Machine learning without a processor: Emergent learning in a nonlinear analog network.

Standard deep learning algorithms require differentiating large nonlinear networks, a process that is slow and power-hungry. Electronic contrastive local learning networks (CLLNs) offer potentially fast, efficient, and fault-tolerant hardware for analog machine learning, but existing implementations are linear, severely limiting their capabilities. These systems differ significantly from artificial neural networks as well as the brain, so the feasibility and utility of incorporating nonlinear elements have not been explored. Here, we introduce a nonlinear CLLN-an analog electronic network made of self-adjusting nonlinear resistive elements based on transistors. We demonstrate that the system learns tasks unachievable in linear systems, including XOR (exclusive or) and nonlinear regression, without a computer. We find our decentralized system reduces modes of training error in order (mean, slope, curvature), similar to spectral bias in artificial neural networks. The circuitry is robust to damage, retrainable in seconds, and performs learned tasks in microseconds while dissipating only picojoules of energy across each transistor. This suggests enormous potential for fast, low-power computing in edge systems like sensors, robotic controllers, and medical devices, as well as manufacturability at scale for performing and studying emergent learning.

Identifying microscopic factors that influence ductility in disordered solids.

There are empirical strategies for tuning the degree of strain localization in disordered solids, but they are system-specific and no theoretical framework explains their effectiveness or limitations. Here, we study three model disordered solids: a simulated atomic glass, an experimental granular packing, and a simulated polymer glass. We tune each system using a different strategy to exhibit two different degrees of strain localization. In tandem, we construct structuro-elastoplastic (StEP) models, which reduce descriptions of the systems to a few microscopic features that control strain localization, using a machine learning-based descriptor, softness, to represent the stability of the disordered local structure. The models are based on calculated correlations of softness and rearrangements. Without additional parameters, the models exhibit semiquantitative agreement with observed stress-strain curves and softness statistics for all systems studied. Moreover, the StEP models reveal that initial structure, the near-field effect of rearrangements on local structure, and rearrangement size, respectively, are responsible for the changes in ductility observed in the three systems. Thus, StEP models provide microscopic understanding of how strain localization depends on the interplay of structure, plasticity, and elasticity.

Hierarchical bubble size distributions in coarsening wet liquid foams.

Coarsening of two-phase systems is crucial for the stability of dense particle packings such as alloys, foams, emulsions, or supersaturated solutions. Mean field theories predict an asymptotic scaling state with a broad particle size distribution. Aqueous foams are good model systems for investigations of coarsening-induced structures, because the continuous liquid as well as the dispersed gas phases are uniform and isotropic. We present coarsening experiments on wet foams, with liquid fractions up to their unjamming point and beyond, that are performed under microgravity to avoid gravitational drainage. As time elapses, a self-similar regime is reached where the normalized bubble size distribution is invariant. Unexpectedly, the distribution features an excess of small roaming bubbles, mobile within the network of jammed larger bubbles. These roaming bubbles are reminiscent of rattlers in granular materials (grains not subjected to contact forces). We identify a critical liquid fraction [Formula: see text], above which the bubble assembly unjams and the two bubble populations merge into a single narrow distribution of bubbly liquids. Unexpectedly, [Formula: see text] is larger than the random close packing fraction of the foam [Formula: see text]. This is because, between [Formula: see text] and [Formula: see text], the large bubbles remain connected due to a weak adhesion between bubbles. We present models that identify the physical mechanisms explaining our observations. We propose a new comprehensive view of the coarsening phenomenon in wet foams. Our results should be applicable to other phase-separating systems and they may also help to control the elaboration of solid foams with hierarchical structures.

Twist and measure: characterizing the effective radius of strings and bundles under twisting contraction.

We test the standard model for the length contraction of a bundle of strings under twist, and find deviation that is significantly greater than typically appreciated and that has a different nature at medium and large twist angles. By including volume conservation, we achieve better fits to data for single-, double-, and triple-stranded bundles of nylon monofilament as an ideal test case. This gives a well-defined procedure for extracting an effective twist radius that characterizes contraction behavior. While our approach accounts for the observed faster-than-expected contraction up to medium twist angles, we also find that the contraction is nevertheless slower than expected at large twist angles for both nylon monofilament bundles and several other string types. The size of this effect varies with the individual-string braid structure and with the number of strings in the bundle. We speculate that it may be related to elastic deformation within the material. However, our first modeling attempt does not fully capture the observed behavior.

Probing Gardner Physics in an Active Quasithermal Pressure-Controlled Granular System of Noncircular Particles.

To search for experimental signals of the Gardner crossover, an active quasithermal granular glass is constructed using a monolayer of air-fluidized star-shaped particles. The pressure of the system is controlled by adjusting the tension exerted on an enclosing boundary. Velocity distributions of the internal particles and the scaling of the pressure, density, effective temperature, and relaxation time are examined, demonstrating that the system has key features of a thermal system. Using a pressure-based quenching protocol that brings the system into deeper glassy states, signals of the Gardner crossover are detected via cage size and separation order parameters for both particle positions and orientations, offering experimental evidence of Gardner physics for a system of anisotropic quasithermal particles in a low spatial dimension.

Strain localization and failure of disordered particle rafts with tunable ductility during tensile deformation.

Quasi-static tensile experiments were performed for a model disordered solid consisting of a two-dimensional raft of polydisperse floating granular particles with capillary attractions. The ductility is tuned by controlling the capillary interaction range, which varies with the particle size. During the tensile tests, after an initial period of elastic deformation, strain localization occurs and leads to the formation of a shear band at which the pillar later fails. In this process, small particles with long-ranged interactions can endure large plastic deformation without forming significant voids, while large particles with short-range interactions fail dramatically by fracturing at small deformation. Particle-level structure was measured, and the strain-localized region was found to have higher structural anisotropy than the bulk. Local interactions between anisotropic sites and particle rearrangements were the main mechanisms driving strain localization and the subsequent failure, and significant differences of such interactions exist between ductile and brittle behaviors.

Dynamics and thermodynamics of air-driven active spinners.

We report on the collective behavior of active particles in which energy is continuously supplied to rotational degrees of freedom. The active spinners are 3D-printed disks, 1 cm in diameter, that have an embedded fan-like structure, such that a sub-levitating up-flow of air forces them to spin. Single spinners exhibit Brownian motion with a narrow Gaussian velocity distribution function, P(v), for translational motion. We study the evolution of P(v) as the packing fraction and the average single particle spin speeds are varied. The interparticle hydrodynamic interaction is negligible, and the dynamics is dominated by hyperelastic collisions and dissipative forces. As expected for nonequilibrium systems, P(v) for a collection of many spinners deviates from Gaussian behavior. However, unlike translationally active systems, phase separation is not observed, and the system remains spatially homogeneous. We then search for a near-equilibrium counterpart for our active spinners by measuring the equation of state. Interestingly, it agrees well with a hard-sphere model, despite the dissipative nature of the single particle dynamics.

Friction controls even submerged granular flows.

We investigate the coupling between the interstitial medium and granular particles by studying the hopper flow of dry and submerged systems experimentally and numerically. In accordance with earlier studies, we find that the dry hopper empties at a constant rate. However, in the submerged system we observe the surging of the flow rate. We model both systems using the discrete element method, which we couple with computational fluid dynamics in the case of a submerged hopper. We are able to match the simulations and the experiments with good accuracy by fitting the particle-particle contact friction for each system separately. Submerging the hopper changes the particle-particle contact friction from µvacuum = 0.15 to µsub = 0.13, while all the other simulation parameters remain the same.

Divergence of Voronoi Cell Anisotropy Vector: A Threshold-Free Characterization of Local Structure in Amorphous Materials.

Characterizing structural inhomogeneity is an essential step in understanding the mechanical response of amorphous materials. We introduce a threshold-free measure based on the field of vectors pointing from the center of each particle to the centroid of the Voronoi cell in which the particle resides. These vectors tend to point in toward regions of high free volume and away from regions of low free volume, reminiscent of sinks and sources in a vector field. We compute the local divergence of these vectors, where positive values correspond to overpacked regions and negative values identify underpacked regions within the material. Distributions of this divergence are nearly Gaussian with zero mean, allowing for structural characterization using only the moments of the distribution. We explore how the standard deviation and skewness vary with the packing fraction for simulations of bidisperse systems and find a kink in these moments that coincides with the jamming transition.

Bellybutton: accessible and customizable deep-learning image segmentation.

The conversion of raw images into quantifiable data can be a major hurdle and time-sink in experimental research, and typically involves identifying region(s) of interest, a process known as segmentation. Machine learning tools for image segmentation are often specific to a set of tasks, such as tracking cells, or require substantial compute or coding knowledge to train and use. Here we introduce an easy-to-use (no coding required), image segmentation method, using a 15-layer convolutional neural network that can be trained on a laptop: Bellybutton. The algorithm trains on user-provided segmentation of example images, but, as we show, just one or even a sub-selection of one training image can be sufficient in some cases. We detail the machine learning method and give three use cases where Bellybutton correctly segments images despite substantial lighting, shape, size, focus, and/or structure variation across the regions(s) of interest. Instructions for easy download and use, with further details and the datasets used in this paper are available at pypi.org/project/Bellybuttonseg .

Simply solvable model capturing the approach to statistical self-similarity for the diffusive coarsening of bubbles, droplets, and grains.

Aqueous foams and a wide range of related systems are believed to coarsen by diffusion between neighboring domains into a statistically self-similar scaling state, after the decay of initial transients, such that dimensionless domain size and shape distributions become time independent and the average grows as a power law. Partial integrodifferential equations for the time evolution of the size distribution for such phase separating systems can be formulated for arbitrary initial conditions, but these are cumbersome for analyzing data on nonscaling state preparations. Here we show that essential features of the approach to the scaling state are captured by an exactly-solvable ordinary differential equation for the evolution of the average bubble size. The key ingredient is to characterize the bubble size distribution approximately, using the average size of all bubbles and the average size of the critical bubbles, which instantaneously neither grow nor shrink. The difference between these two averages serves as a proxy for the width of the size distribution. Solution of our model shows that behavior is controlled by a signed length δ that is proportional to the width of the initial distribution relative to that in the scaling state. In particular, δ is negative if the initial preparation is too monodisperse, and is positive if it is too polydisperse. To test our approach, we compare with data for quasi-two dimensional dry foams created with three different initial amounts of polydispersity. This allows us to readily identify the critical radius from the average area of six-sided bubbles, whose growth rate is zero by the von Neumann law. The growth of the average and critical radii agree quite well with exact solution, though the most monodisperse sample crosses over to the scaling state faster than expected. A simpler approximate solution of our model performs equally well. Our approach is applicable to 3d foams, which we demonstrate by re-analyzing prior data, as well as to froths of dilute droplets and to phase separation kinetics for more general systems such as emulsions, binary mixtures, and alloys.

Machine learning characterization of structural defects in amorphous packings of dimers and ellipses.

Structural defects within amorphous packings of symmetric particles can be characterized using a machine learning approach that incorporates structure functions of radial distances and angular arrangement. This yields a scalar field, softness, that correlates with the probability that a particle is about to be rearranged. However, when particle shapes are elongated, as in the case of dimers and ellipses, we find that the standard structure functions produce imprecise softness measurements. Moreover, ellipses exhibit deformation profiles in stark contrast to circular particles. In order to account for the effects of orientation and alignment, we introduce structure functions to recover the predictive performance of softness, as well as provide physical insight into local and extended dynamics. We study a model disordered solid, a bidisperse two-dimensional granular pillar, driven by uniaxial compression and composed entirely of monomers, dimers, or ellipses. We demonstrate how the computation of softness via a support vector machine extends to dimers and ellipses with the introduction of orientational structure functions. Then we highlight the spatial extent of rearrangements and defects, as well as their cross correlation, for each particle shape. Finally, we demonstrate how an additional machine learning algorithm, recursive feature elimination, provides an avenue to better understand how softness arises from particular structural aspects. We identify the most crucial structure functions in determining softness and discuss their physical implications.

Anisotropic particles strengthen granular pillars under compression.

We probe the effects of particle shape on the global and local behavior of a two-dimensional granular pillar, acting as a proxy for a disordered solid, under uniaxial compression. This geometry allows for direct measurement of global material response, as well as tracking of all individual particle trajectories. In general, drawing connections between local structure and local dynamics can be challenging in amorphous materials due to lower precision of atomic positions, so this study aims to elucidate such connections. We vary local interactions by using three different particle shapes: discrete circular grains (monomers), pairs of grains bonded together (dimers), and groups of three bonded in a triangle (trimers). We find that dimers substantially strengthen the pillar and the degree of this effect is determined by orientational order in the initial condition. In addition, while the three particle shapes form void regions at distinct rates, we find that anisotropies in the local amorphous structure remain robust through the definition of a metric that quantifies packing anisotropy. Finally, we highlight connections between local deformation rates and local structure.

The sands of time run faster near the end.

Grains exiting an underwater silo exhibit an unexpected surge in discharge rate as they empty. This contrasts with the constant flow rate of dry granular hoppers and the decreasing flow rate of pure liquids. Here we find that this surge depends on hopper diameter and happens also in air. The surge can be turned off by fixing the rate of fluid flow through the granular packing. With no flow control, dye injected on top of the packing gets drawn into the grains. We conclude that the surge is caused by a self-generated pumping of fluid through the packing. The effect is modelled via a driving pressure set by the exit speed of the grains. This highlights a surprising and unrecognized role that interstitial fluid plays in setting the discharge rate, and perhaps in controlling clog formation, for granular hoppers whether in air or under water.

Effect of interstitial fluid on the fraction of flow microstates that precede clogging in granular hoppers.

We report on the nature of flow events for the gravity-driven discharge of glass beads through a hole that is small enough that the hopper is susceptible to clogging. In particular, we measure the average and standard deviation of the distribution of discharged masses as a function of both hole and grain sizes. We do so in air, which is usual, but also with the system entirely submerged under water. This damps the grain dynamics and could be expected to dramatically affect the distribution of the flow events, which are described in prior work as avalanche-like. Though the flow is slower and the events last longer, we find that the average discharge mass is only slightly reduced for submerged grains. Furthermore, we find that the shape of the distribution remains exponential, implying that clogging is still a Poisson process even for immersed grains. Per Thomas and Durian [Phys. Rev. Lett. 114, 178001 (2015)PRLTAO0031-900710.1103/PhysRevLett.114.178001], this allows for an interpretation of the average discharge mass in terms of the fraction of flow microstates that precede, i.e., that effectively cause, a stable clog to form. Since this fraction is barely altered by water, we conclude that the crucial microscopic variables are the grain positions; grain momenta play only a secondary role in destabilizing weak incipient arches. These insights should aid ongoing efforts to understand the susceptibility of granular hoppers to clogging.

Onset of sediment transport is a continuous transition driven by fluid shear and granular creep.

Fluid-sheared granular transport sculpts landscapes and undermines infrastructure, yet predicting the onset of sediment transport remains notoriously unreliable. For almost a century, this onset has been treated as a discontinuous transition at which hydrodynamic forces overcome gravity-loaded grain-grain friction. Using a custom laminar-shear flume to image slow granular dynamics deep into the bed, here we find that the onset is instead a continuous transition from creeping to granular flow. This transition occurs inside the dense granular bed at a critical viscous number, similar to granular flows and colloidal suspensions and inconsistent with hydrodynamic frameworks. We propose a new phase diagram for sediment transport, where 'bed load' is a dense granular flow bounded by creep below and suspension above. Creep is characteristic of disordered solids and reminiscent of soil diffusion on hillslopes. Results provide new predictions for the onset and dynamics of sediment transport that challenge existing models.

Deformation-driven diffusion and plastic flow in amorphous granular pillars.

We report a combined experimental and simulation study of deformation-induced diffusion in compacted quasi-two-dimensional amorphous granular pillars, in which thermal fluctuations play a negligible role. The pillars, consisting of bidisperse cylindrical acetal plastic particles standing upright on a substrate, are deformed uniaxially and quasistatically by a rigid bar moving at a constant speed. The plastic flow and particle rearrangements in the pillars are characterized by computing the best-fit affine transformation strain and nonaffine displacement associated with each particle between two stages of deformation. The nonaffine displacement exhibits exponential crossover from ballistic to diffusive behavior with respect to the cumulative deviatoric strain, indicating that in athermal granular packings, the cumulative deviatoric strain plays the role of time in thermal systems and drives effective particle diffusion. We further study the size-dependent deformation of the granular pillars by simulation, and find that different-sized pillars follow self-similar shape evolution during deformation. In addition, the yield stress of the pillars increases linearly with pillar size. Formation of transient shear lines in the pillars during deformation becomes more evident as pillar size increases. The width of these elementary shear bands is about twice the diameter of a particle, and does not vary with pillar size.

Dynamics of normal and superfluid fogs using diffusing-wave spectroscopy.

The dynamics of normal and superfluid fogs are studied using the technique of diffusing-wave spectroscopy. For a water fog generated with a 1.75 MHz piezoelectric driver below the liquid surface, the 7 microm diameter droplets are found to have diffusive dynamics for correlation times long compared to the viscous time. For a fog of 10 microm diameter superfluid helium droplets in helium vapor at 1.5 K the motion appears to be ballistic for correlation times short compared to the viscous time. The velocity correlations between the helium droplets are found to depend on the initial velocity with which the droplets are injected from the helium surface into the fog.

Kinetics of gravity-driven water channels under steady rainfall.

We investigate the formation of fingered flow in dry granular media under simulated rainfall using a quasi-two-dimensional experimental setup composed of a random close packing of monodisperse glass beads. Using controlled experiments, we analyze the finger instabilities that develop from the wetting front as a function of fundamental granular (particle size) and fluid properties (rainfall, viscosity). These finger instabilities act as precursors for water channels, which serve as outlets for water drainage. We look into the characteristics of the homogeneous wetting front and channel size as well as estimate relevant time scales involved in the instability formation and the velocity of the channel fingertip. We compare our experimental results with that of the well-known prediction developed by Parlange and Hill [D. E. Hill and J. Y. Parlange, Soil Sci. Soc. Am. Proc. 36, 697 (1972)]. This model is based on linear stability analysis of the growth of perturbations arising at the interface between two immiscible fluids. Results show that, in terms of morphology, experiments agree with the proposed model. However, in terms of kinetics we nevertheless account for another term that describes the homogenization of the wetting front. This result shows that the manner we introduce the fluid to a porous medium can also influence the formation of finger instabilities. The results also help us to calculate the ideal flow rate needed for homogeneous distribution of water in the soil and minimization of runoff, given the grain size, fluid density, and fluid viscosity. This could have applications in optimizing use of irrigation water.

Drag force scaling for penetration into granular media.

Impact dynamics is measured for spherical and cylindrical projectiles of many different densities dropped onto a variety non-cohesive granular media. The results are analyzed in terms of the material-dependent scaling of the inertial and frictional drag contributions to the total stopping force. The inertial drag force scales similar to that in fluids, except that it depends on the internal friction coefficient. The frictional drag force scales as the square-root of the density of granular medium and projectile, and hence cannot be explained by the combination of granular hydrostatic pressure and Coulomb friction law. The combined results provide an explanation for the previously observed penetration depth scaling.