Source-linked AI summary
Separating Engineering Reasoning from DEXPI Serialization in LLM-Based Greenfield Surface-Process Design: A Three-Case Study for Underground Gas Storage
Qingchuan Zhu, Shuyue Tong, Pengju Ren
TL;DR
The study examines whether deferring DEXPI serialization changes representation burden and failure modes in constrained underground gas-storage process design. Across three cases, Engineering IR reduced prompt burden and was structurally valid in all cases, while engineering inconsistencies remained possible.
Problem
The study asks how representation burden and failure modes differ when engineering design and standards-level DEXPI serialization are generated jointly rather than separately.
Source: §1 · pp. 1–2Method
The study compares Direct DEXPI generation with a lightweight Engineering IR on three underground-gas-storage cases using one fixed model snapshot and one completed generation per condition.
Source: §3.2 · p. 3, §9 · p. 9Results
Takeaways & Limitations
Within these evaluated method bundles, deferring DEXPI serialization substantially reduced representation burden and helped isolate serialization failures, but did not eliminate engineering inconsistencies.
Source: §7 · p. 8, §9 · p. 9The comparison covers one fixed model snapshot and three cases with one completed generation per condition, so it does not support statistical or broad population-level inference.
Source: §8 · p. 8
Abstract
from arXiv · showhide
Large language models can produce engineering descriptions and structured process representations, but standards-level serialization can substantially increase the generation burden. This diagnostic study examines whether separating engineering reasoning from Data Exchange in the Process Industry (DEXPI) serialization changes where representation and engineering failures occur in constrained greenfield surface-process design for underground gas storage. We compare Direct DEXPI generation with generation of a lightweight Engineering Intermediate Representation (IR) on three cases: single-pressure injection, withdrawal and export, and dual-pressure injection. All six conditions use one fixed model snapshot, qwen3.8-max-0902, with one completed hosted generation per condition. Direct outputs are XSD-valid in 2 of 3 cases, while all 3 Engineering IR outputs are structurally valid under a minimal validator. Direct prompt inputs contain approximately 121.8k-121.9k tokens, compared with 617-699 tokens for the Engineering IR prompts. Engineering feasibility does not uniformly favor the IR: one IR output is rejected for an explicit cooling-state contradiction. The cases also reveal two distinct Direct DEXPI failure modes: engineering inconsistency and standards-level serialization failure. The observed comparison shows that, in these evaluated method bundles, deferring DEXPI serialization substantially reduces representation burden and helps isolate serialization failure, but reducing representation burden alone does not eliminate engineering inconsistencies.
1. Introduction
The study asks how combining engineering design with DEXPI serialization compares with deferring serialization in constrained underground-gas-storage process design. Its diagnostic aim is to distinguish the engineering and serialization failures exposed by the two approaches.
1. Introduction
- Greenfield design must satisfy operation-sequence, material-flow, state, and boundary-condition requirements while also representing the process in a standards-level format.Joint generation therefore combines engineering reasoning with serialization.Source: §1 · p. 1
- The constrained UGS cases cover injection, withdrawal for export, and simultaneous delivery to two destinations with different pressure requirements.Capacity, pressure, temperature, and boundary constraints are supplied; the model chooses major operations and intermediate states.Source: §1 · p. 1
- Direct generation produces a complete DEXPI 2.0 Process XML document, whereas Engineering IR produces a lightweight JSON representation of nodes and streams.The IR condition postpones standards-level serialization.Source: §1 · p. 2
- The study asks how representation burden and observable failure modes differ when DEXPI serialization is performed jointly with design or deferred.Source: §1 · p. 2
- The comparison is diagnostic: it observes how the two representation strategies expose engineering and serialization failure modes.Source: §1 · p. 2
2. Background and Related Work
Related work applies language models and agents to process simulation, flowsheet correction, and generation, often using structured representations. Prior studies also use intermediate representations and DEXPI-targeted workflows; this paper does not claim the intermediate-representation idea as novel.
2. Background and Related Work
- Chemical-process studies use LLMs or agents for simulation, flowsheet autocorrection, and automated generation, including structured representations that can be parsed and checked.The cited examples include Liang et al. 2026, Balhorn et al. 2024, and Bahamdan et al. 2026.Source: §2 · p. 2
- Prior work uses intermediate representations in DEXPI-targeted P&ID generation and machine-readable flowsheet graphs, so neither this use of IRs nor DEXPI-targeted workflows is claimed as new.Gowaikar et al. 2024 describe translation from an intermediate domain-specific language to DEXPI/Proteus XML; Laub et al. 2026 use SFILES with standardized formats including DEXPI Process.Source: §2 · p. 2
- Valid Direct DEXPI could support exchange with compatible tools and downstream automation, but standards-level representation requires encoding objects and relationships within the information model and XML conventions.Source: §2 · p. 2
- The study compares Direct DEXPI with lightweight Engineering IR under the same model snapshot and cases, recording representation validity, engineering feasibility, and method-specific completion.It asks how failure modes differ when serialization is joint or deferred.Source: §2 · pp. 2–3
3. Study Design
The study compares two method bundles on shared design bases, using one fixed hosted-generation protocol and preserving the six generated outputs and their validation records as raw evidence.
3.1 Generation methods
- Direct prompts request complete DEXPI 2.0 XML with official schema and model references, while Engineering IR prompts request a five-field JSON object with node and stream data.The method bundles differ in reference context, schema size, and output contract, so the comparison does not isolate XML versus JSON alone.Source: §3.1 · p. 3
3.2 Generation settings and protocol
- Each case-method condition has one completed hosted generation, with no result-dependent retry, repair, or regeneration.The formal protocol uses one fixed model snapshot; the seed is a reproducibility control, not a guarantee of bitwise deterministic inference.Source: §3.2 · p. 3
3.3 Evidence preservation
- The reproducibility package preserves all six generated XML or IR JSON files, run metadata, and validation records without normalizing or rewriting the artifacts.Source: §3.3 · p. 3
4. Test Cases
The three test cases cover single-pressure injection, withdrawal and export, and dual-pressure injection. Across them, the model selects operations and intermediate states subject to the supplied design basis.
4. Test Cases
- The cases are single-pressure injection, withdrawal and export, and dual-pressure injection to two groups with different pressure requirements.The model independently determines major operations, their sequence, and intermediate states needed to satisfy the boundary conditions.Source: §4 · p. 4
- One passage gives 8.0–10.0 MPa and approximately 2 °C; another reports export at 6.5–7.5 MPa in Case 3, dual-pressure injection.Source: §4 · p. 4
- The dual-pressure case specifies 900,000 standard m³/day total, split as 300,000 to Group A and 600,000 to Group B.Source: §4 · p. 4
- No additional gas-quality treatment is specified for the withdrawal-and-export case.Source: §4 · p. 4
5. Evaluation
Evaluation separates representation validity from engineering feasibility, then combines them in a method-specific outcome; engineering review is qualitative rather than rigorous process simulation.
5. Evaluation
- Direct representation passes only when XML parses and validates against the frozen XSD; IR passes when JSON satisfies its minimal structural validator.Neither gate alone establishes full DEXPI semantic conformance or engineering validity.Source: §5 · p. 4
- Engineering review checks topology, flow and state continuity, operation/state consistency, and boundary conditions, but not rigorous simulation or detailed equipment design.It omits full equation-of-state and fluid-property calculations.Source: §5 · p. 4
- Engineering Feasibility labels are assigned in a post-hoc qualitative review of frozen outputs, not as a prespecified statistical endpoint.The review applies the stated topology, continuity, consistency, and boundary criteria without modifying raw outputs.Source: §5 · p. 4
- Within-Method Outcome combines method-specific output completion with engineering acceptability, and no deterministic IR-to-DEXPI conversion is evaluated.A structurally valid IR can still fail when its engineering states are inconsistent.Source: §5 · p. 5
6. Results
Across three underground-gas-storage cases, Direct DEXPI and Engineering IR differ in both representation and engineering outcomes. IR reduced prompt burden, but engineering feasibility did not uniformly favor it.
6.1 Case 1: single-pressure injection
- Case 1 Direct DEXPI was XSD-valid but rejected for engineering errors, while Engineering IR was structurally valid and warned over an unexplained 34.8-to-35.0 MPa step.The Direct output included a feed self-loop, missing flow values, and an aftercooler temperature increase.Source: §6.1 · p. 5
6.3 Case 3: dual-pressure injection
- Case 3 Direct DEXPI failed XSD validation despite an engineering warning, while Engineering IR passed with both dual-pressure branches represented without an obvious operation/state contradiction.The IR final states were 20 MPa / 40 °C for Group A and 30 MPa / 40 °C for Group B.Source: §6.3 · p. 7
6.4 Unified result table
- Direct DEXPI was XSD-valid in two of three cases, while all three Engineering IR outputs were structurally valid; engineering outcomes did not consistently favor either method.The unified outcomes were Case 1: Direct FAIL / IR WARN; Case 2: Direct WARN / IR FAIL; Case 3: Direct FAIL / IR PASS.Source: §6.4 · p. 7
6.5 Token and latency comparison
- Deferring DEXPI serialization reduced prompt burden across all cases, with Direct-to-IR prompt-token ratios of approximately 197×, 185×, and 174× for Cases 1–3.Latency is reported descriptively, not as a strict performance conclusion, because server load, caching, and backend scheduling can affect it.Source: §6.5 · p. 8
7. Discussion
The two method bundles expose different failure modes: structural validity and engineering feasibility are distinct, while Engineering IR makes process details easier to inspect and greatly reduces prompt burden.
7. Discussion
- Direct DEXPI passes XSD validation in Case 1 but fails engineering review, while Case 3 has coherent topology but fails XSD validation and omits the required standard-volume flow basis.Structural validity does not establish engineering consistency, and engineering coherence does not ensure standards-level validity.Source: §7 · p. 8
- The methods expose different levels of process detail, but their differing compression stages cannot be attributed to representation alone because the methods are bundles.Case 1 uses three Direct compressor objects and two interstage coolers versus two IR compression stages; in Case 3, Direct encapsulates stages that IR expands as nodes.Source: §7 · p. 8
- Engineering IR is structurally valid in all three cases, but engineering outcomes range from WARN to REJECT to ACCEPT; Case 2 includes cooling that raises temperature from 12 °C to 35 °C.Its compact node-and-stream representation makes the Case 2 inconsistency straightforward to localize.Source: §7 · p. 8
- Direct prompts contain approximately 121.8k–121.9k tokens versus 617–699 tokens for Engineering IR, or approximately 174×–197× more in the Direct condition.These evaluated method bundles differ in reference context and serialization responsibility; deferring DEXPI serialization reduces prompt burden and helps distinguish serialization from engineering failures.Source: §7 · p. 8
8. Limitations
The comparison is limited to three cases and one generation per condition under one model snapshot, and its method bundles do not isolate representation format. Engineering review and structural validation are also limited proxies for broader engineering and standards conformance.
8. Limitations
- The study covers one fixed model snapshot, three test cases, and one completed generation per case-method condition, so results support run-specific observations rather than statistical or broad population claims.No result-dependent retry, repair, or regeneration was applied.Source: §8 · p. 8
- Direct DEXPI and Engineering IR differ in prompt context, output contract, validator, and serialization responsibility, so observed differences cannot be attributed to XML versus JSON alone.Direct includes the large frozen DEXPI reference context, unlike the smaller IR schema.Source: §8 · p. 8
- The post-hoc qualitative engineering review is not a full simulation or detailed equipment design, and neither IR structural validity nor Direct XSD validity establishes full semantic conformance.The review omits full equation-of-state and fluid-property calculations; no deterministic IR-to-DEXPI converter is evaluated.Source: §8 · p. 9
9. Conclusion
Across three cases, Direct DEXPI was XSD-valid in two and all Engineering IR outputs were structurally valid, yet engineering feasibility varied. The comparison indicates that deferring DEXPI serialization reduces representation burden but does not itself eliminate engineering inconsistencies.
9. Conclusion
- Direct DEXPI was XSD-valid in two cases, while all three Engineering IR outputs were structurally valid in this three-case diagnostic comparison.The Direct outputs showed both engineering inconsistency and standards-level serialization failure.Source: §9 · p. 9
- Engineering IR reduced prompt burden from approximately 121.8k–121.9k tokens to 617–699 tokens, but one output still contained an explicit cooling-state contradiction.The representation also made process objects and state transitions more compact to inspect.Source: §9 · p. 9
- In these evaluated method bundles, separating engineering semantics from DEXPI serialization substantially reduces representation burden and helps isolate serialization failures, but does not eliminate engineering inconsistencies.The observations concern completed runs and do not support statistical or population-level inference.Source: §9 · p. 9
Data and Code Availability
The reproducibility package preserves the study setup, generated outputs, and validation records; reruns require API calls and are not expected to reproduce identical output bytes.
Data and Code Availability
- The public repository preserves the formal runner, prompts, case requirements, model configuration, validators, frozen reference artifacts, engineering review, six raw outputs, validation records, and SHA-256 manifest.Outputs are preserved without repair, normalization, or result-dependent retry.Source: Data and Code Availability · p. 9
- The frozen DEXPI 2.0.0 reference artifacts are redistributed under CC BY 4.0, with source attribution and third-party licensing recorded in THIRD_PARTY_NOTICES.md.Source: Data and Code Availability · p. 9
- A new hosted run consumes API calls and is not expected to reproduce identical output bytes; no real API key is included in the repository.REPRODUCIBILITY.md documents how to inspect the frozen setup and start a new hosted run.Source: Data and Code Availability · p. 9