Engineering AI-Powered Digital Twins: Setting the Right Boundaries for Accurate Insights

To get more out of your AI questions and responses, focus on the way the application you are using is built. AI is like electricity and our particular use case for buildings has to have a particular application structure.

The Problem with Unbounded AI

Much of the disappointment of asking questions directly into current Large Language Models (LLMs) is that they're unbounded. When you type "How is my HVAC system operating?" into a general AI tool, the results will be extremely disappointing and generic.

Why? Because the AI has no context about:

• Which building you're referring to
• What type of HVAC system you have
• The specific equipment models and configurations
• Current operational parameters
• Historical performance data

AI is Like Electricity

Just as electricity needs to be channeled through specific devices and circuits to be useful, AI needs to be channeled through properly structured applications to deliver accurate, actionable insights.

For the AI responses to be more accurate, the application has to apply bounds for the AI responses through a digital twin framework. For creative solutions, the bounds have to be in place on our unique building and building systems.

The Three Bounds Framework

How do applications like Syyclops bound the AI responses to deliver accurate insights? Through a three-tiered boundary system:

First Bound: Standard Ontology

The data of your building gets converted into a standard ontology—a preset organization structure that reflects our buildings. We recommend OmniClass 23 for this, as it's the most comprehensive classification system available.

Example: HVAC System Organization

• Classification: OmniClass 23-35 00 00
• Category: HVAC Systems
• Subcategories include:
  • Air Handling Units (23-35 11 00)
  • Boilers (23-35 31 00)
  • Chillers (23-35 33 00)
  • Pumps (23-35 51 00)

This first bound ensures that when you ask about HVAC, the AI knows exactly what categories of equipment to consider.

Second Bound: Asset-Specific Knowledge Graph

Taking the exact assets of our unique building and putting them into the Digital Twin's knowledge graph creates the second boundary layer.

This goes deeper to capture:

• Specific Equipment Details - Make, model, serial numbers
• Component Specifications - That particular air handler has three coils: preheat, reheat, and chilled water
• Capacities and Ratings - Each component's specific operational parameters
• Relationships - How equipment connects and interacts

Example: Air Handler Asset Profile

• ID: AHU-01
• Type: Air Handling Unit
• Manufacturer: Trane
• Model: TAM150
• Components:
  • Preheat Coil: Hot Water, 150 MBH capacity
  • Reheat Coil: Steam, 100 MBH capacity
  • Chilled Water Coil: 250 tons capacity
• Serves: Floor 1 and Floor 2

Third Bound: Operational Sequences and Real-Time Data

The final bound compares real-time data across all of the spaces to give nuanced answers. This involves:

• Live Sensor Data - Temperature, pressure, flow rates, energy consumption
• Operational Sequences - How the system should be running
• Performance Baselines - What "normal" looks like for this specific building
• Comparative Analysis - Cross-referencing data across zones and equipment

Why This Framework is Critical

This bounded approach is crucial for two reasons:

1. Accurate Diagnostics

When you ask "How is my HVAC system operating?", the bounded AI can respond with:

"AHU-01 is currently operating at 87% efficiency. The preheat coil is consuming 15% more energy than baseline due to increased outdoor air ventilation in Zone 2B. This is compensating for elevated CO2 levels detected at 1,250 ppm, which is above your 1,000 ppm threshold."

Instead of a generic response about HVAC systems in general.

2. Actionable Recommendations

The next critical question—"How do I fix it and where do I make my adjustments?"—becomes answerable:

"To optimize performance: 1) Review Zone 2B occupancy schedule to verify ventilation requirements. 2) Check damper position on Outdoor Air intake for AHU-01. 3) Consider adjusting CO2 setpoint to 1,100 ppm during peak hours. Estimated energy savings: 12-15% on this unit."

Building the Bounds: Implementation Steps

Step 1: Establish Your Ontology

Choose a standard classification system (like OmniClass 23) and map all building assets to it.

Step 2: Create the Knowledge Graph

Build a comprehensive database of your specific equipment, including all specifications, relationships, and dependencies.

Step 3: Integrate Real-Time Data

Connect IoT sensors and BMS systems to feed live operational data into the digital twin.

Step 4: Define Operational Sequences

Document how systems should operate under different conditions and scenarios.

Step 5: Train and Validate

Work with domain experts to validate AI responses and continuously improve the bounds.

Real-World Application

Unbounded AI Query:

Question: "Should I replace my boiler?"

Generic AI Response: "Boilers typically last 15-30 years. Consider factors like age, efficiency, and repair costs..."

Bounded AI Query (Through Digital Twin):

Question: "Should I replace my boiler?"

Bounded AI Response: "Your Cleaver-Brooks CB-700-300 boiler, installed in 2012, is currently operating at 78% efficiency—down from its rated 85%. Based on your natural gas costs ($0.85/therm) and current run-time of 6,200 hours annually, a new high-efficiency boiler (95% AFUE) would save approximately $18,500/year. ROI period: 4.2 years. However, your burner assembly was serviced 3 months ago. Recommend monitoring efficiency for 2 more months before making replacement decision."

The Future of Bounded AI

As AI technology evolves, the sophistication of these bounds will increase:

• Self-Learning Bounds - Systems that automatically refine their understanding based on outcomes
• Dynamic Bounds - Boundaries that adjust based on building usage patterns
• Collaborative Bounds - Sharing insights across similar buildings while maintaining specificity
• Predictive Bounds - Anticipating what information will be needed before it's asked

Conclusion

The power of AI in building operations isn't just in the AI itself—it's in how we structure the applications that deliver AI insights. By creating proper boundaries through:

1. Standard ontologies
2. Asset-specific knowledge graphs
3. Operational sequences with real-time data

We can drastically increase AI response reliability and actionability.

Remember: AI without bounds is like electricity without circuits—it has potential energy but no practical application.

"Come engineer with me." - Raj Setty

Want to see how bounded AI can transform your building operations? Schedule a demo to experience the Syyclops platform.

Read the original LinkedIn post here.