ModPlant-RTN

A resource-task network framework for modeling and scheduling complex recipes in modular plants.

Paper Implementation

This repository contains the reference implementation and the evaluation data for:

B. Chen, T. Miny, and T. Kleinert, "A Resource-Task Network Framework for Modeling and Scheduling Complex Recipes in Modular Plants," in Proc. IECON 2026 — 52nd Annual Conference of the IEEE Industrial Electronics Society, Doha, Qatar, 2026.

The framework normalizes a branched batch recipe into a finite task graph, maps it together with a machine-readable plant description onto a resource-task network (RTN), and decides branch selection, unit assignment, material routing, and timing in a single constraint-programming solve. Every schedule is re-checked by an independent validator and can be exported as an equipment-assigned master recipe in BatchML.

What This Repository Contains

• Recipe normalization and enrichment — parallel regions, exclusive and inclusive alternatives, and bounded loops are unrolled into a finite acyclic task graph. Missing runtime decisions fail closed instead of being silently assumed.
• RTN mapping — units and ports become unary resources, materials and recipe milestones become state resources, and every recipe operation becomes a task with signed coefficients.
• CP-SAT planner — one model decides branch selection, unit assignment, unified-flow routing, staged inventory balance, purity, and connection actions, maximizing a profit objective that prices makespan, operation cost, energy, and emissions.
• Lazy auxiliary transfers — transfer candidates are generated only from detected inventory violations, with complete enumeration as fallback.
• Independent schedule validator — five event-simulation checks (branch consistency, precedence, resource overlap, material balance, objective re-computation) that reuse no planner code.
• BatchML import and export — ISA-88 general recipes in, equipment-assigned master recipes with OPC UA method bindings out.
• AAS plant descriptions — plant capabilities, ports, capacities, and inventories are read from Asset Administration Shell files.

Repository Layout

ModPlant-RTN.ipynb           end-to-end pipeline: recipe -> RTN -> schedule -> validation -> export
ModPlant-RTN-UI.ipynb        interactive notebook front end for the same pipeline
modplant_rtn/                computational kernel shared with the desktop application
  models.py                  plant model, capabilities, ports, inventories
  aas.py                     AAS/AASX import and export
  service.py                 planning service used by the notebooks
  visualization.py           SFC, Gantt, and connection-graph rendering
scripts/                     planner core
  rtn.py                     RTN data model (Pantelides formalism)
  recipe_ir.py               recipe intermediate representation and auto-enrichment
  recipe_to_rtn.py           recipe graph -> RTN mapping
  cp_sat_planner.py          CP-SAT model: variables, constraint families, objective
  schedule_validator.py      solver-independent schedule validation
  pipeline.py                orchestration of the full run
  isa88_recipe.py            BatchML general and master recipe I/O
  notebook_helpers.py        shared notebook kernel
shared/modplant_recipe/      BatchML recipe graph package with normalization and validation
tests/                       91 unit tests and the benchmark notebook
  ModPlant-Benchmarks.ipynb  thirteen recipe-plant combinations reported in the paper
evaluation/                  raw results behind the numbers reported in the paper

Requirements

• Python 3.10 or newer (tested on 3.13)
• Google OR-Tools CP-SAT 9.15 or newer
• See requirements.txt for the full list

python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Run The Notebook

jupyter notebook ModPlant-RTN.ipynb

The notebook loads the four-module demo plant (HC10–HC40), builds the Module_Parallel_Choice_Example recipe, solves it, validates the result, and writes the general and master recipe XML files to generated/. It also renders the selected schedule as a Gantt chart and the resulting inter-module connection configuration as a connection graph. ModPlant-RTN-UI.ipynb offers the same pipeline behind an interactive form.

To reproduce the benchmark table of the paper:

jupyter notebook tests/ModPlant-Benchmarks.ipynb

Expected Runtime Behavior

On an Apple M1 Max (10 cores, 64 GB RAM) with OR-Tools CP-SAT 9.15, a 0.1 % relative gap tolerance, a 120 s wall-time limit, eight workers, and a fixed seed:

Case	Result
Reference case (parallel dosing + fast/standard choice)	proven optimal in 2.6 s, makespan 2283 s
Thirteen benchmark combinations	all proven optimal, median 0.5 s, maximum 18.3 s
Capacity-infeasible order	rejected in 0.01 s with a diagnostic

Solving is deterministic: identical inputs reproduce identical schedules.

Tests

python -m pytest tests/ -q

All 91 tests cover the recipe intermediate representation, the RTN mapping, the CP-SAT model, the validator, runtime control structures (loops, jumps, inclusive choices), and the BatchML round trip.

Scope and Limitations

• The planner operates offline on one batch order for one plant configuration at a time. Multi-batch campaign scheduling is future work.
• Cost, duration, energy, and emission coefficients in the demo data are assumed example values. They can be replaced by measured plant data without changing the model.
• The exported master recipe demonstrates equipment assignment at the recipe level. Closed-loop execution on an orchestration environment is outside the scope of this repository.
• The evaluation is a proof of concept on laboratory-scale configurations of three to five units.

License & Acknowledgments

Released under the MIT License, see LICENSE.

Developed at the Chair of Information and Automation Systems (IAT), RWTH Aachen University. The Honeycomb (HC) modular laboratory plant family used in the evaluation is described here.