ECORE: Diverse Applications in Modern Research

Updated 6 July 2026

ECORE is an umbrella term for distinct constructs in EMF metamodeling, electrochromic optical recording, energy‐conscious edge routing, and cellular network clustering.
In the Eclipse Modeling Framework, Ecore defines metamodel structures—such as EPackages, EClasses, and relationships—that drive instance modeling and automated code generation.
In applied settings, ECORE enables high-sensitivity optical recording of bioelectrical signals, energy-efficient AI inference, and optimized clustering in cellular networks through tailored heuristics and LLM-assisted workflows.

ECORE denotes several unrelated technical constructs in contemporary research. In model-driven engineering, the closely related term Ecore names the metamodeling language of the Eclipse Modeling Framework (EMF). In bioelectric imaging, ECORE expands to electrochromic optical recording, a label-free method for measuring extracellular electrical activity. In edge AI, ECORE names an Energy-Conscious Optimized Routing framework for dispatching inference requests across heterogeneous devices. In cellular networking, eCORE denotes enhanced Clustering Optimization for Resources’ Efficiency, a Device-to-Device clustering heuristic for FDD-LTE systems (Petrovic et al., 7 Mar 2025, Ahmed et al., 30 Jan 2026, Alqahtani et al., 8 Jul 2025, Kollias et al., 2017).

1. Terminological scope

The similarity between ECORE, eCORE, and Ecore creates a recurrent ambiguity across otherwise unrelated literatures. The main usages documented in recent work are summarized below.

Term	Domain	Meaning
Ecore	EMF / model-driven engineering	The metamodeling language of the Eclipse Modeling Framework
ECORE	Bioelectric microscopy	Electrochromic optical recording
ECORE	Edge computing	Energy-Conscious Optimized Routing for Deep Learning Models at the Edge
eCORE	Cellular networking	enhanced Clustering Optimization for Resources’ Efficiency

In the EMF literature, Ecore functions as the structural basis for metamodels, instance models, transformations, and code generation. In the bioelectric literature, ECORE refers to optical measurement of electrical signals using electrochromism in PEDOT:PSS films. In systems and networking, the acronym is reused for optimization frameworks whose objectives are, respectively, energy-aware inference routing and spectral-efficient clustering (Petrovic et al., 7 Mar 2025, Nakasone et al., 2023, Alqahtani et al., 8 Jul 2025, Kollias et al., 2017).

2. Ecore in the Eclipse Modeling Framework

Ecore is the metamodeling language of the Eclipse Modeling Framework. An Ecore metamodel defines EPackages (namespaces), EClasses (domain concepts), EAttributes and EReferences (properties and relations), Inheritance (EClass generalizations), and Data types and multiplicities. In domain-specific metamodeling, Ecore serves as the “grammar” of a modeling language, and automotive MBSE tools such as SysML-on-AUTOSAR rely on hand-crafted Ecore definitions to capture vehicle concepts including sensors, actuators, and control units (Petrovic et al., 7 Mar 2025).

A more formal account models an Ecore model as a typed rooted graph. In one formulation, $\mathrm{EcoreModel} = (P, C, A, R, O),$ where $P$ is the set of packages, $C$ the set of classifiers, $A$ the set of attributes, $R$ the set of references, and $O$ the set of operations. Each EClass is characterized by a name, supertypes, structural features, and operations; each EAttribute and EReference carries bounds and type information; and each EOperation has a return type and parameters (He et al., 5 Dec 2025).

Ecore also underpins EMF instance modeling and XMI serialization. One account views a metamodel $M$ as a tuple $M=(C,P,R)$ , where $C$ is the set of EClasses, $P$ the set of EAttributes, and $P$ 0 the set of EReferences. An EMF instance model then consists of typed objects with unique xmi:id values, attribute assignments, and reference links that respect the metamodel. XMI provides the XML serialization layer, but its deep nesting and namespace requirements make direct generation difficult for LLMs, which has motivated intermediate representations and compiler-based workflows (Pan et al., 28 Mar 2025).

3. LLM-mediated construction of Ecore models

Recent work uses LLMs to automate Ecore metamodel construction from natural-language requirements. An automotive workflow combines Retrieval-Augmented Generation to split a large requirements document into chunks, an LLM service that updates an Ecore snippet based on new or revised requirements, automatic PlantUML generation for visualization, and human-in-the-loop feedback to correct or enhance the partial metamodel. The loop initializes an Ecore template with root class “Vehicle,” iteratively unions LLM-generated deltas into the current metamodel, renders PlantUML, and incorporates expert corrections into either the requirements set or the partial Ecore. A prototype was implemented as a Python + Flask web service with two REST endpoints, POST /updateMetamodel and GET /getCurrentMetamodel, using GPT-4o as the underlying LLM. In three reported iterations, the system processed 19 requirements in 647 tokens in 9.6 s, then 14 requirements in 1102 tokens in 11.5 s, and finally 3 requirements in 1113 tokens in 9.3 s; GPT-4o achieved 100% class recall and approximately 100% attribute coverage across the reported categories, whereas DeepSeek 7B showed 0% subclass generalization on actuators and approximately 70% attribute coverage. The authors conclude that the approach reduces manual effort from minutes to seconds per iteration, while noting token/window limits, hallucination risk, and the need for OCL-based validation (Petrovic et al., 7 Mar 2025).

A second line of work argues that domain modeling should not be attempted with a single prompt. The proposed decomposition separates class and enumeration generation, association and aggregation generation, and inheritance generation, with multi-turn refinement, embedded UML/Ecore knowledge, strict line-based output formats, and rule-based validation before PyEcore assembles the model. Evaluated on 20 systems from the DomSys and NLPSet benchmarks, with GPT-3.5-Turbo and Llama-3-8B over 50 runs per system, the decomposed method improved average GPT-3.5-Turbo class F1 from 0.542 to 0.608, attribute F1 from 0.200 to 0.234, and association/aggregation F1 from 0.099 to 0.191. Two-turn class generation improved F1 by 20% for classes and 15% for attributes, while splitting relationship generation improved F1 by 7% for associations and 248% for inheritances. The system was integrated into the standard Ecore editor through an Eclipse plug-in backed by a Python service using OpenAI’s API and PyEcore (Chen et al., 2024).

LLM-assisted instance-model generation has followed a related two-step strategy. Instead of asking a model to emit XMI directly, the system first generates a format-independent conceptual instance model in JSON, containing instance IDs, types, attributes, compositions, and references, and then compiles that representation into valid XMI through PyEcore. On 24 held-out metamodel-and-specification pairs, direct XMI generation yielded low validity rates—4% for GPT-4o, 8% for o1-preview, 0% for Llama 3.1-8B, and 20% for Llama 3.1-70B—whereas the two-step method achieved 100% validity for all models. Under the two-step pipeline, GPT-4o reached grammatical accuracy 99%, semantic precision 95%, semantic recall 96%, and semantic accuracy 93%; Llama 3.1-70B reached recall 95% but lower precision 84%, indicating stronger coverage but more hallucinated content (Pan et al., 28 Mar 2025).

4. Querying, deriving, and generating artifacts from Ecore

Ecore models are not only constructed; they are also queried, transformed, and compiled into other artifacts. For large automotive software models, one study compares direct prompting against an agentic ReAct-style workflow with general-purpose file-access tools. The root.ecore metamodel used in the experiments contains 384 EClassifiers, approximately 13,572 lines of XML, and about 118,000 tokens under the OpenAI tokenizer. Direct prompting inlines the full model into context, whereas the agentic approach uses eight tools—list_directory, find_file, search_directory, open_file, go_to_line, search_file, scroll_up, and scroll_down—over 50-line windows with up to 100 tool calls per question. On 20 hand-crafted questions, direct prompting accuracy ranged from 45% to 90%, while the agentic method preserved 90% accuracy for GPT-4.1 mini and o4-mini but used only about 657–1083 total tokens versus about 119,000–138,000 for direct prompting, yielding an efficiency of approximately 0.5%–0.9%. The dominant failure modes were agentic recursion errors and incomplete inheritance traversal, which suggests that modeling-aware browsing tools could improve precision (Mazur et al., 16 Jun 2025).

Ecore also serves as the anchor for hybrid code generation. In iEcoreGen, requirements are decomposed into operation specifications, the Ecore model is converted into PlantUML, EMF’s template-based generator emits initial Java code and serializes specifications into docstrings, and an LLM completes and fixes unimplemented methods. The hybrid scheme preserves structural correctness for class declarations, fields, getters, setters, and factory methods while delegating method-body synthesis to the LLM. Across five LLMs and twenty tasks, iEcoreGen improved pass@1 by 5–52 points, with a 29-point average gain over the best LLM-only baseline, and improved pass@3 by 11–36 points, with a 22-point average gain, while remaining comparable on compilation@k. On Deepseek-V3.2-Exp, the full system achieved pass@1 0.75, pass@3 0.89, comp@1 1.00, and comp@3 1.00; ablations showed large degradations when requirement decomposition, code compression, context extraction, or iterative code fixing were removed (He et al., 5 Dec 2025).

Outside LLM-centric workflows, Ecore has been used as an intermediate representation in model-derivation pipelines for software analysis. One framework derives a Java-bytecode metamodel called JBCMM from bytecode, applies EMF model-to-model transformations such as loop augmentation, recursion elimination, and timing injection, and then maps the enriched model to an Ecore-based UPPAAL timed-automata metamodel before emitting UPPAAL XML. JBCMM organizes Project, Class, Method, Instruction, and ControlFlowEdge elements and enforces OCL invariants such as sameMethod, hasReturn, and indexNonNegative. In the LiveGraph case study, the model contained 11,795 Instruction nodes, 11,740 ControlFlowEdges, and 33 detected loops, and the full pipeline required 117 s. For Groove Simulator, with 1,482 classes and 203,030 instructions, total time grew to about 55 minutes, which the authors characterize as near-linear scaling in practice (Yildiz et al., 2017).

5. Electrochromic optical recording

In electrophysiology and optical biophysics, ECORE stands for electrochromic optical recording. The method uses a thin film of the conjugated polymer PEDOT:PSS, whose neutral and oxidized states have different electronic populations in the $P$ 1-conjugated backbone. A local electrical potential $P$ 2 at the film/electrolyte interface drives coupled electron and ion motion, altering the optical absorption spectrum through changes in the complex refractive index $P$ 3. In total internal reflection mode, the relative reflectance change obeys $P$ 4 and the microscope-integrated implementation reports experimental linearity over $P$ 5 mV with $P$ 6 This instrumentation uses a 510 nm superluminescent diode, a Wollaston polarizer, balanced detection, and a 100×, NA = 1.49 oil-immersion objective to deliver and collect a TIR probe beam on a $P$ 7 nm PEDOT film (Ahmed et al., 30 Jan 2026).

The microscopy implementation was designed to simplify older prism-based ECORE systems while retaining high sensitivity and enabling simultaneous brightfield or fluorescence imaging. In cell-free tests, a 1 mV, 1 Hz square wave in PBS yielded mean SNR = 334 over 100 s. The reported detection limit at unity SNR was approximately 3 $P$ 8V, slightly improving on the prior ECORE value of 3.3 $P$ 9V. At the sample, the oblique beam formed a 1/e $C$ 0-elliptical spot of 12.9 $C$ 1m × 9.4 $C$ 2m semi-axes, supporting single-cell and subcellular probing. On human iPSC-derived cardiomyocytes, raw traces showed a sharp extracellular-action-potential spike of about 100 $C$ 3V followed 10–20 ms later by a slower mechanical-contraction signal; digital Bessel filtering preserved electrical spikes while suppressing motion artifacts, and 20 $C$ 4M blebbistatin gradually abolished the mechanical signal over about 15 min while leaving electrical spikes intact. The authors identify temporal response, photothermal limits, sequential rather than simultaneous multi-spot recording, and a noise floor about an order of magnitude above shot noise as current limitations (Ahmed et al., 30 Jan 2026).

A compact ECORE apparatus addressed portability and alignment complexity rather than absolute sensitivity. It uses a Thorlabs CPS650F 657 nm diode laser, an autobalancing differential photodetector, a seven-element optical path, and a small aluminum housing of 23 cm × 20 cm × 11 cm. The laser is focused to an elliptical spot of 31 $C$ 5m × 72 $C$ 6m, and the system achieves a sensitivity of 16.7 $C$ 7V, within a factor of 5 of the shot-noise limit; after filtering out power-line interference, operation within about 1.5 dB of the shot-noise floor is reported. In recordings from HiPSC-derived cardiomyocytes, action-potential SNR was about 38, AP half-width was 0.97 ms, and waveform stability was maintained for more than 10 minutes of continuous recording (Nakasone et al., 2023).

6. Energy-conscious optimized routing for deep learning models at the edge

In edge computing, ECORE denotes a routing framework for distributing image-processing requests across heterogeneous edge device-model pairs. The system comprises edge cameras or sensors, a central gateway, and backend edge nodes. The gateway performs lightweight preprocessing, estimates scene complexity through object count, executes the routing algorithm, and forwards frames to a selected backend. Each profiled pair $C$ 8 has associated energy $C$ 9, detection accuracy $A$ 0, latency $A$ 1, CPU utilization $A$ 2, and memory usage $A$ 3. The generic optimization objective is $A$ 4 subject to latency, CPU, and memory constraints, while the implemented policy uses an accuracy slack $A$ 5: it filters the profiled pairs within an object-count group to those with accuracy at least $A$ 6 and then greedily selects the one with minimum energy. Object-count estimation is performed by one of three strategies: Edge-Detection, SSD-Front-End, or Output-Based (Alqahtani et al., 8 Jul 2025).

The evaluation uses YOLO, SSD, and EfficientDet models on Jetson Orin Nano, Raspberry Pi 4 and 5, and TPU accelerators, over COCO validation, a Balanced-Sorted image set, and a pedestrian video sequence. With $A$ 7, the framework reports that context-aware routing strategies can reduce energy consumption and latency by 45% and 49%, respectively, while incurring only a 2% loss in detection accuracy compared to accuracy-centric methods. On full COCO, Edge-Detection incurred a 2% mAP drop versus the highest-mAP-per-group baseline while using far less energy than accuracy-only schemes; on Balanced-Sorted inputs, Output-Based routing approached oracle behavior with less than 1% mAP drop and the lowest energy among the proposed methods; and on video, Output-Based again delivered the most favorable energy–accuracy–latency trade-off. The stated limitations are static profiling, single-request greedy decisions without global load balancing or Pareto optimization, and lack of support for batch or multi-stage pipelines (Alqahtani et al., 8 Jul 2025).

7. enhanced Clustering Optimization for Resources’ Efficiency in cellular networks

In cellular networking, eCORE is a low-complexity clustering algorithm for FDD-LTE systems with Device-to-Device communication. The underlying problem is to form single-hop clusters, each with one cluster head and several members, so as to minimize the total number of Physical Resource Blocks required in downlink and uplink. User equipments are associated to base stations through binary variables $A$ 8, clustered membership is represented by $A$ 9, and the objective minimizes $R$ 0 under assignment, no-self-loop, and relay-feasibility constraints. The problem is NP-hard. eCORE therefore uses a heuristic based on clustering gain, $R$ 1 where positive $R$ 2 indicates that clustering UE $R$ 3 under cluster head $R$ 4 reduces total PRB consumption (Kollias et al., 2017).

The heuristic first identifies candidate cluster heads with positive gain and then resolves conflicts by sorting users by descending best gain and comparing aggregate savings across competing assignments. Its complexity is $R$ 5 in the number of UEs. In the reported simulations, eCORE achieved downlink spectral efficiency 1.80 bps/Hz versus 1.26 with no clustering and 1.73 for CORE, while the optimum ILP reached 1.95. Total spectral efficiency was within 5% of the ILP optimum, and throughput gain relative to no clustering was +47.2%, compared with +36.6% for CORE. Average cluster statistics for 60 UEs were about 11.6 clusters of average size 2.75. The paper also introduces two complementary algorithms: CaLB, which creates non-spectral-efficient clusters when they result in capacity increase, and CEEa, which manages the trade-off between capacity enhancement and early battery drain of cluster heads. The main limitations noted for eCORE are single-hop clustering, dependence on reliable SINR estimates, and reduced gains under light load or when traffic is not downlink-heavy (Kollias et al., 2017).