FedTP: Federated Learning by Transformer Personalization.

Federated learning is an emerging learning paradigm where multiple clients collaboratively train a machine learning model in a privacy-preserving manner. Personalized federated learning extends this paradigm to overcome heterogeneity across clients by learning personalized models. Recently, there have been some initial attempts to apply transformers to federated learning. However, the impacts of federated learning algorithms on self-attention have not yet been studied. In this article, we investigate this relationship and reveal that federated averaging (FedAvg) algorithms actually have a negative impact on self-attention in cases of data heterogeneity, which limits the capabilities of the transformer model in federated learning settings. To address this issue, we propose FedTP, a novel transformer-based federated learning framework that learns personalized self-attention for each client while aggregating the other parameters among the clients. Instead of using a vanilla personalization mechanism that maintains personalized self-attention layers of each client locally, we develop a learn-to-personalize mechanism to further encourage the cooperation among clients and to increase the scalability and generalization of FedTP. Specifically, we achieve this by learning a hypernetwork on the server that outputs the personalized projection matrices of self-attention layers to generate clientwise queries, keys, and values. Furthermore, we present the generalization bound for FedTP with the learn-to-personalize mechanism. Extensive experiments verify that FedTP with the learn-to-personalize mechanism yields state-of-the-art performance in the non-IID scenarios. Our code is available online https://github.com/zhyczy/FedTP.

Access to Disentangled Ultrahigh Molecular Weight Polyethylene via a Binuclear Synergic Effect.

Disentangled ultrahigh molecular weight polyethylene (dis-UHMWPE) has excellent processability but can be achieved under extreme conditions. Herein, we report ethylene polymerization with the binuclear half-sandwich scandium complexes C1-Sc2 and C2-Sc2 to afford UHMWPE. C1-Sc2 bearing a short linker shows higher activity and gives higher molecular weight PEs than C2-Sc2 containing a flexible spacer and the mononuclear Sc1 . Strikingly, all UHMWPEs isolated from C1-Sc2 under broad temperature range (25-120 °C) and wide ethylene pressures (2-13 bar) feature very low degree of entanglement as proved by rheological test, DSC annealing study and SEM. These dis-UHMWPEs are facilely mediated solid-state-process at 130 °C and their tensile strength and modulus reach up to 149.2 MPa and 1.5 GPa, respectively. DFT simulations reveal that the formation of dis-UHMWPE is attributed to the binuclear synergic effect and the agostic interaction between the active center and the growing chain.

Shape Prediction of the Sheet in Continuous Roll Forming Based on the Analysis of Exit Velocity.

Continuous roll forming (CRF) is a new technology that combines continuous forming and multi-point forming to produce three-dimensional (3D) curved surfaces. Compared with other methods, the equipment of CRF is very simple, including only a pair of bendable work rolls and the corresponding shape adjustment and support assembly. By controlling the bending shapes of the upper and lower rolls and the size of the roll gap during forming, double curvature surfaces with different shapes can be produced. In this paper, a simplified expression of the exit velocity of the sheet is provided, and the formulas for the calculation of the longitudinal curvature radius are further derived. The reason for the discrepancy between the actual and predicted values of the longitudinal radius is deeply discussed from the perspective of the distribution of the exit velocity. By using the response surface methodology, the effects of the maximum compression ratio, the sheet width, the sheet thickness, and the transverse curvature radius on the longitudinal curvature radius are analyzed. Meanwhile, the correction coefficients of the predicted formulas for the positive and negative Gaussian curvature surfaces are obtained as 1.138 and 0.905, respectively. The validity and practicability of the modified formulas are verified by numerical simulations and forming experiments.

Plastic Forming of Sandwich Panels and Numerical Analyses of the Forming Processes Based on Elastoplastic Equivalent Model.

This paper studies the plastic forming of sandwich panels and proposes a universal elastoplastic equivalent method suitable for sandwich panels. To verify the generality of the equivalent method, according to the different core structures, the cores of bi-directional trapezoidal sandwich (BTS) panels and aluminum foam sandwich (AFS) panels are equated to orthotropic and isotropic (special orthotropic) single-layer panels respectively. Through the finite element (FE) numerical simulation of the mesoscopic model of the sandwich panel, the elastoplastic constitutive relationship of the equivalent core model is established, and then the macroscopic equivalent model of the sandwich panel is established. The FE numerical simulation of plastic forming was carried out for the mesoscopic model and equivalent model of BTS panel and AFS panel, and plastic forming experiments were conducted for the sandwich panel through a multi-point forming (MPF) test machine. The results show that the relative errors of the section average stress at the same position of the equivalent model and the mesoscopic model of sandwich panels are all within 4%; compared with the experimental results, the equivalent model of the sandwich panel has high forming accuracy and small shape error, which verifies the high accuracy and generality of the equivalent method. Moreover, using the sandwich panel equivalent model effectively reduces the calculation time of the numerical simulation.