Casting parts retrieval driven by the integration of process constraints and geometric features

The formation of castings depends on numerous process parameters, making traditional casting process design heavily reliant on experience and complex calculations. Efficiently retrieving similar parts from a case library for process reuse can significantly enhance design efficiency. This study has proposed a tailored retrieval scheme for casting parts, ensuring that retrieved parts provide valuable references for process design. First, the study has introduced the concept of process constraint values to quantify the impact of differences in key semantic information—material, casting method, and production batch size—on the applicability and effective transferability of existing processes. These constraint values are incorporated as conditions in subsequent similarity calculations, ensuring that the retrieved parts offer practical reference value for casting process design. For geometric similarity measurement, simulation experiments and flowability analyses are conducted to evaluate the influence of volume and modulus on similarity computation. Considering the characteristics of casting production, shape-structure features are extracted from both the external solid model and internal cavity models. These features are then integrated into a combined encoding, providing a parametric representation of the part’s geometric characteristics. Finally, the overall geometric similarity between parts is simultaneously measured based on volume, modulus, and shape-structure feature encoding. The effectiveness of the proposed retrieval algorithm was validated using Expected Reciprocal Rank (ERR) and F1-score. Results indicate that the first highly similar part appeared at the 1.23th position in the ranked retrieval sequence, allowing users to efficiently find relevant parts at the top of the retrieval rankings. Furthermore, separating the feature extraction and similarity calculation of the external solid and internal cavity improves retrieval performance compared to direct feature extraction from the original part model, especially in high-relevance scenarios with few candidate results.