SeismicShield-RL: A Preregistered and Auditable Infrastructure for Reinforcement Learning and Multi-Objective Seismic Friction-Damper Co-Design

Abstract

Reinforcement learning is an attractive candidate for seismic retrofit optimization because the design space is discrete, multi-objective, computationally expensive, and affected by earthquake and structural uncertainty. The same characteristics, however, make algorithm comparisons vulnerable to unequal simulation budgets, data leakage, post-hoc model selection, source-code drift, and ambiguous restart semantics. This paper reports SeismicShield-RL, a preregistered and auditable research infrastructure for friction-damper co-design rather than a completed algorithm-ranking study. The frozen benchmark contains 136 processed Engineering Strong Motion records from 34 events, partitioned into 52 training, 20 validation, 16 pilot, and 48 confirmatory records; 16 structural states spanning 3-, 6-, 10-, and 20-story buildings; three objectives (retrofit cost proxy, maximum inter-story drift ratio, and peak floor acceleration); six stochastic methods (random search, scalar genetic algorithm, NSGA-II, PPO, IPPO, and MAPPO); and eight fixed random seeds. A runtime preflight reproduced the SHA-256 identities of all 136 processed records and successfully exercised four Tier-1 and four Tier-2 pilot fixtures without inspecting confirmatory structural-response outcomes. Deterministic planning decomposed the registered experiment into 475 atomic shards and 2,820,160 structural-response calls. Stage A alone requires 2,780,992 Tier-1 calls, corresponding to approximately 1,026.68 hours of projected sequential simulation-call time in the tested environment; the 39,168-call Tier-2 campaign projects to approximately 21.64 hours. Because the full registered computation exceeded the intended no-cost compute envelope, the full Stage-A selection campaign, the 768-item selection freeze, and confirmatory Tier-2 evaluation were deferred rather than altering the preregistered semantics. No confirmatory claim of algorithmic superiority or seismic efficacy is made. The contribution is a preserved experimental foundation that can be resumed, audited, or adapted by future researchers.