Trustworthy model adaptation for manufacturing process certification under corrupted data via knowledge-guided unlearning

The National Science Foundation Division of Civil, Mechanical, and Manufacturing Innovation awarded Florida State University $299,999 on January 1, 2027, under the Engineering program (CFDA 47.041) to develop knowledge-guided methods for correcting predictive manufacturing models trained on corrupted or mislabeled data.

The project will integrate hierarchical knowledge graphs, semantic consistency analysis, and localized model correction to remove invalid learned relationships from manufacturing process models while preserving useful process knowledge. The work targets high-value, low-volume production environments where rebuilding corrupted models is costly or infeasible. Researchers will extract relationships among defects, process conditions, materials, observable attributes, and manufacturing outcomes from validated engineering documents, images, and manufacturing data, using language models as constrained extractors of localized evidence and semantic clustering to identify inconsistencies. The approach aims to improve reliability of data-driven process certification, reduce defective production and material waste, and strengthen confidence in manufacturing systems that rely on predictive models for diagnosis, quality prediction, and certification decisions.

The project will train graduate and undergraduate students through research, coursework, and industry collaboration. Research results, educational materials, and shareable software will be disseminated to academic and manufacturing communities.

Performance extends from January 1, 2027, through December 31, 2029, with work performed in Tallahassee, Florida.

Abstract

Advanced manufacturing systems increasingly rely on predictive computer models to connect process observations with diagnosis, quality prediction, and certification decisions. These models are often developed from simulations, historical records, sensor measurements, and engineering assumptions, then deployed within broader digital manufacturing workflows. When some of these sources contain mislabeled examples, outdated information, or flawed simulation results, invalid relationships become embedded in the model and propagate into downstream monitoring and certification decisions. Correcting such errors is especially difficult in high-value, low-volume production, where new data are scarce and rebuilding a model may be costly or infeasible. This project will develop a knowledge-guided approach that integrates manufacturing data, structured engineering knowledge, and localized model correction to remove invalid learned relationships while preserving useful process knowledge. The work will support more reliable process certification, reduce defective production and wasted materials, and strengthen confidence in data-driven manufacturing systems. The project will also train graduate and undergraduate students through research, coursework, and industry collaboration. Research results, educational materials, and shareable software will be disseminated to academic and manufacturing communities. The research will integrate hierarchical knowledge graphs, semantic consistency analysis, and localized model correction within a run-to-run certification workflow. Relationships among defects, process conditions, materials, observable attributes, and manufacturing outcomes will be extracted from validated engineering documents, images, and manufacturing data. Language models will serve as constrained extractors of local evidence, while semantic clustering, physical consistency checks, data compatibility analysis, and selective expert review will assemble and refine the knowledge graphs. These graphs will provide reliable reference knowledge for detecting the effects of corrupted data by comparing observed sample attributes and assigned labels against known process-defect relationships, without requiring ground-truth labels or complete historical baselines. Retain and forget knowledge will then be separated at the attribute-association level, preserving shared valid features while targeting invalid relationships for removal. Knowledge-guided pseudo-samples will probe whether a deployed digital twin still responds to invalid attribute patterns, and gated fine-tuning will update only the model parameter regions most responsible for those responses. The methods will be evaluated under controlled corruption and limited-data conditions against standard model-updating and machine-unlearning approaches. Industrial validation will use in-process and characterization data from powder-bed additive manufacturing to support run-to-run quality prediction and certification decisions. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the foundation's intellectual merit and broader impacts review criteria.- subawards are not planned for this award.