jwhite510 / CUDA-FDTD-simulation

Calculation of the Finite Difference Time Domain method in parallel on a GPU

CUDA-FDTD-simulation

Finite Difference Time Domain Simulation written in CUDA, with python Bindings. All of the CUDA Tensors for the Electric and Magnetic fields are exposed as numpy arrays in python.

CUDA is used to calculate each point of the computational grid in parallel on a GPU.

The FDTD uses a discretized form of Maxwell's equations to simulate propagation of a source through vacuum or a material described by a complex refractive index n(x,y,z).

Maxwell's Equations discretized to 3D space

The computational process is extremely demanding because the computational grid must be resolved on the order of the wavelength of the EM waves being simulated.

import matplotlib.pyplot as plt
import numpy as np
from Cfdtd import FDTD

if __name__=="__main__":
    N_x=200
    N_y=200
    N_z=200
    dx=50e-9
    c=2.998e8
    dt=dx/(2*c)
    freq=500e12;
    omega =2*np.pi*freq;
    tau=10*dt

    fdtd = FDTD(N_x,N_y,N_z,dx,dt,omega,tau)

    time_span=15e-15;
    tmax_steps=int(time_span/dt);

    # define source term
    _x=np.arange(N_x).reshape(-1,1,1)
    _y=np.arange(N_y).reshape(1,-1,1)
    _z=np.arange(N_z).reshape(1,1,-1)
    fdtd.J.z[:,:,:]=np.exp(-(_x-100)**2 / 5)*np.exp(-(_y-100)**2 / 5)*np.exp(-(_z-100)**2 / 5)
    fdtd.J.z[:,:,:]+=np.exp(-(_x-75)**2 / 5)*np.exp(-(_y-75)**2 / 5)*np.exp(-(_z-100)**2 / 5)


    plt.figure(1)
    for n in range(0,tmax_steps):

        fdtd.timestep(5) # run 5 timesteps

        plt.gca().cla()
        # show the x component of the magnetic field
        plt.imshow(fdtd.H.x[:,:,100])
        plt.pause(0.01)


    fdtd.delete()