To ensure robust signal recovery and efficient data transmission in wireless communication systems, accurate channel estimation plays a vital role, especially under dynamic and complex conditions. Machine learning-based channel estimation is explored for binary phase shift keying (BPSK) and quadrature phase shift keying  (QPSK) modulation schemes over Rayleigh, Rician, and Gaussian fading models. In this work, a framework using Convolutional Neural Networks (CNN) and Multilayer Perceptrons (MLP) is developed to predict channel coefficients and analyze the impact on bit error rate (BER), throughput, and spectral efficiency for binary modulations. A comprehensive performance comparison of BPSK and QPSK under ML-based estimation across various fading conditions is provided. The results show that CNNs are effective in tracking time-varying coefficients, while MLPs often yield lower mean squared error (MSE). The study emphasizes practical applications in low-SNR environments and supports energy-efficient designs aligned with SDG goals. Key simulation results include BER vs SNR, throughput, and spectral efficiency comparisons between BPSK and QPSK under ML estimation.