Environmental, Social, and Governance (ESG) investment analytics has matured from screening-based exclusions to integrated, multi-objective portfolio construction that explicitly quantifies sustainability–financial trade-offs. This paper develops a rigorous framework for sustainable portfolio optimization that (1) characterizes ESG data uncertainty and provider divergence, (2) integrates ESG indicators within mean–variance and multi-objective optimization formulations, and (3) leverages machine-learning methods to extract material ESG signals and to enhance return and risk forecasts. We propose an uncertainty-aware ESG penalty term that augments classical mean–variance optimization and show how varying the ESG-aversion parameter produces a continuum from financially-efficient to sustainability-efficient allocations. To account for heterogeneous ESG data and methodological differences across providers, we incorporate ensemble ESG scoring and robust optimization techniques, and we evaluate performance using out-of-sample backtests across multiple markets and time horizons. The framework is extended to a multi-objective evolutionary solver that jointly optimizes expected return, downside risk (CVaR), and aggregated ESG impact metrics, while an explainability module maps portfolio exposures to material E, S, and G drivers. Empirical results indicate that (i) accounting explicitly for ESG uncertainty meaningfully alters allocations, (ii) ML-driven indicator selection improves forecasting accuracy and portfolio ex-post sustainability profiles, and (iii) regulatory developments and methodological biases in ESG scoring critically influence risk–return–sustainability trade-offs. The paper closes with a discussion of implementation challenges, data governance, and directions for future research...