Complex models and simulations are widely used to represent engineering systems. They are often large and exhibit irregular behavior under variable conditions, requiring advanced methods for decision-making. Best designs are typically found through optimization, while resilient designs require optimization under uncertainty. Unveiling vulnerabilities is very informative for building a resilient yet convenient system. Still, this task spans multiple objectives and multiple resilience metrics, from risk-prone to risk-averse, and thus implies a wide exploration of possible scenarios. Regrettably, such an exploratory approach is computationally intractable with traditional sampling methods, despite parallelization. Surrogates can address sample-efficiency issues. They emulate complex models with limited computational resources, but they suffer from accuracy and scalability issues. Here, we propose applying the recently developed Multi-Level Informed Optimization (MLIO), an adaptive, accurate, and scalable machine learning method, built on decomposed Kriging surrogates. Once trained on a small subset of self-selected scenarios, MLIO can perform different multi-objective optimizations under uncertainty on large models. Multiple Values at Risk (VaR) for resilience and performance metrics can be evaluated across a statistically significant number of scenarios, ultimately exposing system vulnerabilities. The search for a climate-resilient European energy system is used as a representative case study. It comprises dozens of design variables and uncertain parameters, and the model’s behavior is highly nonlinear across scenarios. Optimal design solutions are trade-offs between two objectives: total cost (performance) and value of lost load (resilience). MLIO is adaptively trained on the physical model to map objectives to variables and parameters, using only a few hundred scenarios. Then, three VaR optimizations (5%,50%,95%) are conducted through the surrogate, involving ten thousand climate scenarios. The three corresponding Pareto fronts are compared to identify vulnerable design choices. Finally, comprehensive, risk-aware, and quantitative guidance is provided to identify optimal compromise solutions that structurally improve system resilience.