Adaptive Sampling and Variability—Machine Learning for Optimizing Additive Manufacturing Processes

Additive Manufacturing (AM) has been the subject of active research in recent years due to its potential use in many applications and the flexibility in designs it offers. However, the sensitivity of AM printed materials and their net properties to the processing conditions remains a challenge to the widespread adoption of these emerging techniques. A need, therefore, exists to optimize AM processes to identify the best printing conditions and unlock the true potential of this manufacturing capability. Recent studies show that traditional optimization approaches are inapplicable as AM materials possess significantly high variability in their net properties when the printing is repeated with the same processing parameters. In terms of optimization, sequential designs can be used to explore the high-dimensional AM processing parameter spaces efficiently. In such approaches, the optimization algorithm learns from previously collected data and adaptively proposes new experiments to selectively sample the space where needed the most to optimize the AM process. Sequential designs usually rely on optimizing a given metric in the design space to choose the experiment in the next iteration. However, the AM costs associated with performing one experiment per iteration are typically high, and an efficient novel optimization algorithm that can identify several experiments per a given iteration is needed to lower the costs associated with the optimization.<br/><br/>Herein, we develop an adaptive optimization method that implements machine learning and integrates the variability in the properties of AM materials. Our approach learns from previous experiments and adaptively proposes new locations in the input domain that maximize expected information gain. Gaussian Process Regression and probability calculus are used to explicitly account for variability in the analysis and optimization. In addition to variability, the novelty of our approach lies in the use of utility theory to define the optimization criteria upon which the new experiments are chosen and the flexibility in identifying the number of new design sites and their respective AM samples to be tested during each optimization iteration. The proposed method is tested on simulated data to showcase its performance. More specifically, multimodal response surfaces for both the mean and variability of AM process parameters are used to benchmark our optimization approach. This study is expected to result in an efficient, variability-embracing adaptive optimization method.