QSDsan Engine MCP

A universal wastewater treatment simulation engine exposing QSDsan capabilities through dual adapters for AI agent integration.

Motivation

Commercial wastewater simulation platforms offer sophisticated biological models but impose a significant bottleneck: GUI-driven workflows that limit iteration speed, parallelization, and reproducibility. Process engineers spend substantial time navigating interfaces rather than exploring designs.

QSDsan Engine MCP inverts this paradigm by making natural language the primary interface. Instead of clicking through dialogs, engineers describe what they want:

"Build an MLE process with 4000 m3/d influent, simulate for 15 days, and explain why ammonia removal is low"

This enables:

• Collapsed iteration cycles: Build -> run -> diagnose -> patch -> rerun without GUI navigation
• Massive scenario enumeration: DOE, Monte Carlo, and optimization workflows become natural since everything is code
• Reproducible, diffable runs: Version-controlled session specs with deterministic metadata
• Structured diagnostics: Validation warnings, model compatibility checks, and actionable error messages surfaced directly to agents

The goal is not to replace domain expertise, but to remove friction so engineers can focus on design decisions rather than tool mechanics.

Architecture: Dual Adapters

The engine exposes identical functionality through two adapters:

                    +-------------------------------------+
                    |       QSDsan Engine Core            |
                    |  (Templates, Models, Converters)    |
                    +-----------------+-------------------+
                                      |
              +-----------------------+---------------------+
              |                       |                     |
              v                       v                     v
     +----------------+      +----------------+     +----------------+
     |   MCP Adapter  |      |   CLI Adapter  |     |  Python API    |
     |   (server.py)  |      |   (cli.py)     |     |  (direct use)  |
     +----------------+      +----------------+     +----------------+
              |                       |
              v                       v
     +----------------+      +----------------+
     |  MCP Clients   |      |  Agent Skills  |
     |  (Claude, etc) |      |  (Claude Code) |
     +----------------+      +----------------+

MCP Adapter (server.py)

For MCP-compatible clients (Claude Desktop, Cline, etc.):

python server.py

CLI Adapter (cli.py)

For CLI-based agent runtimes and Agent Skills:

python cli.py --help

Tool Surface

Core Simulation Tools

Flowsheet Construction Tools

Build custom treatment trains dynamically:

Session Management Tools

Modify flowsheets without starting over:

Discoverability Tools

Explore models and validate configurations before simulation:

Artifact Retrieval Tools

Access simulation outputs programmatically:

Supported Models

Pre-built Templates

Advanced Simulation Features

SRT-Controlled Simulation (Phase 12)

For systems with MBR or clarifier that decouple HRT and SRT, the engine supports target SRT setpoints where sludge wasting is automatically controlled to achieve the desired SRT:

# CLI: Run MLE-MBR with target SRT of 15 days
python cli.py simulate \
  --template mle_mbr_asm2d \
  --influent influent.json \
  --target-srt 15 \
  --srt-tolerance 0.1

How it works:

• Uses scipy.brentq root-finding to find optimal Q_was (waste activated sludge flow)
• Enforces minimum simulation time of 2× target SRT for dynamics equilibration
• Validates mass balance: Q_was ≤ Q_in (since Q_in = Q_was + Q_effluent)
• Achieves target SRT within specified tolerance (e.g., 0.1 = 10%)

Test results: Target SRT 5.0 days → Achieved SRT 5.01 days (0.23% error)

Run-to-Convergence Simulation

For accurate steady-state simulation, use convergence-based stopping:

# CLI: Run until steady state (auto-detected)
python cli.py flowsheet simulate \
  --session my_plant \
  --run-to-convergence \
  --convergence-atol 0.1 \
  --max-duration 100

Features:

• Abs+rel tolerance: |slope| < atol + rtol × max(|mean|, floor)
• Multi-stream tracking: effluent (nutrients) + WAS (biomass)
• Auto-detection of effluent and sludge streams
• Oscillation detection for non-converged systems

Quick Start

Using CLI

# List templates
python cli.py templates --json-out

# Create influent file
cat > influent.json << 'EOF'
{
  "flow_m3_d": 4000,
  "temperature_K": 293.15,
  "concentrations": {"S_F": 75, "S_A": 20, "S_NH4": 35, "S_PO4": 9}
}
EOF

# Run MLE-MBR simulation (use file path for --influent, --duration-days not --duration)
python cli.py simulate \
  --template mle_mbr_asm2d \
  --influent influent.json \
  --duration-days 15 \
  --report

# Build custom flowsheet
python cli.py flowsheet new --model ASM2d --id my_plant
python cli.py flowsheet add-stream --session my_plant --id influent \
  --flow 4000 --concentrations '{"S_F": 75, "S_A": 20, "S_NH4": 35}'
python cli.py flowsheet add-unit --session my_plant --type CSTR --id anoxic \
  --params '{"V_max": 1000}' --inputs '["influent"]'
python cli.py flowsheet add-unit --session my_plant --type CSTR --id aerobic \
  --params '{"V_max": 2000, "aeration": 2.0}' --inputs '["anoxic-0"]'
python cli.py flowsheet build --session my_plant
python cli.py flowsheet simulate --session my_plant --duration 15

Using MCP

Configure in your MCP client (e.g., Claude Desktop config.json):

{
  "mcpServers": {
    "qsdsan-engine": {
      "command": "python",
      "args": ["/path/to/qsdsan-engine-mcp/server.py"]
    }
  }
}

Then use natural language:

"Create an MLE process treating 4000 m3/d of municipal wastewater and simulate for 15 days"

Unit Registry

49 unit operations available across categories:

• Reactors: CSTR, AnaerobicCSTR, PFR, ActivatedSludgeProcess, AnaerobicDigestion
• Separators: CompletelyMixedMBR, AnMBR, PolishingFilter, MembraneDistillation
• Clarifiers: FlatBottomCircularClarifier, PrimaryClarifier, IdealClarifier
• Sludge: Thickener, Centrifuge, SludgeDigester, DryingBed
• Junctions: ASM2dtoADM1, ADM1toASM2d, mADM1toASM2d (model converters)
• Utilities: Splitter, Mixer, Tank, StorageTank, DynamicInfluent

# List all units
python cli.py flowsheet units --json-out

# Filter by model compatibility
python cli.py flowsheet units --model mADM1

# Filter by category
python cli.py flowsheet units --category reactor

Pipe Notation

Connect units using BioSTEAM pipe notation:

# Output notation: "A1-0" -> unit A1, output port 0
# Input notation: "1-M1" -> unit M1, input port 1
# Direct: "U1-U2" -> U1.outs[0] -> U2.ins[0]
# Explicit: "U1-0-1-U2" -> U1.outs[0] -> U2.ins[1]

Output

Simulations produce:

• JSON results with effluent quality, removal efficiencies, and deterministic metadata (solver settings, library versions, timestamps)
• SVG flowsheet diagrams showing unit operations and streams
• Quarto reports (optional) with comprehensive analysis
• Time-series data for tracked streams

Installation

# Clone repository
git clone https://github.com/puran-water/qsdsan-engine-mcp.git
cd qsdsan-engine-mcp

# Create virtual environment
python -m venv venv
source venv/bin/activate  # or venv\Scripts\activate on Windows

# Install dependencies (either method works)
pip install -r requirements.txt
# OR
pip install -e .

Dependencies

Python packages (installed automatically):

• Python 3.10+
• QSDsan 1.3+
• BioSTEAM 2.40+
• FastMCP (for MCP adapter)
• Typer + Rich (for CLI adapter)
• Jinja2 (for report generation)
• Matplotlib (for time-series plots)

External tools (install separately):

• Graphviz: Required for flowsheet diagrams
• Linux: sudo apt install graphviz
• macOS: brew install graphviz
• Windows: https://graphviz.org/download/
• Quarto CLI (optional): For rendering .qmd reports to HTML/PDF
• https://quarto.org/docs/get-started/

License

University of Illinois/NCSA Open Source License - see LICENSE.txt for details.

This is a derivative work based on QSDsan, licensed under the same terms.

Acknowledgments

Built on QSDsan by the Quantitative Sustainable Design Group.