Search for: All records

Federated learning is a framework for distributed optimization that places emphasis on communication efficiency. In particular, it follows a client-server broadcast model and is appealing because of its ability to accommodate heterogene- ity in client compute and storage resources, non-i.i.d. data assumptions, and data privacy. Our contribution is to offer a new federated learning algorithm, FedADMM, for solving non-convex composite optimization problems with non-smooth regularizers. We prove the convergence of FedADMM for the case when not all clients are able to participate in a given communication round under a very general sampling model.

We report the first experimental demonstration of a vertical superjunction device in GaN. P-type nickel oxide (NiO) is sputtered conformally in 6μm deep n-GaN trenches. Sputter recipe is tuned to enable 1017 cm −3 level acceptor concentration in NiO, easing its charge balance with the 9×1016 cm −3 doped n-GaN. Vertical GaN superjunction p-n diodes (SJ-PNDs) are fabricated on both native GaN and low-cost sapphire substrates. GaN SJ-PNDs on GaN and sapphire both show a breakdown voltage (BV) of 1100 V, being at least 900 V higher than their 1-D PND counterparts. The differential specific on-resistance (RON,SP) of the two SJ-PNDs are both 0.3mΩ⋅ cm 2 , with the drift region resistance (RDR,SP) extracted to be 0.15mΩ⋅ cm 2 . The RON,SP∼BV trade-off is among the best in GaN-on-GaN diodes and sets a new record for vertical GaN devices on foreign substrates. The RDR,SP∼BV trade-off exceeds the 1-D GaN limit, fulfilling the superjunction functionality in GaN.

This work considers the problem of learning the Markov parameters of a linear system from ob- served data. Recent non-asymptotic system identification results have characterized the sample complexity of this problem in the single and multi-rollout setting. In both instances, the number of samples required in order to obtain acceptable estimates can produce optimization problems with an intractably large number of decision variables for a second-order algorithm. We show that a randomized and distributed Newton algorithm based on Hessian-sketching can produce ε-optimal solutions and converges geometrically. Moreover, the algorithm is trivially parallelizable. Our re- sults hold for a variety of sketching matrices and we illustrate the theory with numerical examples.

Twin boundary (TB) strengthening in nanotwinned metals experiences a breakdown below a critical spacing at which softening takes over. Here, we survey a range of nanotwinned materials that possess different stacking fault energies (SFEs) and understand the TB strengthening limit using atomistic simulations. Distinct from Cu and Al, the nanotwinned, ultralow SFE materials (Co, NiCoCr, and NiCoCrFeMn) intriguingly exhibit a continuous strengthening down to a twin thickness of 0.63 nm. Examining dislocation slip mode and deformation microstructure, we find the hard dislocation modes persist even when reducing the twin boundary spacing to a nanometer regime. Meanwhile, the soft dislocation mode, which causes detwinning in Cu and Al, results in phase transformation and lamellar structure formation in Co, NiCoCr, and NiCoCrFeMn. This study, providing an enhanced understanding of dislocation mechanism in nanotwinned materials, demonstrates the potential for controlling mechanical behavior and ultimate strength with broadly tunable composition and SFE, especially in multi-principal element alloys.

Learning a dynamical system from input/output data is a fundamental task in the control design pipeline. In the partially observed setting there are two components to identification: parameter estimation to learn the Markov parameters, and system realization to obtain a state space model. In both sub-problems it is implicitly assumed that standard numerical algorithms such as the singular value decomposition (SVD) can be easily and reliably computed. When trying to fit a high-dimensional model to data, even computing an SVD may be intractable. In this work we show that an approximate matrix factorization obtained using randomized methods can replace the standard SVD in the realization algorithm while maintaining the finite-sample performance and robustness guarantees of classical methods.