Development and Assessment of Machine Learning Advisory Methods in Continuous Casting

Continuous casting is a widely adopted industrial process for the production of aluminum alloy billets; however, the selection of process parameters is still largely guided by empirical knowledge, operator experience, and iterative trial-and-error procedures. Such practices often result in suboptimal operating conditions, increased material waste, and reduced process efficiency. This work addresses these limitations through the development of a data-driven digital advisory framework capable of recommending optimal process parameter configurations under realistic industrial constraints, enabling stable, repeatable, and high-quality production while supporting real-time decision-making.

The proposed methodology is based on an Arbitrary Conditioning Variational Autoencoder (AC-VAE), a generative modeling framework designed to learn the joint distribution of process parameters and corresponding process-state features. The model allows conditional generation of feasible parameter sets given a subset of fixed inputs, reflecting practical operational constraints. Training is conducted using a dataset of high-fidelity CFD simulations and computationally efficient ROMs, ensuring physical consistency and scalability. A multimodal extension is explored, incorporating image-based representations of grain structure to enrich the feature space and capture microstructural characteristics.

The results demonstrate that the AC-VAE outperforms a k-beam search baseline in solution quality, while achieving lower and more stable inference times, making it suitable for real-time advisory applications. Pre-training on ROM-generated data reduces dependence on computationally expensive CFD datasets; however, the observed performance gains remain moderate due to distributional discrepancies. The integration of image-based inputs enhances feature reconstruction but does not substantially improve predictive performance, although it provides additional flexibility in defining quality metrics during inference.

Overall, the findings highlight the effectiveness of generative modeling approaches, particularly AC-VAE architectures, in addressing inverse design and parameter recommendation challenges in continuous casting. The proposed framework establishes a robust foundation for real-time digital advisory systems, contributing to more adaptive, data-driven, and resource-efficient manufacturing processes aligned with the vision of smart industrial systems.