Machine learning-assisted in-situ adaptive strategies for the control of defects and anomalies in metal additive manufacturing

In metal additive manufacturing (AM), the material microstructure and part geometry are formed incrementally. Consequently, the resulting part could be defect- and anomaly-free if sufficient care is taken to deposit each layer under optimal process conditions. Conventional closed-loop control (CLC) engineering solutions which sought to achieve this were deterministic and rule-based, thus resulting in limited success in the stochastic environment experienced in the highly dynamic AM process. On the other hand, emerging machine learning (ML) based strategies are better suited to providing the robustness, scope, flexibility, and scalability required for process control in an uncertain environment. Offline ML models that help optimise AM process parameters before a build begins and online ML models that efficiently processed in-situ sensory data to detect and diagnose flaws in real-time (or near-real-time) have been developed. However, ML models that enable a process to take evasive or corrective actions in relation to flaws via on the fly decision-making are only emerging. These models must possess prognostic capabilities to provide context-sensitive recommendations for in-situ process control based on real-time diagnostics. In this article, we pinpoint the shortcomings in traditional CLC strategies, and provide a framework for defect and anomaly control through ML-assisted CLC in AM. We discuss flaws in terms of their causes, in-situ detectability, and controllability, and examine their management under three scenarios: avoidance, mitigation, and repair. Then, we summarise the research into ML models developed for offline optimisation and in-situ diagnosis before initiating a detailed conversation on the implementation of ML-assisted in-situ process control. We found that researchers favoured reinforcement learning approaches or inverse ML models for making rapid, situation-aware control decisions. We also observed that, to-date, the defects addressed were those that may be quantified relatively easily autonomously, and that mitigation (rather than avoidance or repair) was the aim of ML-assisted in-situ control strategies. Additionally, we highlight the various technologies that must seamlessly combine to advance the field of autonomous in-situ control so that it becomes a reality in industrial settings. Finally, we raise awareness of seldom discussed, yet highly pertinent, topics relevant to adaptive control. Our work closes a significant gap in the current AM literature by broaching wide-ranging discussions on matters relevant to in-situ adaptive control in AM.