Abstract We present a novel hybrid computational framework that couples classical finite-difference timedomain (FDTD) simulation with machine learning (ML) to model and forecast nonlinear wave propagation in a cold plasma. In this study, a one-dimensional (1D) FDTD Maxwell–Drude solver is implemented in Google Colab to simulate electromagnetic waves in a cold (collisionless) plasma. The governing equations (Maxwell’s curl equations coupled to the cold-plasma polarization equation) are solved with absorbing boundary conditions. To extract nonlinear features (such as wave steepening) and to predict future wave profiles, we employ an autoencoder to compress the simulated field snapshots into a lowdimensional latent space, and a shallow neural network to forecast the wave evolution in that latent space. Quantitative results are presented, including the dispersion relation versus plasma frequency and time-domain field amplitude growth. We demonstrate that the ML model accurately reproduces ...