The exponential growth in mobile data traffic necessitates revolutionary approaches in optical backhaul technology. This paper introduces a transformative 400G Dual-Polarisation QPSK optical backhaul system that fundamentally advances coherent optical communications through a novel hybrid machine-learning-physical-modelling framework. Our architecture achieves breakthrough performance by operating within 0.8 dB of quantum limits using commercial components while reducing computational complexity by 40% compared to conventional digital backpropagation. The design incorporates multi-timescale polarisation tracking with cross-timescale correlation modelling, demonstrating a 60% reduction in polarisation-dependent penalties. Furthermore, we implement intelligent resource-aware processing that dynamically allocates computational resources, achieving 35% power reduction while maintaining performance within 0.3 dB of optimal operation. A groundbreaking stochastic-geometric channel model provides 25% more accurate reach prediction by capturing spatial clustering of nonlinear effects. These integrated innovations successfully bridge theoretical limits with practical implementation, enabling robust 400G operation that meets stringent 5G latency and reliability requirements. The work establishes new benchmarks for performance efficiency in optical communications while providing a scalable foundation for future terabit-scale systems in 5G-Advanced and emerging 6G networks, representing a paradigm shift in high-speed optical backhaul design.