With advancements in three-dimensional integration technology, wafer stacking has become a critical process for enhancing semiconductor device performance in the post-Moore era. The reliability of interfacial electrical interconnections depends on the bonding overlay accuracy, which is now primarily limited by residuals at the 50 nm level. This study addresses the challenge of bonding residuals in high-precision wafer bonding, which arise from the coupled effects of wafer elastic deformation, clamping constraints, and bond wave propagation. Existing models often lack comprehensive multiphysics coupling or fail to establish a link between specific process parameters and residual formation, limiting their use in process optimization. Therefore, developing a high-fidelity coupled model is essential for understanding the residual generation mechanism and devising effective suppression strategies.

A multiphysics coupling analysis model was developed that comprehensively considers wafer anisotropy, clamping boundary effects, and bond wave propagation behavior. The framework integrates anisotropic thin-plate elasticity (incorporating crystal orientation transformation tensors), gas film dynamics (governed by a modified Reynolds equation with bonding stress), and contact mechanics (solved via the augmented Lagrangian method). Bond wave propagation is governed by an energy criterion at the wavefront, balancing effective bonding energy against strain energy and the work performed by gas film and mechanical contact pressures. A finite element model for the 300 mm wafer bonding process was developed, achieving submicron accuracy. Key numerical strategies included a staggered iterative scheme for updating the wavefront, bonding force, gas pressure, and structural deformation; adaptive time stepping based on residual variations; and stabilization damping to suppress rigid body motion. Model validity was confirmed through comparison between simulation predictions and experimental pattern wafer geometry measurements.

The simulation accurately captured the nonuniform bond wave propagation induced by wafer anisotropy. The stress and residual distributions exhibited a distinct fourfold symmetry consistent with the crystallographic orientation. Residuals were primarily concentrated near the wafer edge, with additional significant residuals observed at the center-consistent with previous reports. Experimental validation showed strong agreement between simulated and measured residual distribution patterns. Systematic parameter studies revealed that using a flat bond head reduced the 2-norm of the residual vector by 41% compared with a spherical head (0.63 μm vs. 1.07 μm). Employing a lower-stiffness material (polyethylene) for the bond head reduced the residual 2-norm by 23% compared with PEEK plastic. Moreover, an increase in the initial wafer gap correlated with a higher residual 3σ value.

This study establishes a robust multiphysics coupling and process co-optimization framework for high-precision wafer bonding. The proposed model effectively captures the combined effects of wafer anisotropy, gas film dynamics, and contact mechanics on residual formation, enabling high-fidelity simulation of the bonding process and quantitative analysis of key process parameters. The findings demonstrate that optimizing the bond head design-with a flat surface and low-stiffness material-and minimizing the initial wafer gap can significantly suppress bonding residuals. This work provides a theoretical basis and practical design guidelines for optimizing wafer bonding processes to achieve superior overlay accuracy.

A self-paced treadmill (SPT) is key human-robot interactive equipment for virtual reality, which can enable a user to walk at the intended speed by using a re-positioning technology. Realizing multimode interactions of SPTs is crucial for enriching their applications. However, existing studies only realizes a few interaction modes. To realize multimode interactions in self-paced treadmills, a novel multilayer control framework is proposed in this paper.

In this study, the control system is divided into two layers: the recognition layer and the control layer. First, a novel hybrid spatial-temporal graph convolutional neural network is proposed to realize user-independent human locomotion mode recognition based on plantar pressure insoles in the recognition layer. The proposed network separates the pressure and acceleration signals and dealt with them individually. It is also utilized to extract the natural spatial topology between pressure nodes. Long short-term memory layers are used to individually extract temporal-dependent features of pressure and acceleration signals and to fuse multimodal features for final recognition. A multilayer perceptron is utilized to map the fusion features to the locomotion modes. By extracting the natural spatial-temporal features of multimodal data during human locomotion, a high generalization capability of the recognition results can be expected. Second, control strategies for different locomotion modes are designed in the control layer according to the stability condition of different human locomotion modes. Meanwhile, a walking speed feedforward control strategy is proposed to re-position the user and ensure natural gaits for the walking mode. Variable gain control strategies are adopted to manipulate the acceleration for the running and back walking modes. A buffer control strategy is proposed to improve the stability during jump landing for the jumping mode. Then, a finite state machine is used to automatically switch the control strategies. The states are transited based on the recognition results.

1) The proposed locomotion mode recognition method was evaluated on a dataset that comprises eight subjects with five locomotion modes through the leave-one-subject-out cross validation. Then, it was compared with the convolutional neural network (CNN) and domain-adversarial neural network (DANN). Experimental results indicated that the mean and standard deviation classification accuracies of the CNN, DANN, and HSTGCN are (90.26±8.54)%, (97.71±3.60)%, and (97.37±1.40)%, respectively. These results validated that the proposed method can achieve high generalization capability without any dependency on the data of target subjects. Hence, the burden of repeated data collection and network training was reduced. 2) Based on the recognition results, experiments on the multi-locomotion mode human-robot interaction were conducted using a finite state machine. Experimental results indicated that a user can freely change the locomotion modes on the treadmill, and the balance was not significantly affected by the treadmill acceleration.

The proposed framework can automatically combine the recognition results with the treadmill control and can realize the control of multi-locomotion mode human-robot interactions. Further, experimental results validate that the proposed multilayer control strategy can achieve a stable and smooth multi-locomotion mode human-robot interaction, ensure natural gaits and posture stability of the user, and meet the requirements of multi-locomotion mode human-robot interactions for self-paced treadmills.

The feedforward controller is crucial to achieving nano-level motion accuracy for the lithography wafer stage under high acceleration and deceleration conditions. Traditional 4-order feedforward is widely used to control precision motion systems because of its intuitive physical meaning and simple parameter tuning. However, its capacity to fit the inverse model is inadequate, and it is difficult to eliminate the repetitive error caused by the input trajectory. Therefore, a feedforward control architecture using the 4-order feedforward and an extra rational fraction compensator is proposed.

In this study, the input signal of the compensator is the higher-order derivative of the reference trajectory, and the numerator and denominator of the compensator use the delay unit as the basis function. Therefore, obtaining the unknown parameters of the basis function is crucial to the design. This paper proposes a data-driven iterative parameter tuning strategy for the compensation controller. The difficulty is that the tuning problem is a nonconvex optimization problem, making global parameter optimization challenging. This paper uses the relevant rules of system identification to address the issue at hand. The purpose of adding compensatory feedforward is to eliminate the residual error after using the 4-order feedforward, which is equivalent to achieving a zero-generalized error. Since the generalized error has a linear connection with the compensator parameters, the original nonconvex optimization problem is successfully transformed into a convex problem by minimizing the 2-norm of the generalized error. Through the above transformation, the global optimal point is obtained by the Gauss—Newton method, and the step size condition for ensuring iterative convergence is provided. In addition, the gradient and Hessian matrix of the objective function need to be incorporated into the parameter updating law, even though their exact values are difficult to obtain. This paper derives their unbiased estimates using two impulse response experiments and 2 trajectory tracking experiments.

The proposed method was applied to the wafer stage of the lithography machine, and the experiment showed the following results: (1) Using the proposed method to tune three compensation controllers with different orders, their error 2-norm almost converged after five iterations. (2) After adding compensation feedforward, the acceleration and deceleration phase errors were reduced from ±35 nm to ±10 nm; the constant velocity phase error was almost equal to the positioning error, and its trajectory tracking effect was very close to that of iterative learning control (ILC) compensation. (3) Compared with the existing compensation controller parameter tuning method, the maximum moving average and moving standard deviation at velocity phase of the proposed method were smaller, and the lower the compensator order, the more obvious the advantage. (4) After changing trajectory, the proposed compensator could still achieve a better control effect than ILC compensation.

The above experiments verify the convergence performance of the proposed parameter tuning algorithm. It is shown that the proposed feedforward compensation architecture can effectively eliminate the residual repetition error of the 4-order feedforward; simultaneously, it can adapt to variable trajectories. In addition, compared to the current compensator tuning result, this method can achieve a superior trajectory tracking control effect while using a low-order compensation controller.