ADVANCED MATERIALS & PROCESSES | SEPTEMBER 2024 18 Design for Additive Manufacturing (DFAM). ML aids in DFAM by enabling the creation of complex structures that are difficult or impossible to fabricate using traditional manufacturing methods. ML algorithms can optimize design parameters for AM, considering the unique capabilities of AM technologies to produce parts in a single step without the need for assembly[2]. Companies such as Nvidia[3] and AutoCAD are providing developmental workflows to transform digitization and boost design agility, and productivity. Process Optimization and Control. ML techniques are used to optimize processing parameters in AM, such as layer thickness, build orientation, and temperature, to improve the quality and final mechanical properties of the manufactured parts[4]. By analyzing data from previous builds, ML algorithms can predict optimal parameters and identify potential issues before they occur (Figs. 2b and 2c). Anomaly Detection and Quality Assurance. ML algorithms can detect anomalies during the manufacturing process by analyzing data from sensors and monitoring equipment. This includes identifying defects[5], predicting mechanical properties, and ensuring consistent quality across production batches. Techniques such as regression, classification, and clustering are employed to classify and evaluate the performance of various ML algorithms in these tasks (Fig. 2). Production Control. In the context of smart manufacturing, ML-integrated assembly processes can automatically adjust errors in real time to prevent waste and improve efficiency (Fig. 2a). This involves using ML for real-time decision-making and process adjustments based on continuous data analysis. ADVANCES IN ARTIFICIAL INTELLIGENCE While significant research is occurring in ceramic AM utilizing ML, these endeavors often rely on conventional trial-and-error methodologies or simplistic design of experiments (DOE) to establish initial parameters, subsequently requiring extensive testing and iterative refinement to predict and optimize. Such processes are not only time-consuming, but also demand considerable human effort. However, advancements in AI present novel avenues for revolutionizing AM. Now that AI is widely available, various techniques can be more universally adopted. Self-supervised Learning for Tool Use. This innovative approach harnesses self-supervised learning techniques empowering language models to autonomously discern the optimal application of various tools without explicit task-specific guidance. This departure from traditional methods could represent a significant leap, reducing the dependency on human intervention. Integration with Pretrained Models. By including pretrained models from language models endowed with billions of parameters, this approach could leverage the inherent language processing capabilities and also extend functionality to include interaction with external tools via application programming interfaces (API) calls. Integration of this type substantially enhances performance and problem- solving capabilities. Superior Zero-shot Performance. This class of supervised learning model exhibits impressive zero-shot performance by predicting missing values across diverse tasks such as anomaly detection and mechanical/ thermal detection surpassing larger models and baseline competitors. It offers a versatile approach to look toward real-time printing scenarios where generalized classification is a major requirement. AI holds promise in AM through its capacity to predict and optimize processing parameters. RL techniques can adaptively learn and refine strategies for AM, enabling systems to autonomously optimize printing processes. Genetic algorithms and evolutionary computational methods offer avenues for exploring vast parameter spaces, facilitating the discovery of optimal configurations efficiently. APPLICATION OF ML AND AI TO CERAMIC AM Natural language processing (NLP), particularly through large language models (LLMs), offers an innovative approach to leverage existing literature and empirical data for predictive analysis in ceramic AM. These advanced models can be trained to meticulously extract and organize information from diverse sources such as research papers and industry reports, enabling them to identify patterns, correlations, and principles that are involved in the ceramic AM process. Additionally, including techniques like few-shot learning can effectively manage data sparsity, thereby improving the accuracy of the database for our applications. After populating and training the LLM, it becomes an invaluable tool for synthesizing knowledge and deriving insights into optimal process parameters for ceramic AM. When provided with specific inputs such as material composition, particle size, printing temperatures, layer height, and printing speed, the model can forecast critical outcomes like dimensional accuracy, printing process parameters, mechanical properties, and surface finish for the printed components. Moreover, incorporating AI techniques, such as RL, augments the model’s predictive capabilities. RL algorithms empower the LLM to iteratively learn from experimen- tation feedback, continually enhancing its predictions to better align with the evolving conditions of the manufacturing environment. This adaptive learning process enables the model to efficiently navigate through complex decision-making scenarios, optimizing process parameters to achieve the desired results, thus expediting the advancement of ceramic AM processes. Additionally, enhancing the predictive model with real-time sensor data from 3D printers allows for on-the-fly adjustments to the process parameters, informed by immediate monitoring of printing conditions, particularly for advanced ceramics[6]. The system’s
RkJQdWJsaXNoZXIy MjMzNTA5MA==