ChemE-as-Code: An Open-Source Platform for Agentic AI in Chemical Engineering
Organizations: Process Intelligence Research Group, Department of Chemical Engineering, Delft University of Technology, Van der Maasweg 9, 2629 HZ Delft, The Netherlands
Abstract
ChemE-as-Code provides an open-source environment in which AI agents built on large language models (LLMs) can carry out chemical engineering work and engineers can check it. Its calculation tools range from catalysis through process simulation to plant layout, HAZOP and life cycle assessment, and skills tell a general-purpose agent how to use them. Engineering files are linked, machine-readable and versioned in Git; calculations run outside the LLM, each value records its origin, and automated checks report inconsistencies and out-of-date results by file and line. In an illustrative ammonia case, agents developed a written task into a preliminary basic-engineering package (simulation files, P&IDs, layout, HAZOP, datasheets, relief) for USD 224 of API cost in the design sessions. Planted faults and engineering changes show that the checks find what their rules cover and miss the rest, so they guide human review rather than replace it. No independent engineer has reviewed the design.
Figures & tables
| Term | Meaning in this paper | Software analogy |
|---|---|---|
| Check, finding | A program tests one rule over the files; where the rule fails, it reports a finding with the file, the line, a severity and a hint | Automated test; failing test |
| Skill | A short instruction file that tells an agent when and how to perform an engineering task, which tools to use, and what is left to people | Runbook, contributor guide |
| Tool | A deterministic program that calculates or checks; it never calls an LLM | Library, compiler |
| Provenance | The record of where a value came from: given (with a source), assumed (with a reason) or calculated (by a named tool) | Lock file that records the origin of each dependency |
| Fingerprint; out of date | A hash of a calculated value’s inputs; when the inputs change, the value is out of date until it is recalculated | Build cache, invalidated when a source changes |
| Gate, sign-off | A gate is a set of checks a project stage must pass; a sign-off is the recorded acceptance of a stage by a reviewer | Required status check; merge approval |
| Plant | 1,500 t/d of refrigerated liquid ammonia from natural gas in the Rotterdam area: desulfurization, prereforming, fired primary and air-blown secondary reforming, water-gas shift, CO 2 removal, methanation, drying, 150 bar synthesis loop with refrigeration and purge-gas recovery; electric drives, steam export |
| Input from the author | Concept stage: a task and a literature selection in a process-library issue. Engineering stage: a written brief (“take ammonia-1 to a complete design: simulation, basic engineering, HAZOP, datasheets, relief”), budget caps and stop rules |
| Stages | Concept stage, 27 October 2026: 3 sessions, 40 min. Engineering stage, 29 to 30 October 2026: 60 sessions, about 27 h |
| Agent and roles | Claude Opus 5.5 in every session; the roles implement, review, fix, sign off and merge differ only by instructions, section task and loaded skills |
| Human interventions | Eight in the engineering stage, counting the brief: budget and stop rules, two platform bug fixes outside the design, and the decisions the agents escalated |
| Output | Process: flowsheet and converged DWSIM simulation. Engineering: DEXPI P&ID sheets, equipment, line and instrument lists, control loops, datasheets, routed 3D layout. Safety: HAZOP study, relief cases. Assessment: AACE class 4 estimate, TEA, LCI. Records: decisions and reviews (counts in SI Section S9) |
| Checks at the reported state | 0 errors, 268 warnings; all seven concept and 13 basic gates pass, three of five detail gates; calculation maturity balanced , the second of four levels (SI Section S3) |
| Section | Sessions | Implementation | Review | Total |
|---|---|---|---|---|
| Concept stage: design basis, balanced flowsheet, preliminary LCI | 3 | $8.71 | $3.30 | $12.02 |
| S1 Concept close-out and sign-off | 5 | $6.50 | $2.61 | $9.11 |
| S2 Process simulation (DWSIM) | 3 | $28.03 | $2.41 | $30.44 |
| S3 P&IDs, 14 sheets | 6 | $28.04 | $6.59 | $34.63 |
| S4 Control and lists | 2 | $5.07 | $1.95 | $7.01 |
| S5 Layout, MTO, estimate, TEA, basic sign-off | 26 | $57.26 | $32.23 | $89.50 |
| Indicator | Design | Reference |
|---|---|---|
| Net energy, GJ/t | 26.97 | 27.6–31.8 1.5 (BREF BAT); 27.6 (IEA BAT) |
| Net energy, electricity at 40%, GJ/t | 30.14 | 27.6–31.8 (BREF BAT); 28.05 (IEA, if delivered; our inference) |
| Gross energy, GJ/t | 30.13 | 32.4 (IEA BAT) |
| Natural gas, GJ/t | 28.01 | 29.3–31.1 (BREF reported, feed plus fuel, not BAT); 32.1 (IEA BAT) |
| Process CO 2 , t/t | 1.164 | 1.15–1.40 (BREF) |
| Direct CO 2 , t/t | 1.558 (1.84 with purchased power at 475 g/kWh) | 1.8 (IEA BAT) |
| Experiment | Result | Description |
| Planted faults | ||
| Faults written knowing the rules | 19/19 | intended rule fired; these rules work as designed |
| Faults written without seeing the rules | 4/12 | 5 more raised only a change signal, 3 raised nothing |
| Implausible but self-consistent values | 0/4 identified | e.g. a datasheet pressure off by a factor of ten; only change signals |
| Engineering changes (files reached when first checked) | ||
| Pump re-rated; re-tagged | 4/19; 13/22 | re-tagging reaches the most; spare, cost and electrical not reached on a re-rating |
Agentic AI
Agentic artificial intelligence studies systems that plan, use tools, learn from feedback, and act toward goals with varying degrees of autonomy.
See all 223 papers