Toward effective generalisation of machine learning for accurate and high-speed dimensional XCT

X-ray computed tomography (XCT) has become a fundamental tool in industrial metrology, offering non-destructive, high-resolution inspection of complex components, as well as enabling accurate dimensional measurements. However, fast XCT scanning - necessary for high-throughput industrial applications - often leads to reduced image quality and compromised measurement accuracy. Machine learning (ML) techniques have demonstrated potential for improving XCT reconstruction and metrological performance, but their reliability depends on the generalisation capability of the model across different scanning conditions. This study investigates the effects of ML generalisation on measurement accuracy in fast XCT, proposing an experimental framework to evaluate how different levels of generalisation across diverse XCT scanning factors influence metrological performance. The proposed framework is outlined in Figure 1a. A reference object consisting of a calibrated hole plate was scanned under standard conditions using a metrological micro-focus XCT system (see the regular CT in Figure 1b). Fast scans were acquired under varying conditions, including material composition, scanning parameters, and part orientation (see the fast CT example in Figure 1b). A generative adversarial network (GAN)-based model was trained to enhance image quality, initially under conditions characterised by limited generalisation. The effect of increasing generalisation was systematically evaluated by progressively incorporating additional scan conditions into the training set and analysing the resulting form error absolute deviation (FEAD), as shown in Figure 1c for the two most influential scanning factors. Such most critical factors affecting measurement accuracy on high-speed XCT scans enhanced by the GAN-based model introduced above were determined to be the part orientation and material. Low-generalisation models exhibited substantial deviations in form errors, particularly for stainless steel components and highly tilted orientations (e.g., 52°), where FEAD values reached up to 0.35 mm. Increasing the model’s generalisation by incorporating a wider range of scanning conditions significantly reduced FEAD by up to 80%. Figure 1d visually demonstrates this accuracy improvement: the high-generalisation (HGD) model produces clearer hole contours compared to the low-generalisation (LGD) model, where contour blurring significantly affects measurement accuracy. These findings provide key insights into ML implementation for XCT metrology, contributing to the standardisation and traceability of AI-enhanced measurement processes. By enhancing ML training strategies, this research lays the groundwork for reliable, ML-enhanced XCT applications that balance accuracy and efficiency in industrial applications.