Hydrocracker catalysts are carefully optimized for selectivity vs. activity, balancing these properties to achieve high efficiency in refining processes. These catalysts typically consist of three key ingredients: amorphous Si/Al, crystalline Si/Al, and base metals. To accelerate the formulation process, a combination of Generative AI, Design of Experiment and optimization models is employed:

LLM Fine-tuning for Knowledge Retrieval:

The first step involves fine-tuning a Large Language Model (LLM) on domain-specific documents and research papers related to catalyst formulation. This accelerates the knowledge retrieval process, allowing experts to quickly explore various ingredient combinations. By integrating historical data, the LLM can suggest optimal ratios of amorphous Si/Al, crystalline Si/Al, and base metals, while rejecting infeasible combinations early in the process. This dramatically reduces the time spent on trial-and-error formulations.

Hypothesis Generation with DOE and AI Validation:

The next step is hypothesis generation through Design of Experiments (DOE), where multiple combinations of the catalyst ingredients are generated. The LLM, fine-tuned for the catalyst domain, is then used to assess these combinations, eliminating infeasible hypotheses before they move forward to experimental validation. This AI-driven pre-screening ensures that only the most promising formulations are tested, saving time and resources.

Optimization Model for Selectivity and Activity:

To fine-tune the catalyst for optimal selectivity and activity, historical experimental data is utilized to build an optimization model. This model considers previous test results and the properties of the ingredients, helping to refine the formulation process. The goal is to achieve the best balance between selectivity and activity, ensuring that the catalyst performs efficiently under the specific conditions of hydrocracking.