Two-dimensional (2D) lead-halide perovskites offer tunable optoelectronic properties governed by the quantum well (QW) width ($n$)1. However, the spontaneous formation of mixed QW width distributions remains a critical bottleneck for optimizing device performance. To address this, we developed a robust machine learning (ML)-based energy model, rigorously benchmarked against first-principles calculations, enabling extensive hybrid Monte Carlo simulations of complex perovskite microstructures. Our simulations on standard butylammonium and phenethylammonium systems reveal a universal transition from homogeneous to energetically favored mixed-phase distributions. This segregation is primarily driven by the enhanced thermodynamic stability of low-$n$ layers and the strong affinity of spacers to the inorganic framework. Demonstrating the predictive capability of this framework, we further extended our simulations to explore novel molecular engineering strategies. Preliminary results indicate that introducing high-entropy spacer configurations can disrupt the thermodynamic driving forces responsible for segregation. These findings highlight the potential of ML-powered multiscale modeling to not only elucidate complex microstructural evolution but also to guide the design of phase-pure, next-generation optoelectronic materials.