The integration of machine learning (ML) and semiconductor technology offers new approaches for accelerating device design and optimization. The adoption of machine learning in semiconductor device modeling remains hindered by three fundamental limitations: high data dependency, poor generalization ability on small datasets, and limited scalability across device variants. While ML offers the potential to accelerate device design, conventional models require large technology computer-aided design (TCAD) generated datasets and fail to transfer knowledge between technologies. To overcome these challenges, we propose a transfer learning framework that enables efficient knowledge reuse from pre-trained models to downstream modeling tasks. Using the Drain extended Metal Oxide Semiconductor (DEMOS) device as a representative testbed, we demonstrate that the proposed method significantly reduces data requirements while enhancing generalization capabilities under novel device conditions. Crucially, the framework eliminates the need for developing individual models for each device variant, establishing a scalable and practical methodology for semiconductor device modeling. The proposed method was validated through silicon wafer fabrication. Research results demonstrate that it delivers high predictive accuracy while substantially reducing development costs and cycle times, thereby establishing a practical pathway for machine-learning-augmented semiconductor design flows.

Laser powder bed fusion (LPBF) is an additive manufacturing process capable of producing intricate structures with high accuracy. Despite this capability, it struggles to achieve the required reliability for mass production-specifically the stability of a production run and repeatability across multiple runs. Parameter optimization, which adjusts process parameters to regulate a specific quantity of interest (QoI), is a crucial means of quality control. Existing methods, however, have not adequately addressed both random and systematic factors in the LPBF process. The stochastic nature of the process is often neglected under the assumption that identical parameter inputs will consistently yield the same QoI. This deviates from reality and is not intended to reduce potential variations in the QoI. Moreover, many studies do not incorporate the systematic neighboring effects between scan tracks into their optimization, so process reliability cannot be guaranteed. To address this issue, this study focuses on optimizing the probability distribution of the QoI. The key idea is not only to increase the likelihood of achieving the ideal QoI but also to reduce its variance. This is achieved by uncertainty-aware modeling and optimization of the LPBF process using machine learning. Specifically, the problem is formulated as maximizing the posterior distribution of scan parameters given an ideal QoI sequence and historical manufacturing data, yielding a large-scale constrained optimization problem. A stochastic, distributed, gradient-based method is proposed to solve this problem, where a coarse-to-fine strategy plays a critical role in accelerating convergence. A case study is then conducted to stabilize the melt pool volume by optimizing laser powers. The solutions are verified in a calibrated finite element-based simulation environment, in which the variations of the melt pool volume are effectively reduced both within a single run and across multiple runs. The implementation of our method is available at https://github.com/qihangGH/uncertainty_aware_param_optim_for_AM. Note to Practitioners-This paper is motivated by the critical reliability issues in laser powder bed fusion (LPBF) additive manufacturing, requiring that the process is stable within a run and repeatable across runs. Parameter optimization is an effective way to control a certain quantity of interest (QoI) to regulate the LPBF process. Existing methods, however, often ignore the stochastic variations and neighboring effects in the LPBF process. The objective solely focuses on minimizing the error between the predicted QoI and ideal one, under the assumption that the environment is deterministic. The optimized parameters, therefore, do not guarantee reliability under the influences of random factors. This paper intends to improve the reliability through uncertainty-aware modeling and optimization of the LPBF process. It focuses on optimizing the probability distribution of the QoI, encouraging the ideal QoI not only more likely to appear but also appear with less fluctuations. By incorporating important neighboring effects and bounded and smooth input constraints, a large-scale constrained optimization problem is formulated. To solve this problem efficiently, a stochastic and distributed gradient-based optimization algorithm is proposed, and a coarse-to-fine strategy is employed to accelerate the optimization process. For validation, a case study is conducted to stabilize the melt pool volume by regulating laser powers. The optimized laser powers are verified in a finite element-based simulation environment. Quantitative and qualitative results show that the variations of the melt pool volume within a single run and across repeated runs are reduced, indicating the potential of the uncertainty-aware optimization for enhancing the reliability of the LPBF process.

Digital light processing (DLP) is a popular additive manufacturing technology that uses light irradiation to fabricate 3-D devices via a projector to achieve laser-sensitive resin curing. However, the performance and reliability of DLP can be affected by internal defects such as printing errors and the accumulation of residual stress. Existing defect detection methods rely on monitoring the printed parts, which leads to resource wastage and struggles to effectively handle imbalanced defect data. In this article, we propose a defect detection method called dual neural network, which involves detecting defects in materials before the printing process to prevent resource wastage and serious consequences. Specifically, to handle the highly imbalanced class distribution problem in online DLP defect detection, dual neural network utilizes a domain learner and balance learner to effectively balance the information of the minority class and learn the generalization knowledge from the imbalanced defect dataset. Experimental results demonstrate the effectiveness of our proposed method, which has also been applied to real-world production equipment successfully.

The quality of the dataset is critical to the performance of neural networks for error prediction in 3-D printing. In order to enlarge the dataset, we propose a customized two-stage framework, cascaded cross-modality generative adversarial networks (CCMGANs), for generating dental crown meshes in an unsupervised manner. At the first stage, a displacement map-guided generative adversarial network (GAN) is used to generate coarse meshes with diverse shapes. At the second stage, fine-grained details are added to the coarse meshes using an image-based GAN. Unlike previous work that integrates a differentiable renderer into the mesh deformation process directly, we adopt a two-step strategy. First, we use a depth image refinement module to achieve the domain transformation from the rendered depth images of the generated meshes to those of the real ones. Then, we propose a mesh refinement module to optimize the coarse meshes in an image-supervised manner. To alleviate the self-intersection problem, we propose a loss to penalize the distances of point pairs in self-intersection regions. Experimental results show that our method is able to generate highly realistic meshes and outperforms the state-of-the-art point cloud generation method TreeGCN in terms of the metrics FDD, MMD-CD, MMD-EMD, and COV-EMD. Furthermore, we utilize the generated data to augment the original dataset, and demonstrate that the generated data can effectively improve the accuracy of the error prediction task in 3-D printing.

Implantable brain–computer interface (BCI) devices are an effective tool to decipher fundamental brain mechanisms and treat neural diseases. However, traditional neural implants with rigid or bulky cross-sections cause trauma and decrease the quality of the neuronal signal. Here, we propose a MEMS-fabricated flexible interface device for BCI applications. The microdevice with a thin film substrate can be readily reduced to submicron scale for low-invasive implantation. An elaborate silicon shuttle with an improved structure is designed to reliably implant the flexible device into brain tissue. The flexible substrate is temporarily bonded to the silicon shuttle by polyethylene glycol. On the flexible substrate, eight electrodes with different diameters are distributed evenly for local field potential and neural spike recording, both of which are modified by Pt-black to enhance the charge storage capacity and reduce the impedance. The mechanical and electrochemical characteristics of this interface were investigated in vitro. In vivo, the small cross-section of the device promises reduced trauma, and the neuronal signals can still be recorded one month after implantation, demonstrating the promise of this kind of flexible BCI device as a low-invasive tool for brain–computer communication.

Anti-saturation attack (ASA) strategy is vital for the survival of a warship group, and attracts the focus of many researchers. In this paper, the dynamics of ASA is formulated as a Markov Decision Process (MDP) with an enhanced states space since those characters are involved, such as the formation and detection and interception areas of warship group. A reinforcement learning method (Double Deep Q-leaning, DDQN) is developed to solve the problem and deal with the curse of dimensionality whereby the cost-to-go value is calculated by a marine engagement simulation system. A heuristic defense algorithm guided by field expert knowledge is designed for comparison. The experimental results show that the DDQN method performs better in anti-saturation attack scenarios.

This paper presents a dual mode neural probe with enhanced microstructure electrode to improve the electrical stimulation and electrophysiology recording performance. The cylinder and cube microstructures with the depth of 1.5 µm are realized by inductively coupled plasma (ICP) etching. Electrochemical characterization indicates the electrode with cube microstructure obtains the largest charge storage capacity (CSC) compared to cylinder and planar one. Meanwhile, the impedance of cube electrode decreases down to 0.9 kΩ at 1 kHz after platinum black modification. The animal experiments reveal the neurons in primary motor cortex would be activated by electrical stimulation. Meanwhile, the electrode with cube microstructure is qualified to neural signal recording in the deep brain. This dual mode neural probe with enhanced microstructure is dedicated to neuroscience research and neurological disease treatment.

The input-output function of neural networks is complicated due to the huge number of neurons and synapses, and some high-density implantable electrophysiology recording tools with a plane structure have been developed for neural circuit studies in recent years. However, traditional plane probes are limited by the record-only function and inability to monitor multiple-brain regions simultaneously, and the complete cognition of neural networks still has a long way away. Herein, we develop a three-dimensional (3D) high-density drivable optrode array for multiple-brain recording and precise optical stimulation simultaneously. The optrode array contains four-layer probes with 1024 microelectrodes and two thinned optical fibers assembled into a 3D-printed drivable module. The recording performance of microelectrodes is optimized by electrochemical modification, and precise implantation depth control of drivable optrodes is verified in agar. Moreover, in vivo experiments indicate neural activities from CA1 and dentate gyrus regions are monitored, and a tracking of the neuron firing for 2 weeks is achieved. The suppression of neuron firing by blue light has been realized through high-density optrodes during optogenetics experiments. With the feature of large-scale recording, optoelectronic integration, and 3D assembly, the high-density drivable optrode array possesses an important value in the research of brain diseases and neural networks.

The electromagnetic properties at the interface of heterostructure are sensitive to the interfacial crystal structure and external field. For example, the two-dimensional magnetic states at the interface of LaAlO3/SrTiO3 are discovered and can further be controlled by electric field. Here, we study two types of heterostructures, TiO2/PbTiO3 and SrTiO3/PbTiO3, using first-principle electronic structure calculations. We find that the ferroelectric polarization discontinuity at the interface leads to partially occupied Ti 3d states and the magnetic moments. The magnitude of the magnetic moments and the ground-state magnetic coupling are sensitive to the polarization intensity of PbTiO3. As the ferroelectric polarization of PbTiO3 increases, the two heterostructures show different magnetic ordering that strongly depends on the electron occupation of the Ti t2g orbitals. For the TiO2/PbTiO3 interface, the magnetic moments are mostly contributed by degenerated d yz/d xz orbitals of interfacial Ti atoms and the neighboring interfacial Ti atoms form ferromagnetic coupling. For SrTiO3/PbTiO3 interface, the interfacial magnetic moments are mainly contributed by occupied d xy orbital because of the increased polarization intensity, and as the electron occupation increases, there exists a transition of the magnetic coupling between neighboring Ti atoms from ferromagnetism to antiferromagnetism via the superexchange interaction. Our study suggests that manipulating the polarization intensity is one effective way to control interfacial magnetic ordering in the perovskite oxide heterostructures.

Ferroelectricity is usually found in compound materials composed by different elements. Here, based on first-principles calculations, we reveal the first example of spontaneous electrical polarization and ferroelectricity in stable two-dimensional elemental materials: elemental Group-V (As, Sb, and Bi) monolayers. The polarization is due to the spontaneous lattice distortion with atomic layer buckling. Interestingly, for Bi monolayer, apart from the ferroelectric phase, we find that it can also host an antiferroelectric phase. The Curie temperatures of these elemental materials can be higher than room temperature, making them promising for realizing ultrathin ferroelectric devices of broad interest.

Oxide heterostructure interfaces create a platform to induce intriguing electric and magnetic functionalities for possible future devices. A general approach to control growth and interface structure of oxide heterostructures will offer a great opportunity for understanding and manipulating the functionalities. Here, it is reported that an electrostatic force, originating from a polar ferroelectric surface, can be used to drive oxide heteroepitaxy, giving rise to an atomically sharp and coherent interface by using a low-temperature solution method. These heterostructures adopt a fascinating selective growth, and show a saturation thickness and the reconstructed interface with concentrated charges accumulation. The ferroelectric polarization screening, developing from a solid-liquid interface to the heterostructure interface, is decisive for the specific growth. At the interface, a charge transfer and accumulation take place for electrical compensation. The facile approach presented here can be extremely useful for controlling oxide heteroepitaxy and producing intriguing interface functionality via electrostatic engineering.

Epitaxial heterostructures are of particular interest owing to their fascinating properties for wide applications in energy, environment and electronic devices. The understanding of epitaxial growth in solution phase, however, remains a fundamental challenge to realize the rational synthesis of heterostructures. Here we report that anatase TiO2 can epitaxially grow on the selective polar surface of single-crystal and single-domain ferroelectric PbTiO3 nanoplates. The interplay of ferroelectric polar surface and corresponding ion adsorptions on them has been revealed experimentally and theoretically to determine the epitaxial growth mode, giving rising to zero-dimensional (0D)/two-dimensional (2D) and 2D/2D heterostructures. A combination of experimental and theoretical calculations indicate that the resulting heterostructures adopt a polarization-dependent photocatalytic performance under visible light irradiation, including hole-based photodegradation and electron-based hydrogen evolution reaction (HER) of water splitting. Such findings allow investigation of the potential of ferroelectric polarization towards tuning epitaxial growth and functionality of heterostructures.

To explore high performance anodes remains a key challenge for rechargeable Na-ion batteries (NIBS). In this work, we predict that layered molybdenum dinitride (MoN2) could be a promising anode material for NIBS by first-principle calculations. The results show that each formula unit of MoN2 monolayer can accommodate four Na atoms, corresponding to a high specific capacity of 864 mA h g(-1). The average open circuit voltage is estimated to be 0.62 V, benefitting for a high open voltage of a full cell. The diffusion barrier for Na on MoN2 monolayer is quite small as 0.12 eV, indicating an excellent rate capability. Moreover, the Na atoms prefer to adsorb on MoN2 monolayer separately with each other, which is essentially important for remaining high performance during cycling. Our results show that the layered MoN2 has great potential to be an excellent anode material for NIBs. (C) 2017 Elsevier B.V. All rights reserved.

Reinforcement learning (RL)-based decoders in brain-machine interfaces (BMIs) interpret dynamic neural activity without patients' real limb movements. In conventional RL, the goal state is selected by the user or defined by the physics of the problem, and the decoder finds an optimal policy essentially by assigning credit over time, which is normally very time-consuming. However, BMI tasks require finding a good policy in very few trials, which impose a limit on the complexity of the tasks that can be learned before the animal quits. Therefore, this paper explores the possibility of letting the agent infer potential goals through actions over space with multiple objects, using the instantaneous reward to assign credit spatially. A previous method, attention-gated RL employs a multilayer perceptron trained with backpropagation, but it is prone to local minima entrapment. We propose a quantized attention-gated kernel RL (QAGKRL) to avoid the local minima adaptation in spatial credit assignment and sparsify the network topology. The experimental results show that the QAGKRL achieves higher successful rates and more stable performance, indicating its powerful decoding ability for more sophisticated BMI tasks as required in clinical applications.

As the downscaling of electronic devices continues, the problems of leakage currents and heat dissipation become more and more serious. To address these issues, new materials and new structures are explored. Here, we propose an interesting heterostructure made of ultrathin SnO layers on Si(001) surface. Our first-principle calculations show that a single layer of SnO on Si(001) surface is a semiconductor, but a bilayer SnO on the same surface is metallic. This metal-semiconductor dichotomy allows construction of single-2D-material-based electronic devices with low contact resistance and low leakage currents. In particular, due to the interaction between Sn and the Si substrate, the semiconducting monolayer-SnO/Si(001) has a highly anisotropic band structure with a much lighter hole effective mass along one direction than that of Si and most other 2D materials, indicating a high carrier mobility. Furthermore, by combining density functional theory and nonequilibrium Green's function method, we directly investigate the transport characteristics of a field effect transistor based on the proposed heterostructures, which shows very low contact resistance, negligible leakage current, and easy gate control at a compact channel length.

Reinforcement learning is an effective algorithm for brain machine interfaces (BMIs) which interprets the mapping between neural activities with plasticity and the kinematics. Exploring large state-action space is difficulty when the complicated BMIs needs to assign credits over both time and space. For BMIs attention gated reinforcement learning (AGREL) has been developed to classify multi-actions for spatial credit assignment task with better efficiency. However, the outliers existing in the neural signals still make interpret the neural-action mapping difficult. We propose an enhanced AGREL algorithm using correntropy as a criterion, which is more insensitive to noise. Then the algorithm is tested on the neural data where the monkey is trained to do the obstacle avoidance task. The new method converges faster during the training period, and improves from 44.63% to 68.79% on average in success rate compared with the original AGREL. The result indicates that the combination of correntropy criterion and AGREL can reduce the effect of the outliers with better performance when interpreting the mapping between neural signal and kinematics.

Reinforcement learning (RL)-based brain machine interfaces (BMIs) enable the user to learn from the environment through interactions to complete the task without desired signals, which is promising for clinical applications. Previous studies exploited Q-learning techniques to discriminate neural states into simple directional actions providing the trial initial timing. However, the movements in BMI applications can be quite complicated, and the action timing explicitly shows the intention when to move. The rich actions and the corresponding neural states form a large state-action space, imposing generalization difficulty on Q-learning. In this paper, we propose to adopt attention-gated reinforcement learning (AGREL) as a new learning scheme for BMIs to adaptively decode high-dimensional neural activities into seven distinct movements (directional moves, holdings and resting) due to the efficient weight-updating. We apply AGREL on neural data recorded from M1 of a monkey to directly predict a seven-action set in a time sequence to reconstruct the trajectory of a center-out task. Compared to Q-learning techniques, AGREL could improve the target acquisition rate to 90.16% in average with faster convergence and more stability to follow neural activity over multiple days, indicating the potential to achieve better online decoding performance for more complicated BMI tasks.

Sequential Monte Carlo estimation on point processes has been successfully applied to predict the movement from neural activity. However, there exist some issues along with this method such as the simplified tuning model and the high computational complexity, which may degenerate the decoding performance of motor brain machine interfaces. In this paper, we adopt a general tuning model which takes recent ensemble activity into account. The goodness-of-fit analysis demonstrates that the proposed model can predict the neuronal response more accurately than the one only depending on kinematics. A new sequential Monte Carlo algorithm based on the proposed model is constructed. The algorithm can significantly reduce the root mean square error of decoding results, which decreases 23.6% in position estimation. In addition, we accelerate the decoding speed by implementing the proposed algorithm in a massive parallel manner on GPU. The results demonstrate that the spike trains can be decoded as point process in real time even with 8000 particles or 300 neurons, which is over 10 times faster than the serial implementation. The main contribution of our work is to enable the sequential Monte Carlo algorithm with point process observation to output the movement estimation much faster and more accurately.

Sensory information, such as the tactile or proprioceptive signals, helps motor brain-machine interface (mBMI) work more naturally. Before applying sensory feedback, we need to explore if the neural activities are discriminative to different stimuli during a BMI task. Previous studies on the cortical discrimination are mainly focused on the rat whisker system. In this paper, we design a BMI task, where the freely behaving rat needs to maintain its position by poking a hole to receive the vibration stimulation on forepaws. Neural signals are collected synchronously from the corresponding sensory cortex when the rat is performing the task. Support Vector Machine (SVM) algorithm is implemented to evaluate the single trial classification of natural stimulus by neural ensembles. We find that the average classification accuracy reaches 80% over 3 rats, which suggests the vibration with different frequencies can be used as tactile feedback to the mBMI system.

Objective. The high-dimensional neural recordings bring computational challenges to movement decoding in motor brain machine interfaces (mBMI), especially for portable applications. However, not all recorded neural activities relate to the execution of a certain movement task. This paper proposes to use a local-learning-based method to perform neuron selection for the gesture prediction in a reaching and grasping task. Approach. Nonlinear neural activities are decomposed into a set of linear ones in a weighted feature space. A margin is defined to measure the distance between inter-class and intra-class neural patterns. The weights, reflecting the importance of neurons, are obtained by minimizing a margin-based exponential error function. To find the most dominant neurons in the task, 1-norm regularization is introduced to the objective function for sparse weights, where near-zero weights indicate irrelevant neurons. Main results. The signals of only 10 neurons out of 70 selected by the proposed method could achieve over 95% of the full recording's decoding accuracy of gesture predictions, no matter which different decoding methods are used (support vector machine and K-nearest neighbor). The temporal activities of the selected neurons show visually distinguishable patterns associated with various hand states. Compared with other algorithms, the proposed method can better eliminate the irrelevant neurons with near-zero weights and provides the important neuron subset with the best decoding performance in statistics. The weights of important neurons converge usually within 10-20 iterations. In addition, we study the temporal and spatial variation of neuron importance along a period of one and a half months in the same task. A high decoding performance can be maintained by updating the neuron subset. Significance. The proposed algorithm effectively ascertains the neuronal importance without assuming any coding model and provides a high performance with different decoding models. It shows better robustness of identifying the important neurons with noisy signals presented. The low demand of computational resources which, reflected by the fast convergence, indicates the feasibility of the method applied in portable BMI systems. The ascertainment of the important neurons helps to inspect neural patterns visually associated with the movement task. The elimination of irrelevant neurons greatly reduces the computational burden of mBMI systems and maintains the performance with better robustness.