Meta-analysis of literature data in metal additive manufacturing: What can we (and the machine) learn from reported data?

Abstract

Obtaining in-depth understanding of the relationships between the additive manufacturing (AM) process, microstructure and mechanical properties is crucial to overcome barriers in AM. In this study, database of metal AM was created thanks to many literature studies. Subsequently meta-analyses on the data was undertaken to provide insights into whether such relationships are well reflected in the literature data. The analyses help reveal the bias and what the data tells us, and to what extent machine learning (ML) can learn from the data. The first major bias is associated with common practices in identifying the process based on optimizing the consolidation. Most reports were for consolidation while data on microstructure and mechanical properties was significantly less. In addition, only high consolidation values was provided, so ML was not able to learn the full spectrum of the process - consolidation relationship. The common identification of process maps based on only consolidation also poses another bias as mechanical properties that ultimately govern the quality of an AM build are controlled not only by the consolidation, but also microstructure. Meta-analysis of the literature data also shows weak correlation between process with consolidation and mechanical properties. This weak correlation is attributed to the stated biases and the non-monotonic and non-linear relationships between the process and quality variables. Fortunately, trained ML models capture well the influence and interactions between process parameters and quality variables, and predicts accurately the yield stress, suggesting that the correlation between process, microstructure and yield strength is well reflected in the data. Lastly, due to the current limitation in the process map identification, we propose to identify the process map based on not only the consolidation, but also mechanical properties.