MET Conference Agenda

Artificial intelligence (AI) and machine learning (ML) are increasingly promoted as universal tools for enhancing measurement and classification systems, including spectroscopic techniques. In many recent studies, AI models have been integrated into existing analytical workflows with the implicit assumption that they will inherently provide higher accuracy, greater robustness, speed, or more "intelligent" behavior than classical approaches. This paper challenges that assumption in the context of deterministic spectros copy systems, with a particular focus on Laser-Induced Breakdown Spectroscopy (LIBS) for aluminium and light-metal sorting. In such systems, key physical processes (laser–matter interaction, plasma formation, chemical element-specific light emission, and optical detection)are governed by well-established laws related to physics- and chemistry. When paired with proper calibration and mathematically rigorous algorithms, these processes support models that are transparent, reproducible, and in many cases, close to theoretically optimal under clearly defined conditions. In contrast, AI models function as statistical approximators. Their predictions depend on finite training datasets, model architectures, and hyperparameter configurations rather than on explicit chemical structures and compositions. This introduces additional layers of uncertainty and opacity to systems whose behavior can already be described using deterministic, chemistry- and/or physics-based principles. In high-throughput, high-precision or safety critical industrial contexts such as LIBS methodology-based metal scrap sorting, this opacity can hinder validation, troubleshooting, and long-term maintainability. The central claim of this paper is that, in measurement and classification tasks where input–output mappings are governed by known physical and chemistry laws and can be modeled using well-founded algorithms, AI does not provide inherently better outcomes. At best, it may approximate the same solution, while adding complexity and uncertainty. A case study -taken from industrial application- will be presented. We argue that the perceived added value of AI in such deterministic systems is frequently overstated, and that method selection should begin with physical/chemical models and proven algorithms, not AI by default. Questions arise - what the users’ preferences are when adopting the advanced LIBS sensor sorting technology into their metal scrap sorting process line?