A deep-learning approach to realizing functionality in nanoelectronic devices

Topological learning in multiclass data sets

We specialize techniques from topological data analysis to the problem of characterizing the topological complexity (as defined in the body of the paper) of a multiclass data set. As a by-product, a topological classifier is defined that uses an open subcovering of the data set. This subcovering can be used to construct a simplicial complex whose topological features (e.g., Betti numbers) provide information about the classification problem. We use these topological constructs to study the impact of topological complexity on learning in feedforward deep neural networks (DNNs). We hypothesize that topological complexity is negatively correlated with the ability of a fully connected feedforward deep neural network to learn to classify data correctly. We evaluate our topological classification algorithm on multiple constructed and open-source data sets. We also validate our hypothesis regarding the relationship between topological complexity and learning in DNN's on multiple data sets.

Bioactive VS4-based sonosensitizer for robust chemodynamic, sonodynamic and osteogenic therapy of infected bone defects

BACKGROUND

Most bone defects caused by bone disease or trauma are accompanied by infection, and there is a high risk of infection spread and defect expansion. Traditional clinical treatment plans often fail due to issues like antibiotic resistance and non-union of bones. Therefore, the treatment of infected bone defects requires a strategy that simultaneously achieves high antibacterial efficiency and promotes bone regeneration.

RESULTS

In this study, an ultrasound responsive vanadium tetrasulfide-loaded MXene (VSM) Schottky junction is constructed for rapid methicillin-resistant staphylococcus aureus (MRSA) clearance and bone regeneration. Due to the peroxidase (POD)-like activity of VS4 and the abundant Schottky junctions, VSM has high electron-hole separation efficiency and a decreased band gap, exhibiting a strong chemodynamic and sonodynamic antibacterial efficiency of 94.03%. Under the stimulation of medical dose ultrasound, the steady release of vanadium element promotes the osteogenic differentiation of human bone marrow mesenchymal stem cells (hBMSCs). The in vivo application of VSM in infected tibial plateau bone defects of rats also has a great therapeutic effect, eliminating MRSA infection, then inhibiting inflammation and improving bone regeneration.

CONCLUSION

The present work successfully develops an ultrasound responsive VS4-based versatile sonosensitizer for robust effective antibacterial and osteogenic therapy of infected bone defects.

Backpropagation-free training of deep physical neural networks

Recent successes in deep learning for vision and natural language processing are attributed to larger models but come with energy consumption and scalability issues. Current training of digital deep learning models primarily relies on backpropagation that is unsuitable for physical implementation. Here, we proposed a simple deep neural network architecture augmented by a physical local learning (PhyLL) algorithm, enabling supervised and unsupervised training of deep physical neural networks, without detailed knowledge of the nonlinear physical layer's properties. We trained diverse wave-based physical neural networks in vowel and image classification experiments, showcasing our approach's universality. Our method shows advantages over other hardware-aware training schemes by improving training speed, enhancing robustness, and reducing power consumption by eliminating the need for system modelling and thus decreasing digital computation.

Machine learning models for efficient characterization of Schottky barrier photodiode internal parameters

We propose ANN-based models to analyze and extract the internal parameters of a Schottky photodiode (SPD) without presenting them with any knowledge of the highly nonlinear thermionic emission (TE) expression of the device current. We train, evaluate and demonstrate the ML models on thirty-six private datasets from three previously published devices, which denote current responses under illumination and ambient temperature of graphene oxide (GO) doped p-Si Schottky barrier diodes (SBDs). The GO doping levels are 0%, 1%, 3%, 5%, and 10%. The illumination ranged from dark (0 mW/cm2) to 30 mW/cm2. The predictions are then made completely at the intensity of 60 mW/cm2. For each diode, some values of the barrier height ([Formula: see text]), ideality factor (n), and series resistance ([Formula: see text]) independently calculated using the Cheung-Cheung method were included in the training dataset. The predictions are done at unspecified intensities on the model development data at 80 and 100 mW/cm2, and on external data at 5% and 20% GO doping which were not part of the development dataset. The ANN achieved a mean square error and mean absolute error score below 0.003 across all datasets. This demonstrates the effective learning capabilities of the ANN models in accurately capturing the photo responses of the photodiodes and accurately predicting the internal parameters of the Schottky Barrier Diodes (SBDs), all without relying on an inherent understanding of the thermionic emission (TE) equation for SBDs. The ANN models achieved high accuracy in this process. The proposed ML models can significantly reduce analysis time in device development cycles and can be applied to other datasets in various fields.

Toward a formal theory for computing machines made out of whatever physics offers

Approaching limitations of digital computing technologies have spurred research in neuromorphic and other unconventional approaches to computing. Here we argue that if we want to engineer unconventional computing systems in a systematic way, we need guidance from a formal theory that is different from the classical symbolic-algorithmic Turing machine theory. We propose a general strategy for developing such a theory, and within that general view, a specific approach that we call fluent computing. In contrast to Turing, who modeled computing processes from a top-down perspective as symbolic reasoning, we adopt the scientific paradigm of physics and model physical computing systems bottom-up by formalizing what can ultimately be measured in a physical computing system. This leads to an understanding of computing as the structuring of processes, while classical models of computing systems describe the processing of structures.

Transfer Learning of Full Molecular Weight Distributions via High-Throughput Computer-Controlled Polymerization

The skew and shape of the molecular weight distribution (MWD) of polymers have a significant impact on polymer physical properties. Standard summary metrics statistically derived from the MWD only provide an incomplete picture of the polymer MWD. Machine learning (ML) methods coupled with high-throughput experimentation (HTE) could potentially allow for the prediction of the entire polymer MWD without information loss. In our work, we demonstrate a computer-controlled HTE platform that is able to run up to 8 unique variable conditions in parallel for the free radical polymerization of styrene. The segmented-flow HTE system was equipped with an inline Raman spectrometer and offline size exclusion chromatography (SEC) to obtain time-dependent conversion and MWD, respectively. Using ML forward models, we first predict monomer conversion, intrinsically learning varying polymerization kinetics that change for each experimental condition. In addition, we predict entire MWDs including the skew and shape as well as SHAP analysis to interpret the dependence on reagent concentrations and reaction time. We then used a transfer learning approach to use the data from our high-throughput flow reactor to predict batch polymerization MWDs with only three additional data points. Overall, we demonstrate that the combination of HTE and ML provides a high level of predictive accuracy in determining polymerization outcomes. Transfer learning can allow exploration outside existing parameter spaces efficiently, providing polymer chemists with the ability to target the synthesis of polymers with desired properties.

Dopant network processing units as tuneable extreme learning machines

Inspired by the highly efficient information processing of the brain, which is based on the chemistry and physics of biological tissue, any material system and its physical properties could in principle be exploited for computation. However, it is not always obvious how to use a material system’s computational potential to the fullest. Here, we operate a dopant network processing unit (DNPU) as a tuneable extreme learning machine (ELM) and combine the principles of artificial evolution and ELM to optimise its computational performance on a non-linear classification benchmark task. We find that, for this task, there is an optimal, hybrid operation mode (“tuneable ELM mode”) in between the traditional ELM computing regime with a fixed DNPU and linearly weighted outputs (“fixed-ELM mode”) and the regime where the outputs of the non-linear system are directly tuned to generate the desired output (“direct-output mode”). We show that the tuneable ELM mode reduces the number of parameters needed to perform a formant-based vowel recognition benchmark task. Our results emphasise the power of analog in-matter computing and underline the importance of designing specialised material systems to optimally utilise their physical properties for computation.

Machine Learning Optimization of Majorana Hybrid Nanowires

As the complexity of quantum systems such as quantum bit arrays increases, efforts to automate expensive tuning are increasingly worthwhile. We investigate machine learning based tuning of gate arrays using the covariance matrix adaptation evolution strategy algorithm for the case study of Majorana wires with strong disorder. We find that the algorithm is able to efficiently improve the topological signatures, learn intrinsic disorder profiles, and completely eliminate disorder effects. For example, with only 20 gates, it is possible to fully recover Majorana zero modes destroyed by disorder by optimizing gate voltages.

A comprehensive deep learning method for empirical spectral prediction and its quantitative validation of nano-structured dimers

Nanophotonics exploits the best of photonics and nanotechnology which has transformed optics in recent years by allowing subwavelength structures to enhance light-matter interactions. Despite these breakthroughs, design, fabrication, and characterization of such exotic devices have remained through iterative processes which are often computationally costly, memory-intensive, and time-consuming. In contrast, deep learning approaches have recently shown excellent performance as practical computational tools, providing an alternate avenue for speeding up such nanophotonics simulations. This study presents a DNN framework for transmission, reflection, and absorption spectra predictions by grasping the hidden correlation between the independent nanostructure properties and their corresponding optical responses. The proposed DNN framework is shown to require a sufficient amount of training data to achieve an accurate approximation of the optical performance derived from computational models. The fully trained framework can outperform a traditional EM solution using on the COMSOL Multiphysics approach in terms of computational cost by three orders of magnitude. Furthermore, employing deep learning methodologies, the proposed DNN framework makes an effort to optimise design elements that influence the geometrical dimensions of the nanostructure, offering insight into the universal transmission, reflection, and absorption spectra predictions at the nanoscale. This paradigm improves the viability of complicated nanostructure design and analysis, and it has a lot of potential applications involving exotic light-matter interactions between nanostructures and electromagnetic fields. In terms of computational times, the designed algorithm is more than 700 times faster as compared to conventional FEM method (when manual meshing is used). Hence, this approach paves the way for fast yet universal methods for the characterization and analysis of the optical response of nanophotonic systems.

Dynamic molecular switches with hysteretic negative differential conductance emulating synaptic behaviour

To realize molecular-scale electrical operations beyond the von Neumann bottleneck, new types of multifunctional switches are needed that mimic self-learning or neuromorphic computing by dynamically toggling between multiple operations that depend on their past. Here, we report a molecule that switches from high to low conductance states with massive negative memristive behaviour that depends on the drive speed and number of past switching events, with all the measurements fully modelled using atomistic and analytical models. This dynamic molecular switch emulates synaptic behavior and Pavlovian learning, all within a 2.4-nm-thick layer that is three orders of magnitude thinner than a neuronal synapse. The dynamic molecular switch provides all the fundamental logic gates necessary for deep learning because of its time-domain and voltage-dependent plasticity. The synapse-mimicking multifunctional dynamic molecular switch represents an adaptable molecular-scale hardware operable in solid-state devices, and opens a pathway to simplify dynamic complex electrical operations encoded within a single ultracompact component.

Dynamical stochastic simulation of complex electrical behavior in neuromorphic networks of metallic nanojunctions

Nanostructured Au films fabricated by the assembling of nanoparticles produced in the gas phase have shown properties suitable for neuromorphic computing applications: they are characterized by a non-linear and non-local electrical behavior, featuring switches of the electric resistance whose activation is typically triggered by an applied voltage over a certain threshold. These systems can be considered as complex networks of metallic nanojunctions where thermal effects at the nanoscale cause the continuous rearrangement of regions with low and high electrical resistance. In order to gain a deeper understanding of the electrical properties of this nano granular system, we developed a model based on a large three dimensional regular resistor network with non-linear conduction mechanisms and stochastic updates of conductances. Remarkably, by increasing enough the number of nodes in the network, the features experimentally observed in the electrical conduction properties of nanostructured gold films are qualitatively reproduced in the dynamical behavior of the system. In the activated non-linear conduction regime, our model reproduces also the growing trend, as a function of the subsystem size, of quantities like Mutual and Integrated Information, which have been extracted from the experimental resistance series data via an information theoretic analysis. This indicates that nanostructured Au films (and our model) possess a certain degree of activated interconnection among different areas which, in principle, could be exploited for neuromorphic computing applications.

Double gate operation of metal nanodot array based single electron device

Multidot single-electron devices (SEDs) can enable new types of computing technologies, such as those that are reconfigurable and reservoir-computing. A self-assembled metal nanodot array film that is attached to multiple gates is a candidate for use in such SEDs for achieving high functionality. However, the single-electron properties of such a film have not yet been investigated in conjunction with optimally controlled multiple gates because of the structural complexity of incorporating many nanodots. In this study, Fe nanodot-array-based double-gate SEDs were fabricated by vacuum deposition, and their single-electron properties (modulated by the top- and bottom-gate voltages; V_T and V_B, respectively) were investigated. The phase of the Coulomb blockade oscillation systematically shifted with V_T, indicating that the charge state of the single dot was controlled by both the gate voltages despite the metallic random multidot structure. This result demonstrates that the Coulomb blockade oscillation (originating from the dot in the multidot array) can be modulated by the two gates. The top and bottom gates affected the electronic state of the dot unevenly owing to the geometrical effect caused by the following: (1) vertically asymmetric dot shape and (2) variation of the dot size (including the surrounding dots). This is a characteristic feature of a nanodot array that uses self-assembled metal dots; for example, prepared by vacuum deposition. Such variations derived from a randomly distributed nanodot array will be useful in enhancing the functionality of multidot devices.

Artificial Neural Network Modelling for Optimizing the Optical Parameters of Plasmonic Paired Nanostructures

The Artificial Neural Network (ANN) has become an attractive approach in Machine Learning (ML) to analyze a complex data-driven problem. Due to its time efficient findings, it has became popular in many scientific fields such as physics, optics, and material science. This paper presents a new approach to design and optimize the electromagnetic plasmonic nanostructures using a computationally efficient method based on the ANN. In this work, the nanostructures have been simulated by using a Finite Element Method (FEM), then Artificial Intelligence (AI) is used for making predictions of associated sensitivity (S), Full Width Half Maximum (FWHM), Figure of Merit (FOM), and Plasmonic Wavelength (PW) for different paired nanostructures. At first, the computational model is developed by using a Finite Element Method (FEM) to prepare the dataset. The input parameters were considered as the Major axis, a, the Minor axis, b, and the separation gap, g, which have been used to calculate the corresponding sensitivity (nm/RIU), FWHM (nm), FOM, and plasmonic wavelength (nm) to prepare the dataset. Secondly, the neural network has been designed where the number of hidden layers and neurons were optimized as part of a comprehensive analysis to improve the efficiency of ML model. After successfully optimizing the neural network, this model is used to make predictions for specific inputs and its corresponding outputs. This article also compares the error between the predicted and simulated results. This approach outperforms the direct numerical simulation methods for predicting output for various input device parameters.

Deep physical neural networks trained with backpropagation

Deep-learning models have become pervasive tools in science and engineering. However, their energy requirements now increasingly limit their scalability1. Deep-learning accelerators2-9 aim to perform deep learning energy-efficiently, usually targeting the inference phase and often by exploiting physical substrates beyond conventional electronics. Approaches so far10-22 have been unable to apply the backpropagation algorithm to train unconventional novel hardware in situ. The advantages of backpropagation have made it the de facto training method for large-scale neural networks, so this deficiency constitutes a major impediment. Here we introduce a hybrid in situ-in silico algorithm, called physics-aware training, that applies backpropagation to train controllable physical systems. Just as deep learning realizes computations with deep neural networks made from layers of mathematical functions, our approach allows us to train deep physical neural networks made from layers of controllable physical systems, even when the physical layers lack any mathematical isomorphism to conventional artificial neural network layers. To demonstrate the universality of our approach, we train diverse physical neural networks based on optics, mechanics and electronics to experimentally perform audio and image classification tasks. Physics-aware training combines the scalability of backpropagation with the automatic mitigation of imperfections and noise achievable with in situ algorithms. Physical neural networks have the potential to perform machine learning faster and more energy-efficiently than conventional electronic processors and, more broadly, can endow physical systems with automatically designed physical functionalities, for example, for robotics23-26, materials27-29 and smart sensors30-32.

Dynamic Processes in a Superconducting Adiabatic Neuron with Non-Shunted Josephson Contacts

We investigated the dynamic processes in a superconducting neuron based on Josephson contacts without resistive shunting (SC-neuron). Such a cell is a key element of perceptron-type neural networks that operate in both classical and quantum modes. The analysis of the obtained results allowed us to find the mode when the transfer characteristic of the element implements the “sigmoid” activation function. The numerical approach to the analysis of the equations of motion and the Monte Carlo method revealed the influence of inertia (capacitances), dissipation, and temperature on the dynamic characteristics of the neuron.

The rise of intelligent matter

Artificial intelligence (AI) is accelerating the development of unconventional computing paradigms inspired by the abilities and energy efficiency of the brain. The human brain excels especially in computationally intensive cognitive tasks, such as pattern recognition and classification. A long-term goal is de-centralized neuromorphic computing, relying on a network of distributed cores to mimic the massive parallelism of the brain, thus rigorously following a nature-inspired approach for information processing. Through the gradual transformation of interconnected computing blocks into continuous computing tissue, the development of advanced forms of matter exhibiting basic features of intelligence can be envisioned, able to learn and process information in a delocalized manner. Such intelligent matter would interact with the environment by receiving and responding to external stimuli, while internally adapting its structure to enable the distribution and storage (as memory) of information. We review progress towards implementations of intelligent matter using molecular systems, soft materials or solid-state materials, with respect to applications in soft robotics, the development of adaptive artificial skins and distributed neuromorphic computing.