Source tolerant iterative domain adaptation for inexpensive modeling of knowledge-deficient manufacturing processes

• Focus on reducing both experiments and knowledge needed for manufacturing modeling. • Created new AI-based framework to address foundational challenge of concept drift. • Modified iterative Transfer Learning with correlation-loss-driven source update. • Achieved 63% lesser error and 73% lesser experiments than conventional AI. • Robust to large concept drift that causes failure of conventional Transfer Learning. Modeling of parameter-attribute relationships in manufacturing processes is resource-intensive and time-consuming due to the need for significant experimental data (experimental burden) or qualitatively accurate knowledge of material/machine behavior (knowledge burden). Transfer Learning shows potential to address this issue but frequently fails under concept drift, i.e., change in qualitative form of the parameter-attribute relationship from source to target. This work proposes a new framework called Source Tolerant Learning via Iterative Domain Adaptation (STOLID) to address this knowledge-experiment tradeoff, a first in manufacturing, whose core innovation lies in iterative domain adaptation for mitigating concept drift via correlation-driven source updating and repeated transfer learning. The AI novelty lies in relaxing the need for high qualitative similarity between source and target in Transfer Learning (to mitigate knowledge burden) while extracting useful information from the source (to mitigate experimental burden) by introducing incremental correlation-loss-driven update of the source with iterative quantitative-loss-driven Transfer Learning. Testing across three mechanistically distinct processes (i.e., additive, subtractive, forming) demonstrates up to 63% enhancement in prediction accuracy and up to 74% reduction in experimental burden relative to conventional brute force Machine Learning, even when incomplete knowledge of the form of the relationship causes the state-of-the-art Transfer Learning to fail under concept drift. The implications for experimentally inexpensive modeling of hitherto intractable processes or attributes are discussed.