In the last two articles, we discussed the First Principles Blending Model (FPBM) and AI / Machine Learning if they can deliver the best blend in the most profitable and efficient ways. Each of the previously discussed methods has pros and cons, one being better than others in certain respects.
Now, we explore the third strategy to make the best of worlds, integrating the First Principles Model and AI / Machine Learning models. One has to be careful about integrating its components. We will discuss this in this article. Let us follow the food chain of the entire blending process from end to end and examine where FPBM and AI/Machine learning can benefit most.
The process progresses as follows:
- Process Streams: Process Streams feeding the component tanks refer to raw material inputs delivered into the tanks for storing blending components. These streams can vary in type and source, and careful monitoring is essential to maintain the quality and quantity of the blending process.
- Components Tank and Qualities: Each component tank holds a specific product type with unique characteristics. The quality of the contents, such as octane number, sulfur content, or viscosity, must be measured and controlled to meet the requirements of the final blend.
- Component Tank Inventories and Product Tank Heel: The inventory levels of each component tank need to be tracked to manage supply and demand effectively. The ‘heel’ refers to the residual amount left in a product tank after draining, which must be accounted for in the blending process to avoid quality deviation.
- Recipe Formulation: This is the process of determining the optimal mix of components to achieve the desired quality for the final product. Formulation considers the qualities of individual components and the specifications of the end product.
- Flow Controllers: These devices regulate the rate at which components are added to the blend. Accurate flow control is crucial to maintaining the integrity of the blend recipe and achieving consistent product quality.
- Recipe Mixing Rules or Blend Model: The blending model outlines the mathematical rules and constraints that govern the mixing process. This rule could include linear or non-linear blending equations that predict how component qualities will combine in the final product.
- Predicted Blend Qualities at the Blend Header: Before blending, predictions are made about the expected qualities of the blend based on the recipe and individual component qualities. This prediction happens at the blend header, where components come together.
- Analyzed Blend Qualities at Blend Header or Lab Analysis: Once the blend has been initiated, samples are taken from the blend header for analysis. This real-time analysis confirms that the blend meets the expected quality criteria or if any adjustments are needed.
- Quality Integration of Product Tank: After blending, the mixed product is stored in a tank where the different qualities of the components integrate over time. This integration must be monitored to ensure the final product remains homogeneous and meets specifications.
- Final Predicted Tank Quality: Using models and initial measurements, the final quality of the product in the tank is predicted. This prediction helps plan subsequent blending operations and fulfill product specifications before lab analysis confirms the quality.
- Lab Analysis of Product Tank: After blending and settling, samples from the product tank are analyzed in a lab to confirm the actual quality of the final product. This step validates the predictions and ensures compliance with the required specifications.
- Bias or Errors in Tank Quality: Bias or Errors in Tank Quality prediction compared to lab analysis refer to the differences between the predicted qualities and the actual lab results. Understanding these biases or errors is essential for refining predictive models and improving future blends.
- Bias Feedback to Step 7 to Correct the Predicted Blend Header Quality: The information gained from understanding biases and errors is fed back into the predictive models at the blend header. This iterative feedback loop aims to correct and improve the accuracy of future blend quality predictions.
These steps constitute a comprehensive blending strategy that ensures the final product meets the desired specifications and quality standards. Each step is interdependent, requiring accurate data and precise control to achieve the best outcomes.
So, where can AI play in the 13 steps outlined above?
AI’s significant contribution is in predicting blend quality at the blend header. It is because multi-plexed online analyzers are installed at the blend header. The multiplexed analyzer scans the blend every 2 minutes, and the component tank exits once every 15-20 minutes, providing near real-time qualities of the components.
AI Models are developed for spectrum-based analyzers to train the spectrum. AI is then trained to predict the blend qualities using linear or non-linear blend models from first principles and online historical data of analysers.
AI models minimize interruptions due to downtime of online analyzers—the first Principle blends model inputs to the AI model. We will discuss more on this in a case study in the following article in the series.
This diagram shows how First Principle Blend Model (FPBM) and AI / Machine Learning can be integrated:
We will discuss more on the subject in the following article.
Integrating First Principle Models with AI / Machine Learning (ML) in a hybrid approach combines the strengths of physics-based modeling with data-driven insights, offering unique advantages and certain limitations.
This hybrid methodology can be applied across various fields, including chemical engineering, materials science, finance, and environmental modeling.
Here are the PROS and CONS of using a hybrid model that integrates First Principle blend models with AI/ML:
PROS
- Improved Accuracy and Predictive Power: By combining the theoretical underpinnings of First Principle models with the pattern recognition capabilities of AI/ML, hybrid models can achieve higher accuracy and predictive power than either approach alone. This is particularly useful in complex systems where purely data-driven models may miss underlying physical laws.
- Reduced Data Requirement: First Principle models require less empirical data to make accurate predictions, as they are based on fundamental laws. When combined with AI/ML, the hybrid model can leverage these laws to make predictions even with limited data, overcoming one of the main limitations of pure data-driven approaches.
- Enhanced Generalization: Integrating fundamental principles helps generalize the model beyond the range of the training data. This is particularly beneficial in scenarios where data is scarce or expensive.
- Increased Robustness and Reliability: Incorporating First Principle models can increase the robustness and reliability of predictions by ensuring that the outcomes adhere to known physical laws, reducing the likelihood of implausible results often produced by black-box AI/ML models.
- Interpretability and Transparency: Hybrid models can offer greater interpretability by linking predictions to underlying physical or theoretical concepts, addressing one of the critical concerns in applying AI/ML models.
CONS
- Model Complexity: Hybrid models can be more complex to develop and implement than First Principle or AI/ML models alone. This increased complexity may require more expertise and computational resources.
- Integration Challenges: Effectively combining First Principle models with AI/ML approaches can be challenging, especially ensuring that the model components interact correctly and that the integration improves performance.
- Data and Model Mismatch: There can be difficulties in ensuring that the data-driven part of the model correctly interprets the outputs of the First Principle models, potentially leading to errors or inefficiencies in the hybrid model.
- Computational Demand: Depending on the complexity of the First Principle models and the AI/ML algorithms used, hybrid models can be computationally demanding, requiring significant processing power and potentially leading to longer training times.
- Overfitting Risk: There is a risk that the AI/ML component may overfit to the noise in the data rather than capturing the underlying trends, primarily if not adequately regulated or if the integration does not effectively leverage the First Principle guidance.
Summary
In summary, hybrid models that integrate First Principle Blend Models with AI/ML offer a powerful approach to modeling complex systems by combining the strengths of both methodologies. However, their development and implementation come with challenges that must be carefully managed to realize their potential benefits fully.