(588cx) A Reinforcement Learning Framework for Automated Hmi Design in Process Industries

In process industries, Human-Machine Interfaces (HMIs) serve as a vital communication channel between operators and processes, directly influencing process safety, product quality, and overall efficiency. A well-designed interface provides operators with clear and context-specific information during normal and abnormal conditions. This allows them to monitor and control tasks effectively while minimizing errors. Conversely, a poor interface hampers operator response, introduces misunderstandings, and can lead to disruptions. Statistics reveal that about 70% of industrial accidents are due to human error (Mannan, 2013). It is therefore essential to create HMIs that support sound decision-making and reduce the likelihood of human error.

The development of industrial HMIs has traditionally been guided by standards from organizations such as the International Society of Automation (ISA). Many HMIs still rely on Piping & Instrumentation Diagrams (P&IDs) for Distributed Control System (DCS) displays (Lee et al., 2017). These displays often group data by physical equipment hierarchies (Jamieson et al., 2007). Despite advances in display technology, this conventional approach shows limited improvements in practical effectiveness and can increase the risk of overlooking early signs of abnormal events (Krajewski, 2014; Shahab et al., 2023). Traditional HMI design methods also depend on expert-driven or vendor-proprietary guidelines, making layouts difficult to generalize or adapt to changing process requirements.

To address these challenges, this work introduces a novel Reinforcement Learning (RL) approach that generates and optimizes HMI layouts—even without a labeled dataset of good HMIs. Recently, RL methods have been employed to automatically generate process flowsheets (Stops et al., 2023), demonstrating the broader applicability of RL in generative tasks for industrial operations. Our approach operates in three key stages. First, we define a minimal markup to capture essential process information—such as critical variables, alarms, control points, and screen slots (discrete placeholders where different widgets can be placed)—thereby providing a structured description of the requirements. Second, we implement an RL-based environment where each episode begins with an empty HMI layout. In some cases, we pre-place a minimal set of required widgets before letting the RL agent optimize the rest. At each step, the agent chooses actions such as adding and rearranging widgets (e.g., charts, control buttons, alarm banners) in specified screen locations to create HMI. A heuristic reward function encodes basic ergonomic and compliance rules—for example, penalizing overlapping elements, rewarding mandatory control widgets, and ensuring critical variables are visible. This reward function obviates the need for a large labeled HMI dataset by allowing the RL agent to discover high-quality layouts through iterative experimentation. Third, we analyze the agent’s learning progression by evaluating improvements in layout clarity and coverage of user requirements. Our experiments show that, within a few thousand training steps, the agent consistently converges on designs that balance minimal clutter with essential process visibility, outperforming random or purely heuristic baselines in both quantitative scores and qualitative inspection.

We demonstrated the potential of the proposed approach for designing HMI for an in vitro transcription process simulator—a typical upstream process for mRNA based therapeutics. The agent learned to place critical HMI widgets (e.g., temperature trend chart, pH monitor, alarm banner) without unnecessary duplication, achieving about a 30–40% decrease in layout penalty metrics (based on clutter and missing widgets) compared to naive initial configurations. The naive baseline was an unoptimized arrangement of widgets without RL-driven placement. Beyond reducing engineering effort, these results show how RL can generate a set of candidate HMI designs that blend clarity, functionality, and low error potential. In future work, we plan to integrate a bioprocess simulator to measure operator performance across varied tasks and identify error-prone layouts. We also intend to evaluate usability using a human digital twin, providing data-driven insights into how different interface designs impact error rates, operator workload, and overall process safety—and ultimately automating the design process with deeper scientific rigor.

Keywords: Reinforcement learning, Human-Machine Interface (HMI), Digitalization

References

Mannan, S. (2013). Lees' Process Safety Essentials: Hazard Identification, Assessment and Control. Butterworth-Heinemann.

Lee, S. T., Kim, S. Y., & Gilmore, D. (2017, October). Human-in-the-loop evaluation of human-machine interface for power plant operators. In 2017 IEEE International Conference on Systems, Man, and Cybernetics (SMC) (pp. 34-39). IEEE.

Jamieson, G. A. (2007). Ecological interface design for petrochemical process control: An empirical assessment. IEEE Transactions on systems, man, and cybernetics-Part A: Systems and Humans, 37(6), 906-920

Shahab, M. A., Srinivasan, B., & Srinivasan, R. (2023). Enhancing Human Machine Interface design using cognitive metrics of process operators. In Computer Aided Chemical Engineering (Vol. 52, pp. 3513-3518). Elsevier.

Stops, L., Leenhouts, R., Gao, Q., & Schweidtmann, A. M. (2023). Flowsheet generation through hierarchical reinforcement learning and graph neural networks. AIChE Journal, 69(1), e17938.s