Resumen: 

Carbon Fiber Reinforced Polymers (CFRPs) are among the most widely used materials in the aviation, automotive, industrial, and aerospace sectors due to their exceptional mechanical properties and high strength to weight efficiency. Fast and reliable Non Destructive Testing (NDT) is critical for the CFRP industry. NDT covers a range of techniques and plays a pivotal role in the development of novel materials by enabling the assessment of their properties and verifying the reliability of new manufacturing methods. During production, NDT is essential for quality assurance, ensuring that manufactured samples comply with required specifications. Additionally, NDT is invaluable for maintenance applications, allowing for the detection of defects caused by operational use. One key defect in CFRPs than can significantly influence mechanical performance is porosity. Porosity can lead to degraded mechanical properties and cause early catastrophic failure of parts, making its precise evaluation essential. The void volume fraction, which quantifies the percentage of a sample’s volume occupied by voids, is a widely recognized metric for assessing porosity. Accurately measuring void volumetric fraction requires NDT techniques capable of high resolution 3D data acquisition, such as X-ray Computed Tomography (XCT). While lab XCT offers unparalleled micron up to submicron resolution and detailed internal structure visualization, it is expensive, time consuming, and demands strict experimental conditions. In contrast, ultrasonic testing (UT) is a highly efficient NDT method extensively used in the CFRP industry for its rapid diagnostic capabilities across varying sample sizes. However, UT faces limitations in resolution and clarity of internal structural representations compared to XCT. To bridge this gap, our research focuses on combining the efficiency of UT with the precision of XCT through artificial intelligence (AI). Specifically, we aim to leverage UT data to predict porosity metrics equivalent to those derived from XCT. We developed an end to end data pipeline, including automated data preprocessing and registration for dataset generation, and several deep learning models designed to predict the void volumetric fraction from UT measurements.