OpenAqua: An automated multi-agent framework for early-stage water treatment train design with retrieval augmentation and critic-based refinement.

Water treatment train design requires integrating heterogeneous public technical knowledge and historical engineering cases under multiple technical and operational constraints. In current practice, this process still relies heavily on expert judgment and manual evidence synthesis, making it difficult to systematically evaluate complex design factors and transparently compare alternative trains. These challenges have motivated interest in large language models (LLMs), which show strong capabilities in technical text understanding and synthesis and have been explored for scientific question answering, technical document analysis, and domain-specific decision support. However, directly applying LLMs to water treatment train design without explicit grounding may produce weakly supported recommendations and incomplete consideration of engineering constraints. To address these limitations, this study proposes OpenAqua, a multi-agent retrieval-augmented framework that supports early-stage water treatment train design with critic-based refinement. OpenAqua decomposes the task into requirement analysis, hierarchical knowledge retrieval, treatment train generation, critic-based verification and result interpretation. It integrates technical knowledge with engineering cases to generate candidate treatment trains supported by explicit evidence, constraint checking, and risk descriptions. To evaluate the framework, we construct WContBench and compare OpenAqua with stand-alone LLMs. OpenAqua improves design quality, structural similarity to reference treatment trains, constraint satisfaction, and evidence support relative to stand-alone LLMs. Ablation experiments further show that the hierarchical knowledge base and retrieval-augmented generation in OpenAqua improved water treatment train design quality. These findings suggest that multi-agent reasoning with critic-based refinement is a promising approach for transparent and practically useful water treatment decision support.