This paper presents an optimized deep learning-based anti-jam beamforming method for radar and communication systems operating in hostile electromagnetic environments. The proposed method employs a specifically engineered one-dimensional convolutional neural network to directly learn the nonlinear mapping between received signal features and optimal beamforming coefficients. This data-driven approach overcomes the limitations of conventional adaptive methods based on covariance matrix inversion by providing a computationally efficient solution. The network architecture is carefully designed to achieve an optimal balance between performance and complexity, enabling real-time implementation on existing hardware platforms. Simulation results demonstrate that the proposed method can generate deep nulls exceeding -60dB in jammer directions while maintaining the desired signal gain at approximately 0dB without distortion. The achieved SINR improvement of approximately +9.6dB and sidelobe levels in the range of -12 to -15dB confirm the method's effectiveness in suppressing multiple interference sources. Comparative analysis with recent state-of-the-art techniques reveals that the proposed method exhibits superior performance in terms of null depth and main-lobe integrity preservation. The method demonstrates robustness against array imperfections and angular uncertainties through integrated diagonal loading and adaptive constraint mechanisms. These characteristics make the proposed approach a reliable and practical solution for beamforming challenges in modern electronic warfare scenarios, offering significant advantages over traditional adaptive algorithms in terms of computational efficiency, interference suppression capability, and implementation feasibility for next-generation anti-jam systems.