Autonomous AI Control of Mission Critical Cooling at Merck

A New Control Paradigm: Continuous Optimization with AI

Merck has been a healthcare research leader in the US since 1891. Merck's focus on responsible operations and modern innovation is paving the way for a more sustainable future with many investments in operational efficiency, including the implementation and scaling of Phaidra’s AI Virtual Plant Operator service.

As one of Phaidra’s first customers, the discussion with Merck leadership began nearly four years ago, in mid-2020. Merck aimed to meet its large goals for cost, energy, and emissions reduction by using Phaidra’s AI control system for mission-critical cooling at their largest manufacturing site. This 7-million-square-foot campus in West Point, PA, has four interconnected plant buildings forming a campus loop chilled water system with a total capacity of 60,000 refrigerant tons. This cooling system provides the precise temperature control (i.e., thermal stability) needed for safe 24/7 production of vaccines and other medicines, consuming almost 20% of the site's total energy.

Historically, these chiller plants were controlled by hard-coded programming logic and managed by rotating teams of operators. These operators monitored BMS control screens to make setpoint adjustments, manual start/stop equipment and maintain, repair or replace any equipment as needed. While these systems have a programmed standard sequence of operations (SOO), human operators often make manual changes based on their knowledge and experience. Every few years, as equipment wears out and production needs change, the control logic is updated with new programming to create a new SOO. This process is costly but necessary for optimizing these complex systems.

Welcome to West Point

West Point, Pennsylvania was chosen as the initial deployment site because it had recently undergone several years of investment projects to optimize the campus. These projects made the West Point facility extremely efficient, stable, and highly optimized. This made it the ideal place to test if AI control could enhance it further, beyond the physical upgrades and recently improved SOO.

Phaidra’s Virtual Plant Operator service offers a dynamic solution for daily operational control and uses artificial intelligence, specifically ‘reinforcement learning,’ to continuously improve energy efficiency while providing plant stability. Phaidra’s cloud-based service analyzes live operational data, calculates the best setpoints, and updates the local BMS directly. This allows Merck’s operators to respond in real-time to changes in production, environment, and equipment capability, ensuring better stability and reduced energy use.

Given the pressure of meeting strict and fast-approaching deadlines for corporate energy and emissions reduction goals, Merck authorized a demonstration of Phaidra’s service at one of the four chiller plant buildings in West Point, PA. To build confidence with the operations team, the demonstration was manual. The AI Virtual Plant Operator (also known as the ‘AI Agent’) sent setpoint recommendations to on-site staff for their review and implementation. Staff provided feedback on any changes they did not implement, allowing Phaidra to improve the AI agent’s understanding of the plant’s nuances. The AI agent was trained on the plant’s historical data and programmed to understand and respect all system constraints.

At the end of a month-long demo in the spring of 2021, the AI agent’s recommendations resulted in 16.2% energy savings compared to the local SOO results. Additionally, the analysis showed a 70.5% improvement in thermal stability and a 50.9% reduction in excess equipment run time. In short, the virtual plant operator provided a control strategy that improved plant safety, reduced equipment wear and tear, and lowered energy consumption.

Merck had already achieved a high level of optimization and safe operation, with out-of-bounds temperature events occurring less than 3% of the time in 24/7/365 operation. While energy efficiency was the top priority, Phaidra's AI agent also enhanced stability during the trial.

Transition from Manual Demo to Autonomous Deployment

Merck immediately approved the transition to a fully autonomous AI control service for all four chiller plant buildings. Both teams worked closely for nearly a year to develop and deploy an AI Virtual Plant Operator at each building. By April 2022, the four new AI agents were deployed and ready for autonomous control.

Collaboration with the Merck-West Point leadership and operations team was critical to helping Phaidra create and deploy the four necessary AI agents to optimize each chiller plant. By the end of 2022, they had been safely and autonomously operating each plant an average of 84% of the time. Phaidra’s service has a built-in bump-less transfer to local control triggered by operators at any time if they need to make equipment updates, changes, or additions.

During these ‘AI Off’ times, the AI agents still receive live operational data, so when reactivated, they understand changes that occurred and respond appropriately. Even though thermal stability improved, excess equipment runtime decreased, and energy consumption was trending downward, the results after the transition to autonomous control of all four plants were not as significant as the manual demonstration in a single plant. The missing piece Merck and Phaidra determined had to do with the fact that all four plants are interconnected and could work together to achieve better results.

To cover this final aspect of systemic optimization, Phaidra developed the 5th AI Virtual Plant Operator, deemed the ‘AI Conductor.’ This new conductor would take in data at a campus-wide level and coordinate each individual agent’s production of chilled water to optimize load balance between all plants accordingly. This AI Conductor ensures the other four agents work in harmony and optimize universally, much like a conductor harmonizes an orchestra of individual instruments.

After going live in early 2023, this AI Conductor balances load dynamically to further improve the energy-efficient production of a stable thermal system. Trends continue to show improvements, even as equipment ages and the agents dynamically respond to maintenance activities. Operators have realized the benefits of having AI virtual teammates and the sharing of institutional knowledge and productivity it enables.

What began as a project to minimize the energy intensity of mission-critical cooling has become a way to modernize and transform district loop chiller plant operations.

Much More to the Story

To learn more about the process of AI control deployment at Merck West Point and the results the AI Virtual Plant Operators have been able to drive over the course of the last 2 years, download our full in-depth case study.