Building hybrid AI models in chemical engineering: A tutorial review

Modern machine learning (ML) methods excel at learning from vast datasets and have demonstrated exceptional performance in conventional applications like text prediction, recommender systems, and chatbots. However, their application in science and engineering is constrained by challenges such as limited explainability, susceptibility to hallucinations, and a lack of grounding in first-principles knowledge. These limitations could be overcome by incorporating symbolic or classical artificial intelligence (AI) methods, which have been applied in chemical engineering for more than four decades. This paper outlines a systematic approach to incorporating domain knowledge into the AI/ML workflow, resulting in the development of hybrid AI models. Our proposed four-stage process includes (1) knowledge assessment, (2) domain-informed problem formulation, (3) selection of an appropriate AI/ML model, and (4) model validation. Additionally, we present six commonly used templates for hybrid AI development: feature engineering, customized knowledge representation, imposition of additional constraints, integration of these approaches, custom model architecture design, and end-to-end domain-specific AI models. These templates are organized by increasing levels of “hybridization”, reflecting progressively more advanced integration of domain knowledge. The goal is to migrate from the paradigm of large language models (LLMs) to large knowledge models (LKMs), which are better positioned to meet the unique demands of science and engineering applications.