and material noise. Final results from iterative comparison and adjustments of the model data, being compared in impedance planes, are shown in Figure 5 and highlight the ability to evaluate the buried nonmetallic inclusion size and depth. Additional steps in the development process resulted in the ADA algorithms providing guidance to the inspectors when features in the data indicate when a fatigue crack is emanating from a nonme- tallic inclusion. The ADA being developed for this application is in its final stages of refinement before it will be evaluated by a formal validation process. The third example combines elements of heu- ristics, simulations, and large dataset analysis to realize a successful outcome on a very complex inspection. The application addresses the lower forward spar cap on C-130 aircraft (Lindgren et al. 2005), as shown in Figure 6. The approach leverages development at the academic level for both the generation and detection of ultrasonic creeping waves (Nagy et al. 1994), plus the use of algorithms to discern the presence of cracks in a less complex, but still challenging, application (Aldrin et al. 2001). As described in Lindgren et al. (2005), the solution included the use of ana- lytical methods to fully represent the propagation paths within the structure simulation tools to explore various attributes of the inspection data as it propagates in the structure plus, the use of advanced processing methods, namely echo dynamics and local correlation functions, to dis- criminate between responses from potential flaws to those from other geometric reflectors found intermittently in the structure. In addition, over 2000 representative inspection opportunities Scan, x (mils) –50 0 50 20 10 0 –10 –200 –50 0 50 Scan, x (mils)ls)i –50 0 50 20 10 0 –100 20 10 0 –100 –200 –50 0 50 Scan, x (mils) –50 0 50 –50 0 50 Scan, x (mils) –50 0 50 10 5 0 –5– –10 –50 0 50 S (mils) – 0 – (m ––20 ––30 0 S – 0 Figure 5. Model fit (a) to experimental data (b) for the vertical (top) and horizontal (bottom) representations of eddy current scans from a sub-surface nonmetallic particle with differences due to probe and material variability. FEATURE |AI/ML Typical crack Typical crack Web Lower panel Beam cap Front beam cap WS 174 Up Aft Outboard Trace of web splice tee WS 178 Bulkhead WS 80.5 Figure 6. Lower forward spar cap of the C-130 illustrating the complexity of the regions to be inspected. 40 M A T E R I A L S E V A L U A T I O N • J U L Y 2 0 2 3 2307 ME July dup.indd 40 6/19/23 3:41 PM Index, y Index, y Index, y Index, y I (m(mils) I (m(mils) (m(mils) (m(mils)

from both harvested and mock-up test articles were used to refine the decision-making process for the ADA algorithms. The inspection process was fully validated by a comprehensive POD study before being deployed. The inspections were accomplished by contractor field teams that would collect the data and ensure it had sufficient quality to be evaluated by the ADA. Suspected indications identified by the ADA were sent to an NDT engineer to make a final determina- tion if the indication was confirmed and needed to be sent to engineering for disposition. The next generation of ADA will expand the capability of algorithms to facilitate the identification of defects to the capability to characterize the defects in ways that are not available today. While inspec- tors can use methods to approximate defect size, attributes like fatigue crack depth are especially chal- lenging. However, using a combination of heuristics, simulations, and data-driven analytical methods, the use of ADA to determine the depth of a fatigue crack from a bolt-hole eddy current inspection was shown to have an average accuracy of 8.5% for fastener holes with minimal variability (Aldrin et al. 2019a). The next steps in the development process are to use this inte- grated approach to address fastener hole variability, such as skew and out-of-round attributes, to provide a crack depth estimated with a statistical bounds on accuracy, to enable rapid disposition of these defects in aerospace structures. Summary There is a continued potential for AI/ML methods to enhance data analysis and diagnostics for NDT data. However, there needs to be a realis- tic approach that includes evaluation of the data quantity, quality, and fidelity. This ensures it has the desired attributes that enable the AI/ML tech- niques to provide outcomes with sufficient statis- tical metrics for the results to be used in engineer- ing decisions. In addition, these outcomes require rigorous validation of the diagnostic capability before they can be trusted to help ensure the integ- rity, or safety, of systems. A representative example illustrated the chal- lenges in using AI/ML techniques for smaller and noisy datasets, highlighting how this can lead to outliers that would imply potentially missed defects if this approach was used for NDT datasets. Additional challenges exist in data variability from equipment, defects, and structure that impact the amount of quality data required for AI/ML approaches. While data for defects can be aug- mented by simulations, these must contain all the anticipated variability and complexity of the NDT evaluation technique to represent nuances and outliers that are challenging for AI/ML, but critical for high-accuracy flaw detection. The challenges of AI/ML when used for NDT data has led AFRL to pursue a hybrid approach that integrates AI/ML with heuristic- and model-based diagnostic algorithms to facilitate and reduce the workload of inspectors while not taking them com- pletely out of the loop. Representative examples for several DAF-related applications have demon- strated the power of combining at least two of these methods to enable complex inspections and diagnostics of NDT data. The ADA algorithms are combined with human analysis to maximize the value of the algorithms by reducing the workload of inspectors so they can focus on the critical data that could be indications of defects being present. Future work includes plans to expand the capabili- ties of ADA algorithms to characterize defects with statistical metrics of accuracy. Initial development efforts have shown the potential of this capability, which would decrease the disposition time of indi- cations and increase availability of the system to the end user. ACKNOWLEDGMENTS The author expresses his deep appreciation for the pioneering contributions and collaboration with Dr. John Aldrin of Computational Tools. Mr. David Forsyth of Texas Research Institute – Austin is recognized for his contribution to implementing ADA algorithms for multiple NDT applica- tions. The AI/ML analysis here would not be possible without the work performed by Mr. Tushar Gautam and Drs. Kirby, Hochhalter, and Zhe of the University of Utah. AUTHOR Eric Lindgren: AFRL and ASNT Fellow eric.lindgren @us.af.mil CITATION Materials Evaluation 81 (7): 35–42 https://doi.org/10.32548/2023.me-04364 ©2023 American Society for Nondestructive Testing REFERENCES: Aldrin, J. C., D. S. Forsyth, and J. T. Welter. 2016. “Design and Demonstration of Automated Defect Analysis Algorithms for Ultrasonic Inspection of Complex Composite Panels with Bonds.” AIP Conference Proceedings 1706. https://doi. org/10.1063/1.4940591. Aldrin, J. C., E. A. Lindgren, and D. S. Forsyth. 2019a. “Intelligence augmentation in nondestructive evaluation.” AIP Conference Proceedings 2102 (1): 020028. https://doi. org/10.1063/1.5099732. Aldrin, J. C., E. K. Oneida, E. B. Shell, V. Sinha, K. Keller, J. K. Na, A. L. Hutson, H. A. Sabbagh, E. Sabbagh, R. K. Murphy, S. Mazdiyasni, M. R. Cherry, and D. M. Sparkman. 2019b. “Model-based inversion of eddy current data for classification and sizing of planar and volumetric discontinuities.” Vol. 44: Electromagnetic Nondestructive Evaluation XXII 44:80–85. J U L Y 2 0 2 3 • M A T E R I A L S E V A L U A T I O N 41 2307 ME July dup.indd 41 6/19/23 3:41 PM