UniOpt

An Experimental Agentic AI Framework Employing Large Language Models for Automated Chemical Process Optimization

UniOpt is a multi-agent LLM pipeline that takes a converged steady-state UniSim Design simulation (XML export + PFD screenshot) and a natural-language optimization request, then automatically:

1. Interprets the process (textual + visual)
2. Formulates a continuous-variable mathematical optimization problem
3. Generates executable Python code that drives the original .usc file via the COM interface

---

Framework Overview

┌─────────────────┐     ┌─────────────────┐
│  XML Interpreter│     │  PFD Interpreter│
│  (DeepSeek-V4-  │     │  (Qwen3.5-397B) │
│   Flash)        │     │                 │
└────────┬────────┘     └────────┬────────┘
         │                       │
         └───────────┬───────────┘
                     ▼
          ┌─────────────────────┐
          │ Optimization        │
          │ Formulator Agent    │
          │ (DeepSeek-V4-Pro,   │
          │  max reasoning)     │
          └──────────┬──────────┘
                     ▼
          ┌─────────────────────┐
          │ Coder Agent         │
          │ (DeepSeek-V4-Pro)   │
          └──────────┬──────────┘
                     ▼
              Python + COM
              

Agents

Agent	Model	Role
XML Interpreter	DeepSeek-V4-Flash	Distills UniSim XML → process report (equipment, streams, topology)
PFD Interpreter	Qwen3.5-397B-A17B	Vision-based cross-check of equipment ordering and connectivity
Optimization Formulator	DeepSeek-V4-Pro (max)	Builds objective, decision variables, constraints, solver settings
Coder	DeepSeek-V4-Pro (max)	Generates and debugs Python/COM optimization code

All instruction prompts are static system prompts.

---

Case Studies

Three steady-state examples are provided:

1. Heat-exchangers in series
   Minimize total energy duty while enforcing T= 80° C.
   The cooler duty is correctly driven to zero.

2. Three-stage nitrogen compression
   Intermediate pressures compared with the classical equal-pressure-ratio rule (Edgar et al., 2001).
   Gap < 1 %.

3. Ethyl chloride manufacturing (with recycle)
   Maximize venture profit by adjusting purge rate.
   Result matches Seider et al. (2017) within 0.11 %.

   Requirements

• Python 3.10+
• Windows (UniSim Design COM interface)
• Honeywell UniSim Design (licensed)
• OpenAI-compatible API access to:
  • DeepSeek-V4-Flash
  • DeepSeek-V4-Pro
  • Qwen3.5-397B-A17B (via SiliconFlow or equivalent)