FacProcessTwin: LLM Process Twin System

• FacProcessTwin is an LLM-based system that constructs complete digital twins of industrial processes by reading narrative documentation and integrating live OPC UA data.
• It employs a deterministic pipeline with human-in-loop checkpoints to resolve safety-critical ambiguities, ensuring accuracy in process modeling and binding.
• Empirical evaluation shows a sixfold speed-up in twin development with high topology F1 scores (>95%) and zero mis-binding for ambiguous tags.

Searching arXiv for the specified paper and closely related digital-twin work. {"^^^^^^^^3query3 OR all:\3^^^^^^^^"FacProcessTwin\"","max_results":5} {"^^^^^^^^3query3 An LLM-Based System for Process Twin Development","max_results":^^^^3ti:\3query3^^^^} FacProcessTwin is an end-to-end, LLM–based system for developing a complete, data-bound process twin directly from the narrative documentation plants already maintain, with human-in-the-loop governance to keep safety-critical bindings correct. In contrast to asset-based digital twins that model a single machine or sensor in isolation, a process twin is a real-time, data-connected digital representation of an entire production process: each step, the equipment it uses, and the product-specific settings across variations. FacProcessTwin reads free-form process documents, recovers the full production process, binds that model to live operational data through OPC UA tags, and renders the result as an interactive process diagram through which operators can monitor and correct autonomous decisions. In a real-world case study at an Australian food manufacturer covering ^^^^3ti:\3^^^^6 production process flows, it achieved a mean topology F^^^^3ti:\3^^^^ of 95.^^^^3 OR all:\3^^^^%, built each twin in roughly a sixth of the manual time, and deferred all genuinely ambiguous safety-critical bindings to the operator, mis-binding none (&&&^^^^3query3^^^^&&&).

^^^^3ti:\3^^^^. Conceptual scope and motivation

FacProcessTwin addresses the cost structure of process twin development as described for manufacturing plants in which the dominant burden is recurring manual effort. Manual development requires reconstructing the entire process from disparate sources, including formal procedures and tacit operator knowledge, and then binding that reconstructed model to live data streams such as PLC/SCADA tags, OPC UA servers, historians, and MES. These tasks recur with every process change, which makes recurring manual effort the dominant cost. FacProcessTwin targets that development bottleneck rather than the isolated modeling of one asset (&&&^^^^3query3^^^^&&&).

The distinction between a process twin and an asset-based digital twin is central. A process twin captures how process steps interact end-to-end and therefore supports interventions that improve throughput, yield, quality, and safety across the whole process rather than at one asset. In the food-manufacturing example used in the paper, the relevant scope extends from raw-material inspection through preparation, cooking or pasteurisation, to filling and packaging, while preserving the safety-critical control points defined in the plant’s HACCP plan.

A second motivation is safety-critical data binding. The paper emphasizes that an agent operating without oversight can silently guess the wrong tag at precisely the step where a wrong binding would be hazardous, with the holding-tube temperature at UHT sterilisation serving as the canonical example. FacProcessTwin therefore combines automation with staged human checkpoints. This suggests that the system is designed not merely as an extraction pipeline, but as a governance structure for bounded autonomy in industrial process modeling.

^^^^3 OR all:\3^^^^. Architecture and end-to-end workflow

FacProcessTwin runs as a single pipeline with deterministic tools under an LLM agent that is advisory-only. The architecture comprises ingestion and reasoning, process model generation, live data integration, an interactive process diagram and user experience, and governance. Every change is executed by deterministic tools; the LLM proposes but never bypasses tools (&&&^^^^3query3^^^^&&&).

The workflow begins with document ingestion. The system reads plant process documentation in PDF or DOCX form, extracting both text and tables. From these inputs, the LLM recovers candidate product flows and interacts with the operator in natural language to select scope and resolve ambiguity. Once a target flow is chosen, the system generates nodes and directed edges for the process graph and computes the layout deterministically to avoid overlaps or mis-links. The resulting model is persisted in a flexible entity–attribute–value store.

The live-data stage then connects to OPC UA servers and discovers tags. The paper describes three fallbacks for practical interoperability: vendor filters, direct reads when listing returns empty, and identifier repair after server restarts. Tag binding follows a staged policy. The LLM proposes candidate bindings based on names, metadata, and known associations, but deterministic tools only bind when the match is unambiguous. Ambiguous or safety-critical bindings are deferred to the operator rather than guessed.

The data flow is described in five steps: documentation is ingested and candidate flows are proposed; the operator selects the target flow and approves the generated layout; the system connects to OPC UA servers and discovers tags; the LLM proposes bindings and the system either binds unambiguously or defers; confirmed bindings are subscribed and streamed onto the diagram, and values are written to the historian.

3. Process representation, extraction, and binding logic

The internal representation is a directed process graph serialized in an EAV store. Its entities include process steps such as washing, cutting, cooking, UHT sterilisation, and aseptic filling; equipment such as tanks, pumps, and cooker units; sensors and parameters such as temperatures, pressures, and fill levels; and control-point labels such as CCP, OCP, and QCP. Attributes include product-specific settings, variant flags such as smooth versus particulate, and HACCP-derived control-point designations. Relationships encode directed product flow, explicit equipment–step association, and bindings from nodes to live tags (&&&^^^^3query3^^^^&&&).

Process variations are represented either as parameter sets or as alternative subpaths. The paper gives cooker unit alternatives and smooth versus particulate flows as examples of this mechanism. In practical terms, this permits flows to share the same step library while differing in instantiated steps and settings.

Model generation is grounded in a documented HACCP step code library of ^^^^3 OR all:\3^^^^6 steps. The LLM is prompted to enumerate product flows, resolve synonyms to the SOP step library, and output an ordered step sequence with control-point labels. Deterministic tooling then enforces topology correctness, canonicalization to the SOP vocabulary, and retention of CCP, OCP, and QCP safety labeling.

Binding logic is explicitly conservative. Ambiguity is detected when competing candidates map to the same parameter or step, or when an SOP does not specify physical unit selection, as in the case of two cooker units. CCP designation strengthens deferral: CCP steps must not be auto-bound if multiple candidates exist. Confirmed bindings are subscribed, streamed into the diagram, and written to the historian; any overwrite of existing data requires explicit operator confirmation, and a final check compares the twin’s bound signals against recorded values before go-live.

4. Human-in-the-loop governance and operator interaction

Governance is not an auxiliary feature of FacProcessTwin; it is the mechanism by which the system constrains autonomous behavior. The paper defines three staged human checkpoints: choosing the product flow, approving the drawn layout, and clarifying ambiguous bindings. These checkpoints ensure safety and correctness at the points where documentation incompleteness, lexical ambiguity, or unit-level uncertainty are most consequential (&&&^^^^3query3^^^^&&&).

The interactive process diagram serves as the main operational surface. The generated graph is rendered with live values streaming from bound tags. Operators can monitor status in real time, view control-point labels per step, approve or reject the layout, request redraws via deterministic layout logic, inspect and correct bindings, resolve ambiguity via targeted questions, and confirm or veto overwrites. The interface also exposes binding health and the historian record of values.

The paper’s correction workflow is highly specific. When uncertainty remains, the agent issues targeted questions such as “Which of T^^^^3ti:\3^^^^ or T^^^^3 OR all:\3^^^^ is the holding-tube temperature?” The operator’s answer locks the binding, and subsequent overwrites require confirmation. Binding provenance is retained, historian writes create an auditable timeline, and operator feedback is persisted with binding metadata so that ambiguity rules can be re-applied on future runs while reducing repeated deferrals.

A common misconception would be to treat FacProcessTwin as a fully autonomous industrial agent. The architecture described in the paper does not support that interpretation. The LLM is advisory-only, deterministic tools control all side effects, and safety-critical uncertainty is resolved by deferral rather than by automatic completion. This suggests that the system is best understood as a governed twin-construction pipeline rather than as open-ended agentic automation.

5. Empirical evaluation

The evaluation is based on a real-world site: a mid-sized Australian food manufacturer with HACCP SOPs. The study covers ^^^^3ti:\3^^^^6 production process flows spanning chilled, frozen, and aseptic shelf-stable categories, and includes process variations within the same product. The SOP step library comprises ^^^^3 OR all:\3^^^^6 codes instantiated ^^^^3 OR all:\3 OR all:\3^^^^5 times across flows, with ^^^^3ti:\3ti:\3^^^^–^^^^3ti:\3 instances per flow and a mean of ^^^^3ti:\3^^^^4. Live data comes from OPC UA servers for a UHT/pasteuriser with ^^^^3ti:\3^^^^5 tags and a cooker with 4 tags across two units; aseptic flows bind both, while the others bind cookers only. Ground truth consists of explicit, ordered step sequences with control-point annotations and hand-labeled tag-to-node mappings (&&&^^^^3query3^^^^&&&).

The reported metrics use the standard definitions PRESERVED_PLACEHOLDER_^^^^3query3^^^^, PRESERVED_PLACEHOLDER_^^^^3ti:\3^^^^, and PRESERVED_PLACEHOLDER_^^^^3 OR all:\3^^^^.

Metric	Result	Note
Mean topology F^^^^3ti:\3^^^^	95.^^^^3 OR all:\3^^^^%	Range 85.7–^^^^3ti:\3query3query3^^^^%
Pooled topology F^^^^3ti:\3^^^^	95.3%	Precision 95.4%, recall 95.^^^^3query3^^^^%
Sequence accuracy	9^^^^3query3^^^^.^^^^3 OR all:\3^^^^%	Exact order on 8/^^^^3ti:\3^^^^6 flows
Control-point accuracy	96.4%	CCP/OCP/QCP labels
Mapping recall	^^^^3ti:\3query3query3^^^^%	All 9^^^^3 OR all:\3^^^^ required tags correctly bound
Mapping precision	74.^^^^3 OR all:\3^^^^%	3^^^^3 OR all:\3^^^^ extra bindings to a second cooker unit
Ambiguous-tag mis-binding	^^^^3query3^^^^%	FacProcessTwin deferred all ^^^^3 OR all:\3query3^^^^ ambiguous tags
Baseline ambiguous-tag error	75.^^^^3query3^^^^%	^^^^3ti:\3^^^^5/^^^^3 OR all:\3query3^^^^ silently mis-bound
Time per flow	5.^^^^3 OR all:\3^^^^ min	Range 3.4–7.5
Manual time per flow	3^^^^3ti:\3^^^^.8 min	Range ^^^^3 OR all:\3^^^^5.^^^^3 OR all:\3^^^^–4^^^^3query3^^^^.

The time reduction is reported as 83.6%, corresponding to $$\left(1 - \frac{5.2}{31.8}\right)\times 100\% \approx 83.6\%$$, or roughly a sixfold speed-up. The safety result is equally central: at ^^^^3 OR all:\3query3^^^^ genuinely ambiguous tags, the single-pass baseline silently mis-bound ^^^^3ti:\3^^^^5, including the holding-tube temperature on aseptic flows when choosing between two temperature tags, whereas FacProcessTwin deferred all ^^^^3 OR all:\3query3^^^^ and mis-bound none. Unambiguous-tag accuracy remained essentially unchanged with governance, at 9^^^^3query3^^^^.^^^^3query3 versus 89.4%, indicating no accuracy trade-off.

Qualitative findings clarify the numerical results. In UHT sterilisation, two temperature tags, T^^^^3ti:\3^^^^ and T^^^^3 OR all:\3^^^^, were exposed by the server. Because the holding-tube measurement is a CCP, FacProcessTwin refused to guess and asked the operator which tag was correct; the answer locked the binding and prevented the baseline’s frequent mis-binding. By contrast, unit-level ambiguity in the cooker servers produced extra but valid bindings to both units because the SOPs did not specify which physical unit a given flow used. The paper identifies ERP/MES enrichment as the natural mechanism for resolving that ambiguity. Robustness tests across GPT-5-mini, DeepSeek-V4-Flash, and MiniMax-M^^^^3 OR all:\3^^^^.7 yielded topology F^^^^3ti:\3^^^^, sequence accuracy, and time-to-twin within about ±^^^^3ti:\3^^^^ percentage point of the default Gemini 3.^^^^3ti:\3^^^^ Flash-Lite.

6. Limitations, generalization, and relation to adjacent digital-twin research

The main limitations stated for FacProcessTwin are documentation quality, tag heterogeneity, domain adaptation, and scaling. Accuracy depends on SOP completeness: missing unit selection or undocumented settings lead to conservative extra bindings or operator deferrals. Noisy or vendor-specific namespaces require robust discovery; the reported fallbacks mitigate but do not eliminate all edge cases. Evaluation is confined to one plant’s SOPs, so generality across sites and industries remains to be validated. Failure modes include dropped steps under the fixed 4^^^^3query3^^^^-step budget and mislabels of control points, with safeguards provided by staged approvals, CCP deferrals, overwrite confirmations, and verification against recorded state (&&&^^^^3query3^^^^&&&).

The paper’s future-work agenda is correspondingly pragmatic. It proposes multi-plant scaling across diverse SOPs and sites, ERP/MES integration to resolve unit selection and undocumented settings, learning from operator feedback to reduce future deferrals, cross-process optimization for scheduling and quality analytics, formal verification against constraints such as physical ranges and process interlocks, and extended integrations including AAS-based information models.

Within the broader digital-twin literature represented in the supplied corpus, FacProcessTwin occupies a distinct layer of the stack. "Fast Online Digital Twinning on FPGA for Mission Critical Applications" (Xu et al., 13 Dec 2025) addresses online model recovery with FPGA-accelerated GRU and sparse dense components for low-latency synchronization in mission-critical environments. "Physics-Informed Neural Network Digital Twin for Dynamic Tray-Wise Modeling of Distillation Columns under Transient Operating Conditions" (Patra et al., 25 Mar 2026) focuses on a physics-informed unit-operation twin in which VLE, MESH, and McCabe–Thiele constraints are embedded in the loss. "Twinac: A Universal Framework for Virtual Accelerator Controls" (Miceli et al., 28 Jul 2025) proposes a facility-agnostic architecture for real-time mirroring, predictive maintenance, uncertainty reduction, and safe deployment of control actions. FacProcessTwin differs from each of these by centering the development problem itself: it constructs the process graph and live bindings from plant documentation and operator natural-language input, and places safety-critical ambiguity resolution inside the twin-building workflow rather than only in runtime supervision.

That positioning has methodological implications. A plausible implication is that FacProcessTwin can be viewed as an upstream orchestration layer for process-twin authoring and grounding, while systems such as FPGA-accelerated online twins, physics-informed unit-operation twins, and facility-agnostic operational frameworks address downstream execution, estimation, or control once a digital representation already exists.