NeuroPDE: A Neuromorphic PDE Solver Based on Spintronic and Ferroelectric Devices

Siqing Fu İD , Lizhou Wu ∗ İD , Tiejun Li ∗ İD , Chunyuan Zhang İD , Sheng Ma İD , Jianmin Zhang İD , Yuhan Tang İD and Jixuan Tang İD · 2025

Read NeuroPDE: A Neuromorphic PDE Solver Based on Spintronic and Ferroelectric Devices by Siqing Fu İD , Lizhou Wu ∗ İD , Tiejun Li ∗ İD , Chunyuan Zhang İD , Sheng Ma İD , Jianmin Zhang İD , Yuhan Tang İD and Jixuan Tang İD (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2507-04677.txt.

Introduction

Siqing Fu and co-authors design a brain-like chip solver for partial differential equations in this 2025 paper. Such equations govern heat, flow, and fields across science. Random walk methods solve them well but need true randomness to shine.

Standard computer chips fake randomness badly and waste energy doing it. Their NeuroPDE design uses spintronic neurons that fire with real chance. Ferroelectric synapses hold the solver's weights without power. The random walk at the heart of the method maps naturally onto that hardware. Fast, frugal equation solving would help sensors and edge devices. The paper pairs device physics with a concrete solver layout. Climate and chip design both lean on such solvers heavily.

Catalog record

CREATORSiqing Fu İD , Lizhou Wu ∗ İD , Tiejun Li ∗ İD , Chunyuan Zhang İD , Sheng Ma İD , Jianmin Zhang İD , Yuhan Tang İD and Jixuan Tang İD
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE34 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2507-04677

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

NeuroPDE: A Neuromorphic PDE Solver Based on Spintronic and Ferroelectric Devices Thanks: This work was supported by the National Key R&D Program of China(2022YFB2803405) and the NSFC(62304257). ( ∗ Corresponding author: Lizhou Wu and Tiejun Li.) Siqing Fu İD , Lizhou Wu ∗ İD , Tiejun Li ∗ İD , Chunyuan Zhang İD , Sheng Ma İD , Jianmin Zhang İD , Yuhan Tang İD and Jixuan Tang İD Abstract In recent years, new methods for solving partial differential equations (PDEs) such as Monte Carlo random walk methods have gained considerable attention. However, due to the lack of hardware-intrinsic randomness in the conventional von Neumann architecture, the performance of PDE solvers is limited. In this paper, we introduce NeuroPDE, a hardware design for neuromorphic PDE solvers that utilizes emerging spintronic and ferroelectric devices. NeuroPDE incorporates spin neurons that are capable of probabilistic transmission to emulate random walks, along with ferroelectric synapses that store continuous weights non-volatilely. The proposed NeuroPDE achieves a variance of less than 1e-2 compared to analytical solutions when solving diffusion equations, demonstrating a performance advantage of 3.48 × \times to 315 × \times speedup in execution time and an energy consumption advantage of 2.7 × \times to 29.8 × \times over advanced CMOS-based neuromorphic chips. By leveraging the inherent physical stochasticity of emerging devices, this study paves the way for future probabilistic neuromorphic computing systems. I Introduction As a fundamental class of mathematical equations, partial differential equations (PDEs) express the relationship between unknown functions containing several variables and their partial derivatives with respect to these variables. PDEs are widely used not only in all areas of physics but also in engineering to solve complex problems such as wave propagation, heat conduction, fluid mechanics, and particle transport [ 1 , 2 , 3 , 4 ] . In recent years, we have also seen a dramatic increase in the use of PDEs in computer sciences particularly for image processing and graphics. Due to the complexity of the equations, solving PDEs in an energy-efficient manner remains a big challenge and thus attracts great research attention. PDE solvers have evolved from classical analytical techniques [ 5 ] and numerical algorithms such as finite difference methods [ 6 ] , finite element methods [ 7 ] , and spectral methods [ 8 ] , to recently neural networks methods [ 9 , 10 , 11 , 12 ] and emerging Monte Carlo (MC) random walk methods [ 13 , 14 , 15 , 16 ] . Analytical and numerical methods struggle with solving high-dimensional and non-linear equations [ 17 ] , whereas neural network methods tackle this challenge through a data-driven approach. However, neural network PDE solvers, which incorporate PDE constraints into training or utilize pre-generated data, can be computationally expensive to train [ 18 ] . Addressing these limitations, novel MC random walk PDE solvers [ 16 ] employ a probabilistic PDE formulation to train unsupervised models using random particle walks, which can effectively solve convection-diffusion, Allen-Cahn, and Navier-Stokes equations. Additionally, Smith et al. [ 15 ] advanced random-walk-based PDE solving methodologies by implementing them on neuromorphic computing chips such as Loihi [ 19 ] and TrueNorth [ 20 ] , showing a substantial improvement in energy consumption. Nevertheless, these CMOS-based neuromorphic chips lack inherent randomness [ 21 ] and thus do not show a clear performance advantage over conventional CPUs and GPUs. To enhance computational performance while reducing energy consumption, we present NeuroPDE, a neuromorphic PDE solver that utilizes spintronic magnetic tunnel junction (MTJ) and ferroelectric tunnel junction (FTJ) devices. This design seamlessly bridges the gap between the probabilistic MC random walk method and the underlying circuit design, exploiting the intrinsic physical randomness of MTJ devices to create a scalable neuromorphic architecture capable of performing random walk functions. Through both SPICE simulations at the circuit level and functional simulations at the system level, we evaluated the ability of NeuroPDE to solve steady-state diffusion equations, showing the solution variance between NeuroPDE and analytical solutions remains below 1e-2. Compared to CMOS-based neuromorphic PDE solvers [ 15 ] implemented on Loihi and TrueNorth [ 19 , 20 ] , NeuroPDE demonstrates performance improvements of 3.48-315 \times and achieves power efficiency improvements of 2.7-29.8 \times . The main contributions of this paper are as follows: 1. We design a synapse based on an FTJ device. Utilizing the memristive property of this device, the synapse non-volatilely stores continuous synaptic weights, enabling stable storage of Markov chain transition probabilities. 2. We propose a neuron design based on an MTJ device. Using its probabilistic switching property, we achieve probabilistic activation, winner-takes-all, and self-inhibition, enabling simulation of random walks on a Markov chain through neural activation. 3. Building on our neuron and synapse designs, we develop the neuromorphic PDE solver NeuroPDE. We perform a comprehensive circuit and system simulations of NeuroPDE to assess its accuracy, execution time, and energy consumption in solving PDEs. As an architectural exploration, NeuroPDE demonstrates the potential of hardware-integrated stochastic PDE solvers, paving the way for practical PDE hardware solutions. The structure of this paper is as follows: Section II reviews the foundational concepts of MC random walks, and describes the novel devices used in the circuit construction. Section III presents the circuit design of NeuroPDE, with a focus on the neuron and synapse circuits. Section IV details the experimental evaluations, including circuit-level and system-level simulations, and presents the performance and energy results. Section V addresses the limitations and suggests directions for future research. Finally, Section VI concludes the paper. II Background II-A MC Random Walk Method The MC random walk method [ 22 ] relates particle trajectory simulations to PDE solutions by capturing stochastic behaviors, such as Brownian motion associated with thermal diffusion. This method involve spatial discretization, the construction of Markov chains to simulate particle movements, and MC sampling to approximate PDE solution spaces. The Feynman-Kac formula [ 23 ] extends the correspondence principle, suggesting a link between solutions of certain PDEs and their associated stochastic differential equations (SDEs). Specifically, the PDEs of interest are those formulated as Equation ( 1 ): \frac{\partial u}{\partial t}+\mu(x,t)\frac{\partial u}{\partial x}+\frac{1}{2}\sigma^{2}(x,t)\frac{\partial^{2}u}{\partial x^{2}}+f(x,t)u=0, (1) where u=u(x,t) is the function to be solved, \mu(x,t) and \sigma(x,t) represent the drift and diffusion terms, and f(x,t) is a given function. The Feynman-Kac formula expresses the solution to the PDE via a stochastic process, as shown in Equation ( 2 ): u(x,t)=\mathbb{E}[\phi(X_{t})\mid X_{t}=x], (2) where X_{t} denotes Brownian motion at time t , which can be efficiently simulated using a Markov chain model. As depicted in Fig. 1 , we define a finite one-dimensional space with N discrete positions as the state space S=\left\{{\mathrm{X}_{1},\mathrm{X}_{2},...,\mathrm{X}_{i},...}\right\} , where each state represents the position of the walker in the one-dimensional space. The Markov chain is characterized by its state transition matrix P , specifying the probability of transitioning from state i to state j : P_{i,i+1}=P_{i,i-1}=P_{g} , P_{i,i}=P_{s} . Boundary conditions are not detailed here, as they vary with the problem. Smith et al. [ 15 ] provided a comprehensive elucidation on solving PDEs through MC random walks, alongside a discussion on its applicability to a spectrum of problems, including stock option pricing, Boltzmann flux density, the heat equation with dissipation, and simple

Read the complete text