Stack Memory Research Articles

Scabbard: An Exploratory Study on Hardware Aware Design Choices of Learning with Rounding-based Key Encapsulation Mechanisms

Recently, the construction of cryptographic schemes based on hard lattice problems has gained immense popularity. Apart from being quantum resistant, lattice-based cryptography allows a wide range of variations in the underlying hard problem. As cryptographic schemes can work in different environments under different operational constraints such as memory footprint, silicon area, efficiency, power requirement, and so on, such variations in the underlying hard problem are very useful for designers to construct different cryptographic schemes. In this work, we explore various design choices of lattice-based cryptography and their impact on performance in the real world. In particular, we propose a suite of key-encapsulation mechanisms based on the learning with rounding problem with a focus on improving different performance aspects of lattice-based cryptography. Our suite consists of three schemes. Our first scheme is Florete, which is designed for efficiency. The second scheme is Espada, which is aimed at improving parallelization, flexibility, and memory footprint. The last scheme is Sable, which can be considered an improved version in terms of key sizes and parameters of the Saber key-encapsulation mechanism, one of the finalists in the National Institute of Standards and Technology’s post-quantum standardization procedure. In this work, we have described our design rationale behind each scheme. Furthermore, to demonstrate the justification of our design decisions, we have provided software and hardware implementations. Our results show Florete is faster than most state-of-the-art KEMs on software platforms. For example, the key-generation algorithm of high-security version Florete outperforms the National Institute of Standards and Technology’s standard Kyber by 47%, the Federal Office for Information Security’s standard Frodo by 99%, and Saber by 57% on the ARM Cortex-M4 platform. Similarly, in hardware, Florete outperforms Frodo and NTRU Prime for all KEM operations. The scheme Espada requires less memory and area than the implementation of most state-of-the-art schemes. For example, the encapsulation algorithm of high-security version Espada uses 30% less stack memory than Kyber, 57% less stack memory than Frodo, and 67% less stack memory than Saber on the ARM Cortex-M4 platform. The implementations of Sable maintain a tradeoff between Florete and Espada regarding software performance and memory requirements. Sable outperforms Saber at least by 6% and Frodo by 99%. Through an efficient polynomial multiplier design, which exploits the small secret size, Sable outperforms most state-of-the-art KEMs, including Saber, Frodo, and NTRU Prime. The implementations of Sable that use number theoretic transform-based polynomial multiplication (SableNTT) surpass all the state-of-the-art schemes in performance, which are optimized for speed on the Cortext M4 platform. The performance benefit of SableNTT against Kyber lies in between 7-29%, 2-13% for Saber, and around 99% for Frodo.

Intelligent Method of Identifying the Nonlinear Dynamic Model for Helicopter Turboshaft Engines

This research focused on the helicopter turboshaft engine dynamic model, identifying task solving in unsteady and transient modes (engine starting and acceleration) based on sensor data. It is known that about 85% of helicopter turboshaft engines operate in steady-state modes, while only around 15% operate in unsteady and transient modes. Therefore, developing dynamic multi-mode models that account for engine behavior during these modes is a critical scientific and practical task. The dynamic model for starting and acceleration modes has been further developed using on-board parameters recorded by sensors (gas-generator rotor r.p.m., free turbine rotor speed, gas temperature in front of the compressor turbine, fuel consumption) to achieve a 99.88% accuracy in identifying the dynamics of these parameters. An improved Elman recurrent neural network with dynamic stack memory was introduced, enhancing the robustness and increasing the performance by 2.7 times compared to traditional Elman networks. A theorem was proposed and proven, demonstrating that the total execution time for N Push and Pop operations in the dynamic stack memory does not exceed a certain value O(N). The training algorithm for the Elman network was improved using time delay considerations and Butterworth filter preprocessing, reducing the loss function from 2.5 to 0.12% over 120 epochs. The gradient diagram showed a decrease over time, indicating the model’s approach to the minimum loss function, with optimal settings ensuring the stable training.

Video Analytic for Human Management and Security and FPGA Accelerated High Concurrency

This paper explores the use of video analytics by leveraging accelerated FPGA technology in combination with high-performance computing, specifically utilizing two Xilinx Alveo U50Lv cards and one U55C card. While many applications exist for motion analysis and detection in videos, the use of FPGAs in this context remains relatively scarce. FPGAs offer significant advantages in terms of energy efficiency and throughput. We present results demonstrating the parallelism capabilities in terms of the number of threads within a single Docker container that shares stack memory, as well as across multiple Docker containers. When operating within a single Docker process, the application shares the same memory space and resources, making it ideal for tasks that require efficient communication or data sharing. In contrast, running in multiple containers isolates processes, each with its own environment, and can significantly increase the number of threads. Our findings show that the combination of these techniques offers optimal performance for video analytics.

Urban air quality index forecasting using multivariate convolutional neural network based customized stacked long short-term memory model

In the context of increasing urban pollution and its adverse health effects, this study focuses on enhancing the air quality prediction model by forecasting future concentrations of various Air Quality Index (AQI) levels to support the development of green smart cities. The primary objective of this study is to develop a robust multivariate time series forecasting model using a combination of Convolutional Neural Networks (CNN) and Customized Stacked-based Long Short-Term Memory (CSLSTM) networks, namely C2SLSTM. The proposed methodology involved preprocessing AQI and meteorological data, scaling the features, and employing a hybrid CNN-LSTM architecture to capture spatial and temporal dependencies in the data. Dynamic AQI pattern changes are computed using online learning. The proposed model was trained on a big dataset containing AQIs such as particulate matter (PM2.5), carbon monoxide (CO), and nitrogen dioxide (NO2), along with meteorological factors like weather count, temperature, wind, humidity, and pressure. Experimental results demonstrate that the model achieved a mean absolute error (MAE) of 0.25, mean square error (MSE) of 0.083, root mean square error (RMSE) of 0.289 and a R2-Score of 0.84 on the test set. Also, compared to Long Short-Term Memory (LSTM), Convolutional Neural Network (CNN), Stacked Long Short-Term Memory (SLSTM), and Multilayer Perceptron (MLP) models, the C2SLSTM model showcases a performance enhancement of 35 %, 50 %, 30 %, and 32 %. Notably, the CNN component effectively extracted spatial features, while the LSTM layers captured temporal patterns, leading to precise predictions. These results indicate that the proposed approach effectively captures temporal dependencies and improves forecasting performance, with R2-Score values tending towards 1 for 12-hour prediction intervals. This research demonstrates the potential of advanced SLSTM architectures in accurately predicting AQI metrics, which can aid in better environmental monitoring and management. Also, the spatiotemporal correlation analysis is done through the designed ir metric.

An intelligent security mechanism in mobile Ad-Hoc networks using precision probability genetic algorithms (PPGA) and deep learning technique (Stacked LSTM)

The Mobile Ad-hoc Networks (MANETs) have gained a significant attention in the recent years due to their proliferation and huge application purposes. To defend from many types of modern cyber dangers like Distributed Denial of Service (DDoS) attacks, Advanced Persistent Threats (APTs), Insider Threats, Ransomware, Zero-Day Exploits, Social Engineering tactics, and etc is not easy when it comes to keeping MANETs security. These complex assaults focus on network infrastructure, take advantage of weaknesses in communication protocols and control user actions. Although there have been improvements in intrusion detection systems (IDS), it is still difficult to fully safeguard MANETs. The purpose of this research is to create advanced methods that can accurately find and decrease attacks inside MANETs. Applying Sensor-based Feature Extraction (SFE) to extract useful network features such as Received Signal Strength Indication (RSSI) and Time of Travel (TOT) from datasets NSL-KDD and CICIDS-2017. Utilizing the fresh method of Precise Probability Genetic Algorithm (PPGA) optimization for removing unrelated details, which enhances precision in detecting attacks. Predicting normal and attacking labels by applying Stacked Recurrent Long Short Term Memory (SRLSTM) method, fine-tuning classifier's parameters in every layer to improve outcomes. In order to authenticate and compare the suggested methods with current attack detection tactics, this study will make use of various evaluation measurements. The NSL-KDD, which is a benchmark dataset in network intrusion detection research, has a wide variety of network traffic data with instances that are labeled as normal and different attacks. CICIDS-2017 is similar because it contains an extensive dataset too - this includes real-world traces from network traffic where there's both regular activity and harmful actions. The purpose is to enhance the existing status of MANET security so as it can withstand more strongly against cyber dangers. According to the outcomes, it is analyzed that the attack detection accuracy has improved greatly 99 % when compared to other methods, as shown by the detailed assessment measurements. Better handling of big datasets with top detection accuracy reduces the time needed 8.9 s for training and testing models. Decrease in misclassification results and better ability to differentiate normal network actions from harmful intrusions. Improved resistance of MANETs to different cyber dangers, guaranteeing the safety and dependability of network communication in changing and non-centralized settings.

Deep learning-based water quality index classification using stacked ensemble variational mode decomposition

Water is crucial to human survival in general, and determining the WQI (water quality index) is one of the primary aspects. The existing water quality classification models are facing various challenges and gaps that are impeding their effectiveness. These challenges include limited data availability, the intricate nature of water systems, spatial and temporal variability, non-linear relationships, sensor noise, and error, interpretability, and explainability. It is imperative to address these challenges to improve the accuracy and efficacy of the models and to ensure that they continue to serve as reliable tools for monitoring and safeguarding water quality. To solve the issues, this paper proposes a Stacked Ensemble efficient long short-term memory (StackEL) model for an efficient water quality index classification. At first, the raw input data is pre-processed to rescale the input data using data normalization and one-hot encoding. After that, the process known as variational mode decomposition (VMD) is applied to get at the intrinsic mode functions (IMFs). Consequently, feature selection is performed using an extended coati optimization (EX-CoA) algorithm to select the most significant attributes from the feature selection. Here, publicly available datasets, namely the water quality dataset from Kaggle, are used for classification and performed using are used to perform the Stacked Ensemble efficient long short-term memory (StackEL) classification process effectively. To further perfect the proposed prediction model, the Dwarf Mongoose optimization (DMO) method is implemented. Several measures of effectiveness are examined. When compared to other existing models, the suggested model can achieve a high accuracy of 98.85% of the water quality dataset.

A Novel Deep Learning Model for Student Performance Prediction Using Engagement Data

Technology-enhanced learning supported by virtual learning environments (VLEs) facilitates tutors and students. VLE platforms contain a wealth of information that can be used to mine insight regarding students’ learning behaviour and relationships between behaviour and academic performance, as well as to model data-driven decision-making. This study introduces a system that we termed ASIST: a novel Attention-aware convolutional Stacked BiLSTM network for student representation learning to predict their performance. ASIST exploits student academic registry, VLE click stream, and midterm continuous assessment information for their behaviour representation learning. ASIST jointly learns the student representation using five behaviour vectors. It processes the four sequential behaviour vectors using a separate stacked bidirectional long short term memory (LSTM) network. A deep convolutional neural network models the diurnal weekly interaction behaviour. It also employs the attention mechanism to assign weight to features based on their importance. Next, five encoded feature vectors are concatenated with the assessment information, and, finally, a softmax layer predicts the high-performer (H), moderate-performer (M), and at-risk (F) categories of students. We evaluate ASIST over three datasets from an Irish university, considering five evaluation metrics. ASIST achieves an area under the curve (AUC) score of 0.86 to 0.90 over the three datasets. It outperforms three baseline deep learning models and four traditional classification models. We also found that the attention mechanism has a slight impact on ASIST’s performance. The ablation analysis reveals that weekly event count has the greatest impact on ASIST, whereas diurnal weekly interaction has the least impact. The early prediction using the first seven weeks of data achieves an AUC of 0.83 up to 0.89 over the three datasets. In yearly analysis, ASIST performs best over the 2018/19 dataset and worst over the 2020/21 dataset.

A milling tool wear predicting method with processing generalization capability

Tool Wear Prediction (TWP) is crucial to ensure machining quality. Compared to traditional prediction methods based on full-life or failure data, labelled data are limited or nonexistent when a new tool or similar tools have emerged. This makes it extremely challenging to predict wear using limited tracking data. This work aims to provide an innovative method for accurate long-term wear prediction of milling processes using limited monitoring data of individual tools. First, the Variational Modal Decomposition (VMD) and Pearson Correlation Threshold (PCT) are combined to adaptively filter the corresponding machining signals. Second, fitness analysis is performed by selecting features strongly related to tool degradation using feature monotonicity metrics, and sensitive information is fused based on Kernel Principal Components Analysis (KPCA). Third, a Stacked Bi-directional Long Short-Term Memory with attention mechanism (AT-SBiLSTM) model is constructed to ascertain the connection between sensitive feature and the future wear state of the tool by multi-step advance rolling prediction. The results of the various milling experiment show that the proposed approach can provide accurate forecasts of tool wear using limited monitoring data.

Ferroelectric Hf0.5Zr0.5O2 for Analog Memory and In‐Memory Computing Applications Down to Deep Cryogenic Temperatures

AbstractLow‐power nonvolatile memories operating down to deep cryogenic temperatures are important for a large spectrum of applications from high‐performance computing, electronics interfacing quantum computing hardware to space‐based electronics. Despite the potential of Hf0.5Zr0.5O2 (HZO), thanks to its compatibility with complementary metal‐oxide‐semiconductor (CMOS) back‐end‐of‐line processing, only few studies of HZO‐based memory devices down to cryogenic operation temperatures exist. Here, analog ferroelectric memory stack fabrication with 10 nm HZO and their detailed characterization under wide range of pulse amplitudes and frequencies down to 4 K are reported. When operated at temperatures below 100 K, HZO devices can support high amplitude voltage pulses, yielding record high Pr of up to 75µC cm−2 at ±7 Vp (14 Vpp) pulse amplitudes accompanied with frequency‐dependent memory window between 6 and 8 V. Devices show excellent endurance exceeding 109 cycles of ±5 Vp (10 Vpp) and Pr of 30 µC cm−2 without significant degradation of coercive voltages or loss of polarization at cryogenic temperatures. At least 20 reproducible analog states for temperatures below 100 K with almost ideal linearity of intermediate polarization states in both pulse directions is observed, demonstrating the high potential of analog cryogenic ferroelectric memories, essential for on‐line training in in‐memory‐computing architecture.

Mixed-Input Residual Network for Air Temperature Forecasting by Low-Power Embedded Devices

Accuracy and model compactness are essential requirements for weather forecasting models designed for operation on low-power embedded devices. This study developed Mixed-Input Residual Network (MIRNet), a compact temperature-forecasting deep neural network model. MIRNet integrates stacked bidirectional long short-term memory layers using concatenated 1-dimensional and 2-dimensional convolutional layers to improve model accuracy. MIRNet was trained and tested on two datasets: one, IfeData, comprising historical weather data from Ile-Ife, Nigeria and the other a standard weather forecasting dataset called the Jena dataset. Training was carried out using 100 epochs of data partitioned in the standard 80:20 ratio, with an adaptable learning rate strategy. The model was tested for Nth-hour-ahead prediction for 1N24; where N are natural numbers, and performance quantified using metrics such as mean absolute percentage error (MAPE) and mean square error (MSE). The model was also implemented on a Raspberry Pi 4 device with a 1.8 GHz 64-bit quad-core ARM Cortex-A72 processor. The model achieved a MSE of 1.00 x 10-3 on the IfeData dataset, and 1.23x10-4 on the Jena dataset for 1-hour ahead forecasting. This is currently the best verifiable result achieved on the Jena dataset by any prediction model globally. For Nth hour ahead forecasting, MIRNet achieved an MSE generally below 2.0x10-3 for all values of N on the standard Jena dataset. The MSE of MIRNet for N-sequential 1-hour ahead and single Nth hour predictions using the Jena dataset reveal quadratic and linear relationships with N respectively. The model compares favourably with existing models for multi-hour predictions. The developed model is compact and has good forecasting properties.

How good are different machine and deep learning models in forecasting the future price of metals? Full sample versus sub-sample

This study aims to forecast metal futures in commodity markets, including gold, silver, copper, platinum, palladium, and aluminium, using different machine and deep learning models. Prevalent models such as Stacked Long-Short Term Memory, Convolutional LSTM, Bidirectional LSTM, Support Vector Regressor, Extreme Gradient Boosting, and Gated Recurrent Unit are utilized. The model performance is assessed by multiple factors such as Root Mean Squared Error, Mean Absolute Error, and Mean Absolute Percentage Error. The study stands out by considering multiple metal commodity futures simultaneously, incorporating both Machine Learning and Deep Learning models, and conducting two sets of experiments with a full sample and subsample analysis. In addition, it uses different inputs of 30- and 60-days periods for robustness checks. Mean Absolute Percentage Error values suggest that different machine and deep learning models are efficient on prediction the future metal prices. However, the model performance varies significantly with the influence of metal choice, sample period, and inputs on prediction performance. Therefore, it suggests that constructing a theory based on machine and deep learning models is challenging.

H3DM: A High-bandwidth High-capacity Hybrid 3D Memory Design for GPUs

Graphics Processing Units (GPUs) are widely used for modern applications with huge data sizes. However, the performance benefit of GPUs is limited by their memory capacity and bandwidth. Although GPU vendors improve memory capacity and bandwidth using 3D memory technology (HBM), many important workloads with terabytes of data still cannot fit in the provided capacity and are bound by the provided bandwidth. With a limited GPU memory capacity, programmers should handle the data movement between GPU and host memories by themselves, causing a significant programming burden. To improve programming ease, GPUs use a unified address space with the host that allows over-subscribing GPU memory, but this approach is not effective in terms of performance once GPUs encounter memory page faults. Many recent works have tried to remedy capacity and bandwidth bottlenecks using dense non-volatile memories (NVMs) and true-3D stacking. However, these works mainly focus on one bottleneck or do not provide a scalable solution that fits future requirements. In this paper, we investigate true-3D stacking of dense, low-power, and refresh-free non-volatile phase change memory (PCM) on top of state-of-the-art GPU configurations to provide higher capacity and bandwidth within the available area and power budget. The higher density and lower power consumption of PCM provide higher capacity through integrating more cells in each 3D layer and enabling stacking more layers. However, we observe that stacking more than six layers of pure-PCM memory violates the thermal constraint and severely harms the performance and power efficiency due to its higher write latency and energy. Further, it degrades the lifetime of GPU to less than one year. Utilizing a hybrid architecture that leverages the benefits of both DRAM and PCM memories has been widely studied by prior proposals; however, true-3D integration of such a hybrid memory architecture especially on top of state-of-the-art powerful GPU architecture has not been investigated yet. We experimentally demonstrate that by covering 80% of write requests in DRAM and eliminating refresh overhead, true-3D stacking of eight 32GB layers of PCM along with two 8GB layers of DRAM is possible resulting in a total of 272GB memory capacity. Based on the explored design requirements, We propose a 3D high-bandwidth high-capacity hybrid memory (H3DM) system utilizing a hybrid-3D (H3D)-aware remapping scheme to reduce expensive PCM writes to under 20% while avoiding DRAM refresh overhead. H3DM improves the performance up to 291% compared to the baseline GPU architecture while remaining within only 3% of an ideal case with DRAM-like access latency, on average. Moreover, by increasing the dataset size above the baseline GPU memory space, H3DM improves performance and power up to 648% and 87% compared to the baseline GPU architecture since it avoids expensive data transfers through off-chip communication links.

Characterization of program and erase speed of memory capacitors with nanolaminated HfO2/Al2O3 stacks for application in non-volatile memories

The programming and erase performance of atomic layer deposited HfO2/Al2O3 memory stacks is investigated depending on the thickness of the tunnel oxide and the annealing of the structures. The as-grown stacks require applying of voltage pulses with duration of about 10-3 s for a measurable electron trapping (programming). The time dependence of the electron trapping process is quite sharp initially - the increase of the charging time from 4×10-4 s to 4×10-3 s provides more than 80% of the saturation value of the trapped charge obtained at durations above 1 s. The annealing increases the time needed for the onset of the electron trapping. For the chosen voltage amplitudes the programming characteristics does not depend on the tunnel oxide thickness. The electron detrapping (erase) is observed even at durations of ~10-7 s, but with a low efficiency. The erase characteristics is more gradual then the program one, and shows some dependence on tunnel oxide and annealing. The complete erase is achieved for times comparable to those for programming the cell. Erase pulses with duration > 10-2 s results to accumulation of positive charge in the capacitors.

Data-model linkage prediction of tool remaining useful life based on deep feature fusion and Wiener process

Accurately predicting the tool remaining useful life (RUL) is critical for maximizing tool utilization and saving machining costs. Various physical model-based or data-driven prediction methods have been developed and successfully applied in different machining operations. However, many uncertain factors affect tool RUL during the cutting process, making it challenging to create a precise physical model to characterize the degradation of tool performance. The success of the purely data-driven technique depends on the amount and quality of the training samples, it does not consider the physical law of tool wear, and the interpretability of the prediction results is poor. This paper presents a data-model linkage approach for tool RUL prediction based on deep feature fusion and Wiener process to address the above limitations. A convolutional stacked bidirectional long short-term memory network with time-space attention mechanism (CSBLSTM-TSAM) is developed in the data-driven module to fuse the multi-sensor signals collected during the cutting process and then obtain the mapping relationship between signal features and tool wear values. In the physical modeling module, a three-stage tool RUL prediction model based on the nonlinear Wiener process is established by considering the evolution law of different wear stages and multi-layer uncertainty, and the corresponding probability density function is derived. The real-time estimated tool wear of the data-driven module is used as the observed value of the physical model, and the model parameters are dynamically updated by the weight-optimized particle filter (WOPF) algorithm under a Bayesian framework, thereby realizing the data-model linkage tool RUL prediction. Milling experiments demonstrate that the proposed method not only improves RUL prediction accuracy, but also has good generalization ability and robustness for prediction tasks under different working conditions.

Multidie 3-D Stacking of Memory Dominated Neuromorphic Architectures

Multidie 3-D Stacking of Memory Dominated Neuromorphic Architectures

3DL-PIM: A Look-Up Table Oriented Programmable Processing in Memory Architecture Based on the 3-D Stacked Memory for Data-Intensive Applications

3DL-PIM: A Look-Up Table Oriented Programmable Processing in Memory Architecture Based on the 3-D Stacked Memory for Data-Intensive Applications

(Invited) Hybrid Bonding from Hype to Reality: How Will Hybrid Bonding Dtw Pave the Way to Large Set of Applications?

Packaging and 3D integration have grown in importance as one of the key technology enablers to compensate for the slowing of two-dimensional scaling. Today, advanced 3D SoC are calling for fine interconnect where microbumps are hitting manufacturing limitation at 10µm pitch and beyond. Hybrid bonding technology is a good candidate for high interconnect density and connects dies in packages using tiny copper-to-copper connections, as opposed to bumps. This technology is already mature for wafer to wafer hybrid bonding and is expected to open new opportunities for die to wafer hybrid bonding, providing the same downscaling in pitch as well as additional flexibility and heterogeneity. This derived technology is of great interest for many applications such as chiplet, memory stacking, photonics with imagers, displays, optic transceivers and RF. However, there are still some hurdles to overcome to get this technology in full production. CEA-Leti has been involved in hybrid bonding process development for more than 20 years with industrial partners in wafer to wafer and today in die to wafer. This presentation will discuss about: Limitations of the technology: specific process challenges, interconnect alignment, design and architecture,...What needs to be developed to pass the hurdle: known good die, throughput & yield,...Key achievements / performances and roadmap at CEA-Leti on this promising technology

(Invited) Die to Wafer Direct Bonding: Overview of Processes for Optoelectronic and 3D at CEA

Direct wafer bonding of wafers is nowadays well established. However, direct bonding of dies to wafers can result in innovative devices for optoelectronic or 3D applications (Memory, High Performance Computing...). The paper will focus on specific fundamental mechanisms involved in direct bonding, III/V die-to-wafer bonding and copper hybrid bonding. Die-To-Wafer (DTW) direct hybrid bonding, using copper/oxide mixed interfaces [1] is foreseen by major microelectronics companies as essential for future memory or HPC stacks and as a support to extend Moore’s Law. It indeed increases functionalities per surface unit by optimizing space in three directions: different dies embedding specific technologies coming from different wafers can closely be assembled in one place[2] [3]. However, there are specific integration challenges such as keeping delicate surfaces clean from any degradation or particle contamination after dicing and during die handling and stacking. We describe here important progresses concerning dies preparation as well as alignment with a SET NEO HB die bonder. DTW bonding with 5µm interconnection pitch was demonstrated by CEA with an alignment of 1µm[4]. The capillary approach used during die self-alignment on wafer will notably be discussed [5]. Integrated transmitters incorporating lasers and modulators on silicon are of prime importance for communications. They are at the same time most challenging to manufacture due to a need for hybrid III-V integration [6]. In order to use III-V materials in low cost silicon platforms, direct bonding is of great interest due to difficulties encountered when growing III-V hetero-epitaxial layers directly onto silicon (presence then of anti-phase boundaries, dislocations and so on). Photonic demonstration devices are often processed using as a template a III-V stack on a III-V wafer that is transferred thanks to direct bonding onto a silicon wafer [7]. The potential low cost model of silicon photonics is however based on making full use of the surface of a 200 or 300 mm diameter SOI photonic wafer. III-V wafer direct bonding is thus not suitable for two reasons. First, the maximum diameter available for III-V wafers is up to now limited to 150 mm, while Si-based wafers are nowadays typically 300 mm in diameter. Second, the III-V stack is needed only in the emitter and receiver areas that cover only a very small fraction of the overall device area. Therefore, most of the blanket III-V stack is wasted when removing, thanks to patterning, materials in unwanted area. To overcome this wafer diameter mismatch and limit the loss of a very expensive starting material, a collective die direct bonding seems therefore well adapted. The idea is to collectively bond III-V chips only where there are needed and over the full photonic SOI surface. Until now, only a few teams focused on such an approach. Collective die bonding with an adhesive carrier wafer has however already been demonstrated [8]. At CEA, after exploring a bonding process based on a silicon die wafer holder [9], we describe here an alternative solution using a tape on a dicing frame (see figure 1). As before, throughput is favoured over die placement accuracy. This is well adapted as the transferred layer can be resized by photolithography and etching after the die substrate removal. Furthermore, this approach with a “flexible” tape presents the great advantage of being extremely tolerant regarding die thicknesses variations and enables the use of multiple die wafer sources. We will describe such DTW bonding processes with InP materials. 3 x 3 mm² dies will be used as standard die size for bonding performance evaluation.[1]L. Sanchez et al, ECTC, pp.1960-1964[2] A. Bond et al, ECTC 2022, pp.168-176[3] A. Jouve et al, ECTC 2021, pp.225-234[4] E. Bourjot et al, ECTC 2021, pp 470-475[5]J. Berthier et al, Journal of Applied Physics, 2010, 108, 054905[6] B. Szelag et al, IEEE Journal of Selected Topics in Quantum Electronics, 2019, 25, 8201210[7] J. Durel at al, IEEE IEDM 2016[8] L. Xianshu et al, Journal of Selected Topics in Quantum Electronics, 2016, 22, 8200612[9] L. Sanchez, ECS transactions, 2018, 86, 223-231 Figure 1

A content-based recommender system using stacked LSTM and an attention-based autoencoder

Recommender systems (RS) are popular in many areas, such as movies, music, news, books, research articles, search queries, and social tagging. Proprietary recommender systems are used by e-commerce websites like eBay, Amazon, and Alibaba to better match customers with products they are likely to purchase. This study suggests a recommender system that combines stacked long short-term memory (LSTM) and an attention-based autoencoder. This system would be used in a self-supervised learning paradigm, and the Amazon product datasets were used to run simulations. The results showed that the proposed method is more accurate, uses less computing power, and can be used on a large scale. No matter how big the data is, it can capture low-dimensional fixed latent representations and use the bare minimum of information already in the items, whether they are new or old. Several evaluation metrics show that the method works and that the cold start problem has been solved.

Modeling the impact of incomplete conformality during atomic layer processing

Atomic layer processing (ALP) is a modern fabrication technique for the deposition or etching of materials, which provides precise control of film thickness, composition, and conformality on a nanometer scale. This makes it crucial for the fabrication of high aspect ratio (HAR) structures, such as 3D NAND memory stacks, as its self-limiting nature provides enhanced conformality compared to traditional processes. However, as the number of NAND stacks grows and the aspect ratio continues to increase, deviations from full conformality can often occur due to precursor desorption from the surface. In this regard, a model for surface coverage during ALD in the presence of desorption, leading to incomplete conformality, has been developed and implemented in process simulation frameworks. This work is an extension of our previous research which concentrated on developing an accurate modeling approach for ALD in HAR structures (L.Aguinsky et al., Solid State Electron. 201, 2023). The model combines existing Knudsen diffusion and Langmuir kinetics methods and includes the Bosanquet formula for gas-phase diffusivity and reaction reversibility. It has been incorporated into academic and commercial level-set-based topography simulators. The parameters for the model have been calibrated using published results for the ALD of Al2O3 from trimethylaluminum (TMA) and H2O in HAR geometries. The temperature dependence of the H2O step is likewise analyzed, revealing an activation energy of 0.178eV, which is consistent with recent experiments. In the TMA step, the Bosanquet formula leads to improved accuracy, and the same parameter set is able to reproduce multiple experiments, demonstrating that the model parameters accurately capture reactor conditions. Finally, the developed model is combined with atomic layer etching (ALE) to simulate the controlled, conformal deposition of HfO2 inside HAR 3D NAND structures.