Conference Agenda

EnergyPlus practitioners face a persistent challenge: accessing EnergyPlus's comprehensive modeling capabilities requires navigating complex object structure and extensive input requirements. Traditional user interface solutions add abstraction layers that simplify the UI, but compromise direct access to EnergyPlus's full capabilities.

This presentation describes the implementation of context-aware AI integration embedded directly within EP3, a commercial EnergyPlus user interface designed for direct interaction with EnergyPlus. Unlike other recent AI tools building simulation tools (ex the EnergyPlus MCP) that operate as external applications or post-processing scripts, this AI integration functions within the modeling environment itself. The AI integration provides a natural-language interface for understanding and manipulating files in real-time during the modeling process, reducing the need for abstraction within EP3.

The team developed a system that enables AI assistance without custom model training. A key insight is that large language models pretrained on EnergyPlus require no additional training to assist with EnergyPlus modeling when:

1. Interactions are structured around native EnergyPlus file formats

2. Prompts are carefully engineered

3. Additional resources are provided to the LLM

4. Suggestions made by the LLM are validated, and returned to the LLM with comments until they pass validation

5. Sufficient context is given to support the LLM

By conducting all AI interactions using standard EPJSON format, the system leverages existing AI knowledge of EnergyPlus. This implementation is based on EP3's pre-existing EnergyPlus import functionality.

The AI operates through an embedded chat interface integrated into the EP3 workspace that supports the following modes:

1. Context-aware model queries

2. Simulation error interpretation and resolution

3. Automated simulation report generation

The system employs dynamic context management through tool-based information retrieval, allowing the AI to request specific model information as needed rather than processing entire files. This enables users to interact with the AI while actively working on their models, with the AI aware of their current selection and modeling context.

All AI suggestions undergo schema validation before presentation to users. The validation checks for hallucinations, and the existence of any connected objects. Invalid suggestions trigger retry with feedback about validation failures in a closed-loop process. AI suggestions are presented to the user before being integrated into the EP3 model using EP3's EnergyPlus import functions, allowing users to review and accept changes without leaving the modeling interface.

Early results from beta testing indicate that this approach successfully reduces barriers to detailed energy modeling without sacrificing precision or access to EnergyPlus's full capabilities. The implementation demonstrates that AI can provide intelligent assistance for complex building simulation embedded directly within a user interface, enabling seamless integration of AI assistance into existing modeling workflows.

Recent advances in large language models (LLMs) have opened new frontiers for data-driven insights and decision-making across a range of domains. This presentation explores how LLMs can be leveraged to accelerate discovery and enable more intelligent building operations. By integrating LLMs with building management systems, sensor networks, and simulation tools, we demonstrate how these models can assist in uncovering hidden patterns, automating fault detection, optimizing energy usage, and supporting adaptive control strategies. The session will highlight case studies where LLMs interpret unstructured data (e.g., maintenance logs, occupant feedback, or building codes) and bridge the gap between human-centric narratives and machine-readable operational insights. Attendees will gain an understanding of the opportunities and limitations of applying LLMs in the built environment, along with perspectives on future research directions at the intersection of AI and sustainable building operation.

Mastery of the text-based input file formats used by building performance simulation tools like EnergyPlus and Radiance is a major hurdle that separates simulation experts from novices. Large language models (LLMs) are well adapted to text-based inputs and might remove this barrier for novice users. This presentation will explore certain nuances of simulation input languages to understand why, for example, EnergyPlus is relatively easy for LLMs to comprehend, while Radiance presents a bigger challenge. Despite this, we will showcase work in which agentic AI successfully generates Radiance input that would be challenging for expert users to produce, such as definitions of novel materials. We will also discuss the effect that AI in this role will likely have on training the future building performance simulation workforce.