arXiv subjects

Zhuonan Lin

Publications and source records attributed to Zhuonan Lin.

5 recordsLinked to original sources

arXiv2024

Periodic phase diagrams in micromagnetics with an eigenvalue solver

This work introduces an approach to compute periodic phase diagram of micromagnetic systems by solving a periodic linearized Landau-Lifshitz-Gilbert (LLG) equation using an eigenvalue solver with the Finite Element Method formalism. The linear operator in the eigenvalue problem is defined as a function of the periodic phase shift wave vector. The dispersion diagrams are obtained by solving the eigenvalue problem for complex eigen frequencies and corresponding eigen states for a range of prescribed wave vectors. The presented approach incorporates a calculation of the periodic effective field, including the exchange and magnetostatic field components. The approach is general in that it allows handling 3D problems with any 1D, 2D, and 3D periodicities. The ability to calculated periodic diagrams provides insights into the spin wave propagation and localized resonances in complex micromagnetic structures.

cond-mat.mtrl-sci↗

arXiv2024

Parameter Efficient Reinforcement Learning from Human Feedback

While Reinforcement Learning from Human Feedback (RLHF) effectively aligns pretrained Large Language and Vision-Language Models (LLMs, and VLMs) with human preferences, its computational cost and complexity hamper its wider adoption. To alleviate some of the computational burden of fine-tuning, parameter efficient methods, like LoRA were introduced. In this work, we empirically evaluate the setup of Parameter Efficient Reinforcement Learning from Human Feedback (PE-RLHF) that leverages LoRA fine-tuning for Reward Modeling, and Reinforcement Learning. We benchmark the PE-RLHF setup on six diverse datasets spanning summarization, harmless/helpful response generation, UI automation, and visual question answering in terms of effectiveness of the trained models, and the training resources required. Our findings show, for the first time, that PE-RLHF achieves comparable performance to RLHF, while significantly reducing training time (up to 90% faster for reward models, and 30% faster for RL), and memory footprint (up to 50% reduction for reward models, and 27% for RL). We provide comprehensive ablations across LoRA ranks, and model sizes for both reward modeling and reinforcement learning. By mitigating the computational burden associated with RLHF, we push for a broader adoption of PE-RLHF as an alignment technique for LLMs and VLMs.

cs.LG↗

arXiv2022

Linearized frequency domain Landau-Lifshitz-Gilbert equation formulation

We present a general finite element linearized Landau-Lifshitz-Gilbert equation (LLGE) solver for magnetic systems under weak time-harmonic excitation field. The linearized LLGE is obtained by assuming a small deviation around the equilibrium state of the magnetic system. Inserting such expansion into LLGE and keeping only first order terms gives the linearized LLGE, which gives a frequency domain solution for the complex magnetization amplitudes under an external time-harmonic applied field of a given frequency. We solve the linear system with an iterative solver using generalized minimal residual method. We construct a preconditioner matrix to effectively solve the linear system. The validity, effectiveness, speed, and scalability of the linear solver are demonstrated via numerical examples.

physics.app-ph↗

arXiv2022

Eigenvalue-based micromagnetic analysis of switching in spin-torque-driven structures

We present an eigenvalue-based approach for studying the magnetization dynamics in magnetic nanostructures driven by spintronic excitations, such as spin transfer torque and spin orbit torque. The approach represents the system dynamics in terms of normal oscillation modes (eigenstates) with corresponding complex eigenfrequencies. The dynamics is driven by a small number of active eigenstates and often considering just a single eigenstate is sufficient. We develop a perturbation theory that provides semi-analytical dynamic solutions by using eigenstates for the case in the absence of damping and spintronic excitations as a basis. The approach provides important insights into dynamics in such systems and allows solving several difficulties in their modeling, such as extracting the switching current in magnetic random access memories (MRAM) and understanding switching mechanisms. We show that the presented approach directly predicts the critical switching current, i.e., switching current for an infinite time. The approach also provides solutions for the switching dynamics allowing obtaining the switching current for a finite switching time, provided that the system symmetry is broken, e.g., by tilting the polarizer, so that switching by a finite pulse is possible.

physics.app-ph↗

arXiv2014

A rigorous two-dimensional model for the stripline ferromagnetic resonance response of metallic ferromagnetic films

In this work we constructed a two-dimensional numerical model for calculation of the stripline ferromagnetic resonance (FMR) response of metallic ferromagnetic films. We also conducted numerical calculations by using this software. The calculations demonstrated that the eddy current contribution to the FMR response decreases with a decrease in the stripline width. The most important manifestations of the conductivity (eddy current) effect are excitation of the higher-order standing spin waves across the film thickness in the materials for which the standing spin wave peaks would be absent in cavity FMR measurements and strong dependence of the off-resonance series conductance of the stripline on the stripline width. Whereas the contribution of the eddy currents to the stripline FMR response can be very significant, because wide striplines (100nm+) are conventionally used for the FMR measurements, it is negligible in the case of excitation of spin waves, just because very narrow stripline transducers (0.5-5micron wide) are required in order to excite spin waves in metallic ferromagnetic films in a noticeable frequency/applied field range.

cond-mat.mtrl-sci↗