AI Meets Kerbal Space Program: Enhancing Space Missions Through AI-Driven Spacecraft Design and Sandbox Games

<p dir="ltr"><b><u>Abstract</u></b>: </p><p dir="ltr">While cislunar space exploration demands mission architectures with flexible interdependencies between novel systems and their operations, heritage data for such Systems-of-Systems are sparse. This gap limits the development of strategies to evaluate the feasibility and performance of emerging concepts, such as On-Orbit Refueling, critical for sustaining reusable spacecraft in cislunar missions. Traditional evaluation methods alone are insufficient, creating a pressing need for approaches that reveal where higher fidelity and deeper analysis is needed. To address these limitations, this work introduces a framework that couples (i) rapid sandbox game-based system design prototyping augmented with fidelity-enhancing modifications, (ii) a mission architecture generation engine featuring advanced astrodynamics for architectural operations, and (iii) surrogate models equipped with Explainable AI for enhanced interpretability. The proposed framework rapidly explores diverse early-stage space system designs within Kerbal Space Program through reinforcement learning and evolutionary algorithms. High-performing designs are then systematically evaluated within a higher-fidelity SoS model, where surrogate models predict mission-critical metrics with exceptional accuracy (R² > 0.99), while Shapley Additive Explanations elucidate which design variables most significantly influence mission outcomes. Counterfactual analysis further provide actionable recommendations - like mass reductions, engine changes, or orbit adjustments - without re-running computationally intensive simulations. To demonstrate the framework's utility, a case study centered on a refueling architecture inspired by Artemis V is investigated. To the author’s knowledge, this is the first approach that merges sandbox-generated designs, cost and performance measures, and explainable surrogates for cislunar logistics into a single auditable pipeline. This framework (i) provides a novel sandbox-based methodology for Pre-Phase A exploration of system design concepts before committing significant resources, (ii) delivers real-time feature importance while quantifying explanation quality through compactness, stability, and cross-method consistency diagnostics, and (iii) generates counterfactual architectural changes that translate model insight into concrete actions. It also sets the stage for dedicated sandbox game-based and XAI frameworks for future space mission research, advancing the explainability of complex models in this field.</p>