In Additive Manufacturing, particularly Laser Powder Bed Fusion, improving part quality and reliability while reducing build failures remains a major industrial challenge. Part properties are heavily influenced by process dynamics, and deviations in parameters such as laser power, scan speed, or environmental conditions can lead to defects including porosity, residual stress accumulation, overheating, and geometric distortion. These issues directly affect the mechanical performance and dimensional accuracy of the final components, which is especially critical in high-value sectors such as aerospace and energy.

While commercial in-situ monitoring systems, such as EOS State OT and EOS State Meltpool, are capable of capturing high-frequency images and detecting real-time variations during the melting process through advanced optical sensors like photodiodes and long-exposure cameras, the interpretation and analysis of this data remains predominantly manual. Expert operators are required to identify anomalies or defects, and this reliance on subjective judgement limits scalability, repeatability, and the speed of quality control. Process corrections are typically implemented only after a failure has already occurred, resulting in a reactive rather than preventive approach to quality management.

Another significant challenge lies in the large volumes of in-situ monitoring data generated during each build, which are often underutilized due to the lack of effective analysis tools. The relationships between process data and the final quality of the part are frequently unclear or weak, especially for complex failure modes such as stress-induced cracks or thermal distortion. Current solutions are limited to simple data visualization and storage or rely on qualitative post-process evaluation, without providing the structured and automated analysis needed to extract actionable quality indicators during the printing process itself.

From an operational perspective, the absence of a unified mechanism to translate raw sensor data into clear, actionable insights for operators constrains the ability of SMEs like f3nice to scale LPBF production efficiently, reduce variability between builds, and accelerate parameter optimization for new materials or geometries. In addition, the available monitoring systems operate with limited integration into daily production workflows. This gap between monitoring capability and practical quality assurance represents the core challenge that the DIAMOND experiment aims to address.

Objective 1:

Develop an automated data management and analysis tool for LPBF in-situ monitoring systems that extracts meaningful quality indicators from sensor data and correlates them with physical characteristics of the printed part, such as porosity, residual stress, and geometric deformation. The system will replace the current manual and expert-dependent interpretation workflow with a structured, scalable, and objective approach to quality assessment. The target is to achieve at least a 30% improvement in defect identification accuracy compared to current semi-automated analysis methods, and a 30% reduction in data interpretation time.

Objective 2:

Develop and validate predictive models that link LPBF process signals to final part properties, enabling proactive quality control through real-time anomaly detection during the build process. By integrating these models into the production workflow, the system will allow operators to identify critical process deviations before defects propagate, supporting earlier intervention and reducing the number of failed builds by at least 20%. This shift from post-process diagnostics to operational decision-support represents a fundamental advancement in the management of LPBF quality.

Objective 3:

Deploy an operator-friendly Digital Intelligent Assistant that integrates an open-source Large Language Model trained on the LPBF ontology, a RASA-based conversational AI, and OVOS voice skills into a unified human-machine interaction framework. The DIA will be deployed via the WASABI Docker Compose framework and made available through the WASABI White Label Shop marketplace, enabling SME evaluation and adoption. The assistant will reduce cognitive load for operators, lower the learning curve for new personnel, and support decision-making via natural language interaction with quality indicators, predictive models, and process anomalies.

The DIAMOND experiment operates within the metal Additive Manufacturing sector, specifically targeting the Laser Powder Bed Fusion process. LPBF is increasingly adopted for the production of high-value, complex metal components in sectors such as aerospace, energy, advanced manufacturing, and industrial research and development. These sectors demand strict quality requirements, including tight tolerances on porosity, residual stress, dimensional accuracy, and mechanical performance, making reliable process monitoring and quality control essential.

The experiment will be carried out at the production facility of f3nice Srl, located in Lombardy, Italy. f3nice operates industry-standard LPBF systems equipped with in-situ monitoring solutions, providing a realistic industrial environment for testing and validating the Digital Intelligent Assistant. The operational environment encompasses both controlled experimental builds, designed to introduce known defects and test the response of the AI tools, and real production conditions using historical and ongoing manufacturing data. This dual approach ensures that the solution is validated not only in laboratory-like conditions but also under the variability and constraints of actual production workflows.

The primary target users of the DIA are LPBF machine operators, process engineers, and quality control personnel at f3nice, who will interact with the system through conversational and voice-based interfaces to monitor builds, interpret process data, and receive real-time alerts and recommendations. Secondary stakeholders include SMEs across the Emilia-Romagna manufacturing ecosystem, who will benefit from the replication and adoption pathways facilitated by DIH-ER through workshops, webinars, and one-to-one innovation assessments.

Relevant constraints include the need for compliance with GDPR and the EU AI Act for all data processing and AI-driven functionalities, as well as alignment with ISO/IEC 42001 and EU Trustworthy AI guidelines. The solution must integrate with existing commercial monitoring systems and their proprietary data formats and APIs, and must operate within the security and data governance requirements of an industrial production environment. Data privacy is ensured through federated learning and end-to-end encryption, allowing decentralized validation without the exchange of raw data between partners.