Abstract
Wear resistance is an important property in the longevity and performance of additively manufactured (AM) dental materials. However, the existing ISO standardisation for wear testing remains limited, with protocols that fail to comprehensively account for the variability in testing conditions, material compositions, and antagonist interactions. The available data is highly scattered, making direct comparisons between studies challenging, and the inconsistencies in laboratory methodologies further complicate correlation with clinical outcomes. Additionally, the scarcity of clinical wear studies and the variability in patient-specific factors hinder the development of reliable in vitro-to-in vivo translation frameworks. Traditional in vitro wear testing is time-intensive, resource-demanding, and often fails to replicate in vivo conditions. One possible alternative to conventional wear testing is the application of machine learning algorithms, such as Long Short-Term Memory (LSTM) neural networks, which offer a more cost-effective and time-efficient approach to predicting material wear while reducing reliance on extensive experimental testing. This study aimed to assess the correlation between three established wear testing methods, develop an AI-driven predictive model for wear behaviour in AM resin-based materials, and explore various post-analysis techniques, with confocal microscopy being the primary method utilized for surface characterisation.
Three distinct wear testing methods—ball-on-disc, block-on-ring, and Reciprocation wear tests—were employed to evaluate AM restorative materials under varying antagonists, loading forces, and surface treatments. Surface characterization was conducted using confocal laser scanning microscopy (CLSM) and profilometry, with CLSM utilized for assessing surface roughness, antagonist evaluation, and volumetric loss. Vertical loss and coefficient of friction (CoF) measurements were obtained directly from the wear testing machine. A predictive model based on an LSTM neural network was trained using experimental wear data and validated through a Leave-One-Material-Out (LOMO) and Leave-One-Group-Out (LOGO) approach. The LSTM model achieved accurate wear predictions due to its ability to capture sequential dependencies and non-linear patterns in the wear progression data.
The results demonstrated low correlation between the three wear testing methods, possibly indicating that each test represents a distinct wear mechanism rather than providing interchangeable results. Glazing improved wear resistance, delaying surface degradation over prolonged testing. CoF remained stable across all tested materials and conditions, confirming its reliability as a mechanical parameter. Surface roughness increased significantly post-wear, with potential implications for bacterial adhesion and plaque accumulation. The LSTM model accurately predicted wear progression, illustrating its capacity to optimise material evaluation while minimising experimental duration.
This study underscores the necessity for method-specific wear testing rather than a universal approach, emphasizing the importance of selecting wear evaluation techniques based on clinical application. Furthermore, the successful implementation of machine learning in wear prediction provides a robust framework for improving the efficiency and standardization of wear testing protocols in AM dental materials. The integration of AI-driven modelling has the potential to enhance material selection, reduce experimental workloads, and refine predictive assessments in digital dentistry. This study is limited by the in vitro design, restricted load ranges, media type and lack of direct clinical data, which requires cautious interpretation of results.