The process industry, characterized by continuous, large-scale production, is vital to modern manufacturing. In recent years, integrating big data and artificial intelligence (AI) into process manufacturing has emerged as a strategic way to boost efficiency, resource utilization, and competitiveness.

While general foundation models like GPT are skilled at text, image, and multimodal tasks, they struggle to meet the unique demands of the process industry. Early applications in discrete manufacturing—such as Siemens’ programmable logic controller (PLC) code generation or OpenAI’s AI-driven robotic control—show promise for artificial general intelligence (AGI). However, the process industry faces distinct challenges: complex physical-chemical mechanisms, stringent control requirements, and lengthy R&D cycles. These make general models inadequate for industrial needs in trustworthiness, adaptability, and interoperability.

To address this gap, Professors Lei Ren and Bohu Li’s team has proposed ProcessFM, the first foundational model framework designed for the process industry, which was published in Engineering recently. ProcessFM is equipped with six core capabilities: mechanism cognition, knowledge Q&A, simulation and generation, process control, optimization and decision-making, and scientific discovery. ProcessFM integrates data, mechanisms, knowledge, and computation across four systematically layered components: the resource level, base level, adaptation level, and application level.

At the resource level, ProcessFM relies on three key elements: diverse industrial data (e.g., sensor signals, images, documents, and computer-aided engineering (CAE) files), embedded domain knowledge such as chemical and physical mechanisms, and strong computational resources. These form the foundation for building scalable and intelligent models, with cloud computing-assisted large-scale training and edge computing-assisted real-time applications.

The base level focuses on core technologies for model training and application. This includes multimodal pre-training to extract unified features from different data types, and mechanism-embedded fine-tuning to incorporate domain knowledge. Human–agent–cyber-physical–social system (CPSS) collaboration also allows intelligent agents to perceive, plan, and act in dynamic industrial environments.

At the adaptation level, ProcessFM uses targeted fine-tuning for task-specific and domain-specific needs. Six key model types are created: mechanism cognition, knowledge Q&A, simulation and generation, process control, optimization and decision-making, and scientific discovery. Post-pre-training and knowledge internalization further adapt foundation models for fields like power, petrochemicals, and metallurgy.

At the top application level, ProcessFM turns foundation models into proactive industrial agents. These agents enable intelligent services, from real-time process control to autonomous decision-making and collaborative optimization. By interacting with CPSS environments, they improve human–machine collaboration and performance in specialized industrial settings.

Looking ahead, ProcessFM is poised to push the boundaries of AI in the process industry with innovations in multi-modal and mechanism-informed learning, explainable decision intelligence, and real-time edge deployment. On one hand, by integrating production data and domain knowledge, the accuracy and reliability of the models will be improved. On the other hand, optimizing edge computing capabilities will enable models to respond quickly to on-site demands. Additionally, the interpretability of the models will help engineers better understand and trust AI-driven decisions. By integrating physical principles and domain expertise into AI models, it will be able to make industrial intelligence not just more capable, but also transparent and trustworthy.

The paper “Foundation Models for the Process Industry: Challenges and Opportunities,” authored by Lei Ren, Haiteng Wang, Yuqing Wang, Keke Huang, Lihui Wang, Bohu Li. Full text of the open access paper: https://doi.org/10.1016/j.eng.2025.03.023. For more information about Engineering, visit the website at https://www.sciencedirect.com/journal/engineering.