Search for: All records

We present a fully integrated AI-driven framework for rapid endurance prediction in NVDRAM ferroelectric capacitors. Endurance testing is one of the most time- and resource-intensive steps in memory characterization, often requiring up to 10¹² cycles per device. To overcome the scarcity of endurance training data, we propose an experimentally calibrated synthetic data generation pipeline using kinetic Monte Carlo (kMC) simulations in Ginestra™, seeded with experimentally extracted defect parameters. We train a transformer-based AI surrogate using this high-fidelity dataset, achieving an R² of 0.992 and enabling ~105x speedup in defect evolution prediction. The surrogate generates large-scale synthetic datasets by sampling initial defect profiles, which are then used to train a hybrid multi-layer perceptron (MLP)- attention model that maps early-life defect characteristics to Weibull endurance distributions. This final endurance prediction model achieves strong agreement with ground truth Weibull parameters, with R² values of >0.98 for η and ~0.9 for β, demonstrating its reliability in capturing endurance distribution characteristics. Wafer-scale prediction of breakdown distributions is demonstrated in one-shot, reducing characterization time by over 10 orders of magnitude. This framework enables scalable, high-throughput reliability screening for ferroelectric memory technologies. 

Abstract The drive toward non‐von Neumann device architectures has led to an intense focus on insulator‐to‐metal (IMT) and the converse metal‐to‐insulator (MIT) transitions. Studies of electric field‐driven IMT in the prototypical VO₂thin‐film channel devices are largely focused on the electrical and elastic responses of the films, but the response of the corresponding TiO₂substrate is often overlooked, since it is nominally expected to be electrically passive and elastically rigid. Here, in‐operando spatiotemporal imaging of the coupled elastodynamics using X‐ray diffraction microscopy of a VO₂film channel device on TiO₂substrate reveals two new surprises. First, the film channel bulges during the IMT, the opposite of the expected shrinking in the film undergoing IMT. Second, a microns thick proximal layer in the substrate also coherently bulges accompanying the IMT in the film, which is completely unexpected. Phase‐field simulations of coupled IMT, oxygen vacancy electronic dynamics, and electronic carrier diffusion incorporating thermal and strain effects suggest that the observed elastodynamics can be explained by the known naturally occurring oxygen vacancies that rapidly ionize (and deionize) in concert with the IMT (MIT). Fast electrical‐triggering of the IMT via ionizing defects and an active “IMT‐like” substrate layer are critical aspects to consider in device applications. 

Abstract Transparent oxide thin film transistors (TFTs) are an important ingredient of transparent electronics. Their fabrication at the back‐end‐of‐line (BEOL) opens the door to novel strategies to more closely integrate logic with memory for data‐intensive computing architectures that overcome the scaling challenges of today's integrated circuits. A recently developed variant of molecular‐beam epitaxy (MBE) called suboxide MBE (S‐MBE) is demonstrated to be capable of growing epitaxial In₂O₃at BEOL temperatures with unmatched crystal quality. The fullwidth at halfmaximum of the rocking curve is 0.015° and, thus, ≈5x narrower than any reports at any temperature to date and limited by the substrate quality. The key to achieving these results is the provision of an In₂O beam by S‐MBE, which enables growth in adsorption control and is kinetically favorable. To benchmark this deposition method for TFTs, rudimentary devices were fabricated. 

Abstract Many real-world mission-critical applications require continual online learning from noisy data and real-time decision making with a defined confidence level. Brain-inspired probabilistic models of neural network can explicitly handle the uncertainty in data and allow adaptive learning on the fly. However, their implementation in a compact, low-power hardware remains a challenge. In this work, we introduce a novel hardware fabric that can implement a new class of stochastic neural network called Neural Sampling Machine (NSM) by exploiting the stochasticity in the synaptic connections for approximate Bayesian inference. We experimentally demonstrate an in silico hybrid stochastic synapse by pairing a ferroelectric field-effect transistor (FeFET)-based analog weight cell with a two-terminal stochastic selector element. We show that the stochastic switching characteristic of the selector between the insulator and the metallic states resembles the multiplicative synaptic noise of the NSM. We perform network-level simulations to highlight the salient features offered by the stochastic NSM such as performing autonomous weight normalization for continual online learning and Bayesian inferencing. We show that the stochastic NSM can not only perform highly accurate image classification with 98.25% accuracy on standard MNIST dataset, but also estimate the uncertainty in prediction (measured in terms of the entropy of prediction) when the digits of the MNIST dataset are rotated. Building such a probabilistic hardware platform that can support neuroscience inspired models can enhance the learning and inference capability of the current artificial intelligence (AI).