Search

Human Activity Recognition (HAR) is critical for applications in healthcare, fitness, and IoT, but deploying accurate models on resource-constrained devices remains challenging due to high energy and memory demands. This paper demonstrates the application of Differentiable Weightless Neural Networks (DWNs) to HAR, achieving competitive accuracies of 96.34% and 96.67% while consuming only 56nJ and 104nJ per sample, with an inference time of just 5ns per sample. The DWNs were implemented and evaluated on an FPGA, showcasing their practical feasibility for energy-efficient hardware deployment. DWNs achieve up to 926,000x energy savings and 260x memory reduction compared to state-of-the-art deep learning methods. These results position DWNs as a nano-machine learning nanoML model for HAR, setting a new benchmark in energy efficiency and compactness for edge and wearable devices, paving the way for ultra-efficient edge AI., Comment: Accepted as a full paper by the 2025 EDGE AI FOUNDATION Austin

Scientific computing applications have benefited greatly from high performance computing infrastructure such as supercomputers. However, we are seeing a paradigm shift in the computational structure, design, and requirements of these applications. Increasingly, data-driven and machine learning approaches are being used to support, speed-up, and enhance scientific computing applications, especially molecular dynamics simulations. Concurrently, cloud computing platforms are increasingly appealing for scientific computing, providing "infinite" computing powers, easier programming and deployment models, and access to computing accelerators such as TPUs (Tensor Processing Units). This confluence of machine learning (ML) and cloud computing represents exciting opportunities for cloud and systems researchers. ML-assisted molecular dynamics simulations are a new class of workload, and exhibit unique computational patterns. These simulations present new challenges for low-cost and high-performance execution. We argue that transient cloud resources, such as low-cost preemptible cloud VMs, can be a viable platform for this new workload. Finally, we present some low-hanging fruits and long-term challenges in cloud resource management, and the integration of molecular dynamics simulations into ML platforms (such as TensorFlow)., Comment: 4 pages, position paper appearing in the Proceedings of the 2021 IEEE 14th International Conference on Cloud Computing (CLOUD)

Molecular dynamics simulations are powerful tools to extract the microscopic mechanisms characterizing the properties of soft materials. We recently introduced machine learning surrogates for molecular dynamics simulations of soft materials and demonstrated that artificial neural network based regression models can successfully predict the relationships between the input material attributes and the simulation outputs. Here, we show that statistical uncertainties associated with the outputs of molecular dynamics simulations can be utilized to train artificial neural networks and design machine learning surrogates with higher accuracy and generalizability. We design soft labels for the simulation outputs by incorporating the uncertainties in the estimated average output quantities, and introduce a modified loss function that leverages these soft labels during training to significantly reduce the surrogate prediction error for input systems in the unseen test data. The approach is illustrated with the design of a surrogate for molecular dynamics simulations of confined electrolytes to predict the complex relationship between the input electrolyte attributes and the output ionic structure. The surrogate predictions for the ionic density profiles show excellent agreement with the ground truth results produced using molecular dynamics simulations. The high accuracy and small inference times associated with the surrogate predictions provide quick access to quantities derived using the number density profiles and facilitate rapid sensitivity analysis., Comment: 14 pages, 8 figures

We use molecular dynamics simulations of the primitive model of electrolytes to study the ionic structure in aqueous monovalent electrolyte solutions confined by charged planar interfaces over a wide range of electrolyte concentration, interfacial separation, surface charge density, and ion size. The investigations are inspired by recent experiments that have directly measured the increase in the decay length for highly-concentrated electrolytes with increase in concentration. The behavior of ions in the nanoconfinement created by the interfaces is probed by evaluating the ionic density profiles, net charge densities, screening factors, and decay length associated with the screening of the charged interface. Results show the presence of two distinct regimes of screening behavior as the concentration is changed from 0.1 M to 2.5 M for a wide range of electrolyte systems generated by tuning the interfacial separation, surface charge density, and ionic size. For low concentrations, the screening factor exhibits a monotonic decay to 0 with a decay length that decreases sharply with increasing concentration. For high concentrations ($\gtrsim 1$ M), the screening factor has a non-monotonic behavior signaling charge inversion and formation of structured layers of ions near the interfaces. The decay length under these conditions rises with increasing concentration, exhibiting a power-law behavior. To complement the simulation results, a variational approach is developed that produces charge densities with characteristics consistent with those observed in simulations. The results demonstrate the relation between the rise in the strength of steric correlations and the changes in the screening behavior., Comment: 15 pages; 16 figures

We explore the idea of integrating machine learning (ML) with high performance computing (HPC)-driven simulations to address challenges in using simulations to teach computational science and engineering courses. We demonstrate that a ML surrogate, designed using artificial neural networks, yields predictions in excellent agreement with explicit simulation, but at far less time and computing costs. We develop a web application on nanoHUB that supports both HPC-driven simulation and the ML surrogate methods to produce simulation outputs. This tool is used for both in-classroom instruction and for solving homework problems associated with two courses covering topics in the broad areas of computational materials science, modeling and simulation, and engineering applications of HPC-enabled simulations. The evaluation of the tool via in-classroom student feedback and surveys shows that the ML-enhanced tool provides a dynamic and responsive simulation environment that enhances student learning. The improvement in the interactivity with the simulation framework in terms of real-time engagement and anytime access enables students to develop intuition for the physical system behavior through rapid visualization of variations in output quantities with changes in inputs., Comment: 10 pages, 6 figures

Designing reconfigurable materials based on deformable nanoparticles (NPs) hinges on an understanding of the energetically-favored shapes these NPs can adopt. Using simulations, we show that hollow, deformable patchy NPs tailored with surface charge patterns such as Janus patches, stripes, and polyhedrally-distributed patches differently adapt their shape in response to changes in patterns and ionic strength, transforming into capsules, hemispheres, variably-dimpled bowls, and polyhedra. The links between anisotropy in NP surface charge, shape, and the elastic energy density are discussed., Comment: 9 pages, 8 figures

Classical molecular dynamics simulations are based on solving Newton's equations of motion. Using a small timestep, numerical integrators such as Verlet generate trajectories of particles as solutions to Newton's equations. We introduce operators derived using recurrent neural networks that accurately solve Newton's equations utilizing sequences of past trajectory data, and produce energy-conserving dynamics of particles using timesteps up to 4000 times larger compared to the Verlet timestep. We demonstrate significant speedup in many example problems including 3D systems of up to 16 particles., Comment: 15 pages, 12 figures; updated content

Transient cloud servers such as Amazon Spot instances, Google Preemptible VMs, and Azure Low-priority batch VMs, can reduce cloud computing costs by as much as $10\times$, but can be unilaterally preempted by the cloud provider. Understanding preemption characteristics (such as frequency) is a key first step in minimizing the effect of preemptions on application performance, availability, and cost. However, little is understood about temporally constrained preemptions---wherein preemptions must occur in a given time window. We study temporally constrained preemptions by conducting a large scale empirical study of Google's Preemptible VMs (that have a maximum lifetime of 24 hours), develop a new preemption probability model, new model-driven resource management policies, and implement them in a batch computing service for scientific computing workloads. Our statistical and experimental analysis indicates that temporally constrained preemptions are not uniformly distributed, but are time-dependent and have a bathtub shape. We find that existing memoryless models and policies are not suitable for temporally constrained preemptions. We develop a new probability model for bathtub preemptions, and analyze it through the lens of reliability theory. To highlight the effectiveness of our model, we develop optimized policies for job scheduling and checkpointing. Compared to existing techniques, our model-based policies can reduce the probability of job failure by more than $2\times$. We also implement our policies as part of a batch computing service for scientific computing applications, which reduces cost by $5\times$ compared to conventional cloud deployments and keeps performance overheads under $3\%$., Comment: 12 pages, 9 figures; to appear at HPDC 2020

Simulating the dynamics of ions near polarizable nanoparticles (NPs) using coarse-grained models is extremely challenging due to the need to solve the Poisson equation at every simulation timestep. Recently, a molecular dynamics (MD) method based on a dynamical optimization framework bypassed this obstacle by representing the polarization charge density as virtual dynamic variables, and evolving them in parallel with the physical dynamics of ions. We highlight the computational gains accessible with the integration of machine learning (ML) methods for parameter prediction in MD simulations by demonstrating how they were realized in MD simulations of ions near polarizable NPs. An artificial neural network based regression model was integrated with MD simulation and predicted the optimal simulation timestep and optimization parameters characterizing the virtual system with $94.3\%$ success. The ML-enabled auto-tuning of parameters generated accurate dynamics of ions for $\approx 10$ million steps while improving the stability of the simulation by over an order of magnitude. The integration of ML-enhanced framework with hybrid OpenMP/MPI parallelization techniques reduced the computational time of simulating systems with thousands of ions and induced charges from thousands of hours to tens of hours, yielding a maximum speedup of $\approx 3$ from ML-only acceleration and a maximum speedup of $\approx 600$ from the combination of ML and parallel computing methods. Extraction of ionic structure in concentrated electrolytes near oil-water emulsions demonstrates the success of the method. The approach can be generalized to select optimal parameters in other MD applications and energy minimization problems., Comment: 16 pages, 10 figures

There is an ongoing debate concerning the best rheological model for liquid flows in elastohydrodynamic lubrication (EHL). Due to the small contact area and high relative velocities of bounding solids, the lubricant experiences pressures in excess of 500 MPa and strain rates that are typically $10^5 -10^7$ $\textrm{s}^{-1}$. The high pressures lead to a dramatic rise in Newtonian viscosity $\eta_{N}$ and the high rates lead to large shear stresses and pronounced shear-thinning. This paper presents detailed simulations of a model EHL fluid, squalane, using nonequilibrium molecular dynamics methods to extract the scaling of its viscosity with shear rate ($10^5 - 10^{10}$ $\textrm{s}^{-1}$) over a wide range of pressure $P$ (0.1 MPa to 1.2 GPa), and temperature $T$ ($150 - 373$ K). Simulation results are consistent with a broad range of equilibrium and nonequilibrium experiments. At high $T$ and low $P$, where $\eta_{N}$ is low, the response can be fit to a power-law, as in the common Carreau model. Shear-thinning becomes steeper as $\eta_{N}$ increases, and for $\eta_{N}\gtrsim 1$ Pa-s, shear-thinning is consistent with the thermally activated flow assumed by another common model, Eyring theory. Simulations for a bi-disperse Lennard-Jones (LJ) system show that the transition from Carreau to Eyring is generic. For both squalane and the LJ system, the viscosity decreases by only about a decade in the Carreau regime, but may fall by many orders of magnitude in the Eyring regime. Shear thinning is often assumed to reflect changing molecular alignment, but the alignment of squalane molecules saturates after the viscosity has dropped by only about a factor of three. In contrast, thermal activation describes shear thinning by six or more decades in viscosity. Changes in the diagonal elements of the stress tensor with rate and shear stress are also studied., Comment: 17 pages, 18 figures

The convergence of HPC and data-intensive methodologies provide a promising approach to major performance improvements. This paper provides a general description of the interaction between traditional HPC and ML approaches and motivates the Learning Everywhere paradigm for HPC. We introduce the concept of effective performance that one can achieve by combining learning methodologies with simulation-based approaches, and distinguish between traditional performance as measured by benchmark scores. To support the promise of integrating HPC and learning methods, this paper examines specific examples and opportunities across a series of domains. It concludes with a series of open computer science and cyberinfrastructure questions and challenges that the Learning Everywhere paradigm presents.

We provide exact expressions for the electrostatic energy of uniformly-charged prolate and oblate spheroidal shells. We find that uniformly-charged prolate spheroids of eccentricity greater than 0.9 have lower Coulomb energy than a sphere of the same area. For the volume-constrained case, we find that a sphere has the highest Coulomb energy among all spheroidal shells. Further, we derive the change in the Coulomb energy of a uniformly-charged shell due to small, area-conserving perturbations on the spherical shape. Our perturbation calculations show that buckling-type deformations on a sphere can lower the Coulomb energy. Finally, we consider the possibility of counterion condensation on the spheroidal shell surface. We employ a Manning-Oosawa two-state model approximation to evaluate the renormalized charge and analyze the behavior of the equilibrium free energy as a function of the shell's aspect ratio for both area-constrained and volume-constrained cases. Counterion condensation is seen to favor the formation of spheroidal structures over a sphere of equal area for high values of shell volume fractions., Comment: 15 pages, 7 figures, Accepted in Physical Review E

In the study of model electronic device systems where electrons are typically under confinement, a key obstacle is the need to iteratively solve the coupled Schr\"{o}dinger-Poisson (SP) equation. It is possible to bypass this obstacle by adopting a variational approach and obtaining the solution of the SP equation by minimizing a functional. Further, using molecular dynamics methods that treat the electronic potential as a dynamical variable, the functional can be minimized on the fly in conjunction with the update of other dynamical degrees of freedom leading to considerable reduction in computational costs. But such approaches require access to a true free-energy functional, one that evaluates to the equilibrium free energy at its minimum. In this paper, we present a variational formulation of the Schr\"{o}dinger-Poisson (SP) theory with the needed free-energy functional of the electronic potential. We apply our formulation to semiconducting nanostructures and provide the expression of the free-energy functional for narrow channel quantum wells where the local density approximation yields accurate physics and for the case of wider channels where Thomas-Fermi approximation is valid., Comment: 7 pages, 2 figures

Manipulating the shape of nanoscale objects in a controllable fashion is at the heart of designing materials that act as building blocks for self-assembly or serve as targeted drug delivery carriers. Inducing shape deformations by controlling external parameters is also an important way of designing biomimetic membranes. In this paper, we demonstrate that electrostatics can be used as a tool to manipulate the shape of soft, closed membranes by tuning environmental conditions such as the electrolyte concentration in the medium. Using a molecular dynamics-based simulated annealing procedure, we investigate charged elastic shells that do not exchange material with their environment, such as elastic membranes formed in emulsions or synthetic nanocontainers. We find that by decreasing the salt concentration or increasing the total charge on the shell's surface, the spherical symmetry is broken, leading to the formation of ellipsoids, discs, and bowls. Shape changes are accompanied by a significant lowering of the electrostatic energy and a rise in the surface area of the shell. To substantiate our simulation findings, we show analytically that a uniformly charged disc has a lower Coulomb energy than a sphere of the same volume. Further, we test the robustness of our results by including the effects of charge renormalization in the analysis of the shape transitions and find the latter to be feasible for a wide range of shell volume fractions., Comment: 6 pages, 4 figures

In simulating charged systems, it is often useful to treat some ionic components of the system at the mean-field level and solve the Poisson-Boltzmann (PB) equation to get their respective density profiles. The numerically intensive task of solving the PB equation at each step of the simulation can be bypassed using variational methods that treat the electrostatic potential as a dynamic variable. But such approaches require the access to a true free-energy functional; a functional that not only provides the correct solution of the PB equation upon extremization, it also evaluates to the true free energy of the system at its minimum. Moreover, the numerical efficiency of such procedures is further enhanced if the free-energy functional is local and is expressed in terms of the electrostatic potential. Existing PB functionals of the electrostatic potential, while possessing the local structure, are not free-energy functionals. We present a variational formulation with a local free-energy functional of the potential. In addition, we also construct a nonlocal free-energy functional of the electrostatic potential. These functionals are suited for employment in simulation schemes based on the ideas of dynamical optimization., Comment: 10 pages

In biological and synthetic materials, many important processes involve charges that are present in a medium with spatially varying dielectric permittivity. To accurately understand the role of electrostatic interactions in such systems, it is important to take into account the spatial dependence of the permittivity of the medium. However, due to the ensuing theoretical and computational challenges, this inhomogeneous dielectric response of the medium is often ignored or excessively simplified. We develop a variational formulation of electrostatics to accurately investigate systems that exhibit this inhomogeneous dielectric response. Our formulation is based on a true energy functional of the polarization charge density. The defining characteristic of a true energy functional is that at its minimum it evaluates to the actual value of the energy; this is a feature not found in many commonly used electrostatic functionals. We explore in detail the charged systems that exhibit sharp discontinuous change in dielectric permittivity, and we show that for this case our functional reduces to a functional of only the surface polarization charge density. We apply this reduced functional to study model problems for which analytical solutions are well known. We demonstrate, in addition, that the functional has many properties that make it ideal for use in molecular dynamics simulations., Comment: 14 pages, 7 figures

Variational principles are important in the investigation of large classes of physical systems. They can be used both as analytical methods as well as starting points for the formulation of powerful computational techniques such as dynamical optimization methods. Systems with charged objects in dielectric media and systems with magnetically active particles are important examples. In these examples and other important cases, the variational principles describing the system are required to obey a number of constraints. These constraints are implemented within the variational formulation by means of Lagrange multipliers. Such constrained variational formulations are in general not unique. For the application of efficient simulation methods, one must find specific formulations that satisfy a number of important conditions. An often required condition is that the functional be positive-definite, in other words, its extrema be actual minima. In this article, we present a general approach to attack the problem of finding, among equivalent variational functionals, those that generate true minima. The method is based on the modification of the Lagrange multiplier which allows us to generate large families of effective variational formulations associated with a single original constrained variational principle. We demonstrate its application to different examples and, in particular, to the important cases of Poisson and Poisson-Boltzmann equations. We show how to obtain variational formulations for these systems with extrema that are always minima., Comment: 15 pages, 1 figure

For charged systems in heterogeneous dielectric media, a key obstacle for molecular dynamics (MD) simulations is the need to solve the Poisson equation in the media. This obstacle can be bypassed using MD methods that treat the local polarization charge density as a dynamic variable, but such approaches require access to a true free energy functional; one that evaluates to the equilibrium electrostatic energy at its minimum. In this letter, we derive the needed functional. As an application, we develop a Car-Parrinello MD method for the simulation of free charges present near a spherical emulsion droplet separating two immiscible liquids with different dielectric constants. Our results show the presence of non-monotonic ionic profiles in the dielectric with lower dielectric constant.