors for electrons covering the whole periodic table and the ranges [0-2]Å−1 and [2-6]Å−1.

Sign in / Sign up

Export Citation Format

Share Document