Accurate prediction of mechanical properties in as-cast aluminium alloys remains a critical challenge in alloy design due to the complex, non-linear interplay between chemical composition, solidification thermodynamics, and microstructure. This study presents a machine learning framework augmented with analytically computed thermodynamic descriptors for predicting Vickers hardness (HV) of Al-6xxx series as-cast alloys. Thermodynamic features including phase fractions (f_Mg₂Si, f_Al₃Sc, f_Al₃Zr, f_FCC_Al), liquidus/solidus temperatures, solidification range, and Gibbs mixing energy at 500°C were calculated using simplified analytical expressions implemented in Python, rather than full CALPHAD software, and integrated with conventional compositional and physics-based descriptors.A curated dataset of 69 alloys (HV range: 40–107 HV; mean 84.6 ± 17.0 HV) was assembled from literature, augmented to 150 training samples via Gaussian noise injection, and used to benchmark 11 machine learning models. CatBoost and Stochastic Gradient Boosting emerged as top performers, achieving test R² values of 0.804 and 0.824 respectively, with RMSE below 4.6 HV. External validation on four novel as-cast alloys confirmed strong generalisability, with mean absolute prediction errors of 1.8% (CatBoost) and 1.5% (Stochastic GB). SHAP analysis identified Trace_total, Atomic_size_mismatch, and Excess_Si as the most influential predictors, followed by Mg and Mn. Thermodynamic descriptors contributed meaningfully, with T_liquidus and f_FCC_Al ranking in the top 10. Three inverse design strategies (gradient-based, genetic algorithm, Bayesian optimisation) identified compositions targeting 95 HV. Pareto front analysis revealed 11 cost-optimal compositions spanning 4.18–10.29/kg, with the genetic algorithm proposing a minimum-cost solution at $4.87/kg. Sensitivity analysis quantified marginal effects of each alloying element, providing actionable guidelines for cost-efficient Al-6xxx alloy design