A Flexible and Accurate Additive Manufacturing Data Retrieval Method based on Probabilistic Modeling and Transformation-Invariant Feature Learning

Additive manufacturing (AM) is gaining prominence across numerous fields, which involves the generation of extensive data at each process stage. A relational database is a useful tool to store such AM data in a key-value manner and streamline data retrieval. Users can specify the value of one AM variable or key and retrieve the corresponding record values of another key. This establishes the correlations between AM variables, and supports applications such as process planning. Nonetheless, such an operation is a “hard” query, which lacks reasoning capabilities and fails to provide useful information when required records are missing. It is urgent to develop a more powerful AM database to handle AM data better, which should support “soft” queries, be scalable to high-dimensional data, and maintain flexible query functionality among multiple keys. In this paper, we upgrade an AM database with probabilistic modeling and transformation-invariant feature learning, which is termed as a probabilistic AM database (PAMDB). The PAMDB allows the selection of any key as a query key, or even multiple keys as query keys, to retrieve the values of other keys, which is adapted to unseen, high-dimensional, and multi-modal AM data. Two case studies were conducted for laser powder bed fusion (LPBF) and vat photopolymerization (VP). Compared with existing methods, experimental results underscore the efficacy of the PAMDBs, both qualitatively and quantitatively, in tasks that includes melt pool size prediction and scan parameter estimation in LPBF, and defect detection for the resin deposition process in VP.