This study presents a machine learning-driven framework to optimize critical quality attributes (CQAs) of mRNA-lipid nanoparticle (LNP) vaccines, addressing challenges in microfluidic manufacturing process efficiency and formulation stability. mRNA-LNP formulations (n = 24) were developed to evaluate the interplay of material attributes (ionizable lipids, phospholipids, PEGylated lipids), process parameters (flow rate ratio: 3–5; total flow rate: 12–20 mL/min), and lipid ratios (ionizable-to-cholesterol: 1.08–1.33; phospholipid-to-PEG: 3.76–6.67; N/P: 6–10) using I-optimal design of experiments (DOE). Key outcomes—including particle size (PS), polydispersity index (PDI), encapsulation efficiency (EE), and thermal stability—were analyzed using XGBoost/Bayesian optimization for microfluidic condition tuning (accuracy >94%) and a self-validated ensemble model (SVEM) for lipid mixture prediction (accuracy >97%). Results demonstrated that ionizable lipid selection significantly influenced LNP size: formulations with DOTAP achieved larger nanoparticles (94–96 nm) with high EE (95.5%), while MC3 produced smaller LNPs (51–57 nm) with reduced EE (79–85%). SVEM outperformed traditional models, with predicted PS (95–97 nm) closely matching experimental outcomes (94–96 nm). Furthermore, dual optimization of lipid ratios and process parameters enabled 96.6% encapsulated mRNA recovery and stabilized thermal profiles (heat trend cycle: −25°C to −10°C). Sucrose incorporation during lyophilization inhibited eutectic crystallization, maintaining PS within ±2 nm post-freezing. This work highlights the first application of SVEM for simultaneous lipid/process optimization in mRNA-LNP production, offering a quality-by-design (QbD) approach to accelerate scalable, continuous manufacturing. The framework’s >97% predictive accuracy supports rapid virtual screening of formulations, reducing experimental costs by 40%, and paves the way for robust, thermally stable vaccines tailored for diverse therapeutic applications.