Model-AI Chemometrics

---

Near-infrared spectroscopy (NIR) has become an established analytical technique for laboratories and continuous industrial processes, especially with the integration of AI chemometrics. A single NIR spectrum can provide comprehensive information regarding several physical and chemical properties, enabling one instrument to estimate composition, quality, and process performance without the lengthy procedures that typically accompany many conventional laboratory methods. This advancement underscores the significance of industrial AI in analytical applications.

However, the spectrometer is just one component of the wider measurement system. An NIR analyzer typically does not measure properties like density, octane number, sulphur, water, distillation point, or product composition directly. Instead, it records the interaction between near-infrared radiation and the sample, with a mathematical model converting the spectral information into estimated property values.

The analyzer's performance relies on three interconnected elements: the quality and stability of the spectrometer, the quality of the reference laboratory data, and the quality, relevance, and maintenance of the chemometric model, often powered by AI chemometrics.

Maintaining the third element—the chemometric model—can be particularly challenging over the analyzer's operational life. Modcon Systems Ltd. has released an updated version of Model-AI, its chemometric software designed for developing, validating, maintaining, and automatically updating models used with laboratory and process NIR analyzers, reflective of the growing trend in industrial AI. This new release is available in both automatic and semi-automatic modes, allowing the same platform to support routine users, laboratory specialists, and experienced chemometricians.

The software automates critical tasks such as data preparation, model training, validation, and intelligent outlier detection, making advanced modeling techniques accessible to operators with varying levels of experience. This automation is essential for real-time process analyzers, ensuring robust analytical performance.

Continuous attention to NIR models is paramount because a chemometric model encapsulates relationships derived from a defined set of spectra and reference laboratory results, and is not a permanent mathematical reality.

Though initial calibration may perform exceptionally well, the process conditions can change, leading to discrepancies with the original dataset over time.

Typical reasons for this drift include changes in crude oil, feedstock origin, product formulation variation, seasonal feed composition changes, introduction of new components or additives, catalyst ageing, equipment modifications, and spectrometer maintenance.

Given that a limited dataset may yield excellent statistics during testing but falter when real processes shift to a broader operational envelope, maintaining model reliability in industrial analytics is critical. Outliers, uninformative samples, and non-representative datasets can detrimentally affect performance, while well-chosen samples can significantly strengthen the model. Thus, model maintenance should be treated as an ongoing responsibility within analyzer lifecycle management rather than an occasional software task.

Traditionally, chemometric model development has been a manual endeavor performed by specialists, including data importing, spectrums matching, timestamps verification, data reviewing, property selection, outlier identification, and validation. While effective, this process is time-consuming and can be further complicated when facilities lack full-time chemometric specialists, often relying on external support.

This dependence can lead to a range of practical issues, including delays in model updates, underutilized reference samples, diminished analyzer representation of the process, accumulation of minor model issues, declining operator confidence in results, and excessive laboratory testing.

The Model-AI software addresses these challenges by automating and streamlining various model development processes, which is increasingly important in the context of real-time process analyzers.

Model-AI simplifies the creation, validation, deployment, and management of mathematical models for correlative analyzers, fostering better integration across laboratory and process-analysis environments. It supports distinct operating approaches, both automatic and semi-automatic, catering to the differing needs of industrial users.

Automatic mode is designed for routine model development and maintenance, ensuring key modeling decisions remain within predefined technical and quality constraints. This mode enhances consistency by standardizing processes across multiple analyzers, laboratories, or production sites.

Data screening and intelligent outlier detection are crucial components of the model-training process. The software verifies the quality of data before training begins, which mitigates the risks associated with using incorrect laboratory results or unrepresentative sample data. In addition, automatic model training efficiently adapts to changes in the provided dataset, balancing factors such as accuracy, robustness, stability, and resistance to overfitting.

Automatic validation ensures that models can reliably predict samples not used during the fitting process. This aspect is vital for real-time process analyzers, where accuracy in varying conditions directly affects operational efficiency. The software employs a structured workflow enabling both automatic and semi-automatic operations, with the latter allowing users, particularly those in specialized roles, to intervene when necessary.

One of the most transformative capabilities introduced in the latest version of Model-AI is support for automatic model updating. In a dynamic industrial environment, an NIR model must evolve alongside process changes while avoiding pitfalls associated with uncontrolled adaptations based on poor data or temporary disturbances. A controlled sequence for updates ensures robust model governance, with ongoing monitoring of various parameters post-deployment.

In laboratory settings, Model-AI enhances the development of models using spectra from benchtop or at-line NIR instruments. Typical applications span across product-quality screening, incoming-material inspection, and formulation analysis. Conversely, process NIR applications face challenges due to the continuous nature of operation, necessitating models capable of distinguishing useful compositional data from unrelated physical variations.

Model transfer between laboratory and process environments requires careful consideration of various factors, including differing instrument characteristics and preprocessing methods. Model-AI maintains the connection between laboratory reference data and in situ analyzer spectra, reinforcing its role in fostering data integrity across diverse applications.

Ultimately, the effectiveness of chemometric models hinges on the reliability of reference data, making it paramount to establish appropriate procedures before development. Automatic modeling processes streamline data management while ensuring that model governance maintains traceability and accountability across updates. 

As NIR spectroscopy plays a pivotal role in modern analytics, particularly with advancements in AI chemometrics and industrial AI, the updated Model-AI empowers users to maintain reliable NIR models seamlessly, adapting them to meet evolving analytical demands. Keeping these models current is not an optional task; it forms an integral part of the measurement chain.