Onyx: A Multifaceted Research Term

• Onyx is a polysemous term that denotes distinct artifacts in SEM graphical modeling, additive manufacturing, HPC, radiosurgery, enterprise AI, and secure ANN systems.
• In SEM, Onyx streamlines model specification by enabling visual diagram creation and automated code export, significantly reducing manual scripting for twin modeling.
• Across engineering domains, Onyx demonstrates notable performance improvements—from high fracture toughness in polymers to enhanced throughput and predictable behavior in HPC and ORAM-secured ANN systems.

Searching arXiv for papers mentioning “Onyx” across the relevant technical domains. Onyx is a polysemous technical term whose meaning depends entirely on disciplinary context. In recent research literature, it denotes a graphical front-end for structural equation modeling interoperable with umx and OpenMx, a chopped-carbon-fiber-reinforced nylon used in fused deposition modeling, a liquid-cooled Cray supercomputer at the ERDC DoD Supercomputing Resource Center, a liquid embolic agent considered in Gamma Knife arteriovenous malformation dosimetry, the organizational context behind an enterprise RAG benchmark, and a disk-oblivious approximate nearest neighbor search system for trusted execution environments (Castro-de-Araujo et al., 11 Dec 2025, Cuesta et al., 2019, Williams et al., 2023, Dagli et al., 2021, Sun et al., 5 May 2026, Rathee et al., 22 Apr 2026). A recurring source of ambiguity is therefore nominal rather than conceptual: identical nomenclature is applied to unrelated artifacts spanning SEM software, materials science, medical devices, high-performance computing, enterprise AI evaluation, and privacy-preserving vector search.

1. Referential scope and disciplinary separation

A concise way to disambiguate the major research uses of the term is to map each referent to its operational role.

Domain	Onyx referent	Role
SEM	$\Omega$nyx/Onyx	GUI tool for SEM exporting OpenMx RAM and algebra code
Additive manufacturing	Onyx	Nylon-based crystalline polymer reinforced with short carbon fibers
HPC	Onyx	Cray XC40/50 production supercomputer
Radiosurgery	Onyx	Liquid embolic agent in AVM dosimetry studies
Enterprise RAG	Onyx	Organization releasing EnterpriseRAG-Bench
Secure ANN	Onyx	Cost-efficient disk-oblivious ANN system

The most common misconception is that these uses share a technical lineage. They do not. The SEM tool $\Omega$nyx is part of a graphical modeling workflow; the polymer Onyx is an FDM feedstock; the HPC system Onyx is infrastructure; the embolic Onyx is a clinical material; the enterprise-RAG use refers to a research and product organization; and the ANN system Onyx is a systems-security design. The shared label is therefore best treated as homonymy across research communities rather than as a unified technical family.

2. $\Omega$nyx/Onyx in structural equation modeling

Within the umx v4.5 ecosystem, $\Omega$nyx is a GUI tool for SEM in which models can be specified visually by drawing elements of a diagram, after which the resulting diagram can be exported as OpenMx RAM and algebra code (Castro-de-Araujo et al., 11 Dec 2025). In this role, Onyx functions as a graphical front-end for specifying path models, twin models, and related biometrical structures by drawing boxes, circles, and arrows rather than writing code directly. Its principal interoperability value is that the exported mxPath() syntax can be read by umxRAM() for general SEM or by umxTwinMaker() for twin models.

The workflow is explicit. A researcher draws a model in $\Omega$nyx, exports it as OpenMx RAM code, comments out the data-related lines and top-level mxModel() construction that $\Omega$nyx generates, retains the mxPath() statements as an R object, and then supplies those paths to umxRAM("modelName", data = myData, my_paths) or umxTwinMaker("modelName", paths = my_paths, mzData = mzData, dzData = dzData). For twin models, umxTwinMaker() interprets the $\Omega$nyx paths as a base single-twin model and expands them into a multiple-group ACE specification with MZ and DZ submodels. The paper states that the final OpenMx object is a multiple groups model, with each submodel including the MZ and DZ specific data, and that this massively reduces scripting code for ACE models.

The integration is strongest for Cholesky AE/ACE-style twin models and ICU models for censored data. In the Cholesky AE example, latent factors such as a1, a2, e1, and e2, growth factors such as icept and slope, and manifests x1, x2, x3 are drawn visually and exported as RAM paths. In the ICU example, a latent X1 loads on X1cont and X1bin, with the latent variance subsequently decomposed into A and E components in the twin extension. In both cases, $\Omega$nyx supplies the latent structure and path layout, while umx supplies the twin grouping, constraints, data incorporation, and summary workflow.

Automatic twin expansion depends on naming conventions. The naming of the A, C, and E variances must follow the pattern a1, a2, a3, ..., c1, c2, c3, ..., and e1, e2, e3, ... so that umxTwinMaker() can appropriately set the MZ and DZ covariance paths and constraints. The implied ACE covariance structure is the standard one: for MZ twins, $$\operatorname{Cov}(A_{T1},A_{T2}) = 1 \cdot \operatorname{Var}(A)$$, $$\operatorname{Cov}(C_{T1},C_{T2}) = 1 \cdot \operatorname{Var}(C)$$, and $\Omega$0; for DZ twins, the corresponding genetic covariance is $\Omega$1.

Its limitations are equally clear. The paper does not claim that all advanced umx model families are directly authored in $\Omega$2nyx. CLPM, MR-DoC, and sex-limitation models are presented primarily through high-level umx functions such as umxCLPM, umxMRDoC, and umxSexLim(). A plausible implication is that $\Omega$3nyx is best understood here as a model-specification and rapid-prototyping layer, rather than as the full authoring environment for all of umx v4.5.

3. Onyx as a chopped-carbon-fiber-reinforced FDM polymer

In additive manufacturing, Onyx is a nylon-based crystalline polymer reinforced with short carbon fibers and printed by FDM through layer-by-layer extrusion (Cuesta et al., 2019). In the fracture study summarized here, it serves as the fiber-reinforced benchmark against which PLA, PP, and ABS are compared. Its fiber morphology, from SEM, is reported as a mean fiber length of approximately $\Omega$4 and a fiber diameter of approximately $\Omega$5. Manufacturer data cited in the summary give a density of $\Omega$6, Young's modulus of $\Omega$7, elongation at failure of $\Omega$8, and flexural strength of $\Omega$9; it is described as approximately $\Omega$0 stronger and stiffer than ABS and approximately $\Omega$1 stiffer than regular nylon due to fiber stiffening. CT quantification gives a mean porosity of approximately $\Omega$2.

The fracture characterization uses the Essential Work of Fracture framework, which partitions total work into essential work associated with surface creation and non-essential plastic work in the surrounding dissipation zone:

$\Omega$3

with

$\Omega$4

Here, $\Omega$5 is the specific essential work of fracture and $\Omega$6 is the effective plastic work parameter. For deeply double-edge notched tensile specimens, the Onyx results are reported as $\Omega$7, $\Omega$8, and $\Omega$9. The study states that Onyx’s fracture energy is about an order of magnitude higher than ABS and significantly higher than PLA and PP.

The distinction between initiation toughness and global plastic dissipation is central. Onyx exhibits much higher $\Omega$0 than the non-reinforced FDM polymers, but its $\Omega$1 is not correspondingly dominant; PP, not Onyx, dominates the plastic term. This indicates that the main reinforcement benefit is in fracture initiation energy rather than in large-scale plastic dissipation. The reported comparisons make this explicit: Onyx/ABS in $\Omega$2 is approximately $\Omega$3, Onyx/PLA is approximately $\Omega$4, and Onyx/PP is approximately $\Omega$5.

The material also shows a specific ductility profile. Using the ductility level

$\Omega$6

the Onyx DDEN-T specimens are placed in the $\Omega$7-$\Omega$8 range, corresponding to the ductile instability regime, on the border to post-yielding. ABS and PLA occupy the same regime, whereas PP lies in the blunting regime with substantially larger deformation. Thus, Onyx combines high fracture initiation resistance with moderate ductility rather than extreme blunting behavior.

A second contribution of the study is methodological. The proposed deeply double-edge notched small punch configuration reproduces the essential work of fracture for Onyx with close agreement to DDEN-T. The reported values are $\Omega$9 for DDEN-T and $\Omega$0 for DDEN-SP, a difference of approximately $\Omega$1, while $\Omega$2 rises from $\Omega$3 to $\Omega$4. The summary therefore concludes that DDEN-SP is reliable for quantitative characterization of $\Omega$5 when only small specimens are available, but not for direct quantitative comparison of the plastic work term.

4. Onyx as an HPC platform

In high-performance computing, Onyx is a large Cray XC40/50 supercomputer at the ERDC DSRC and serves in the cited benchmarking study as the production baseline against which the research-oriented Vulcanite cluster is compared (Williams et al., 2023). The benchmark uses only the GPU-accelerated nodes of Onyx. Across the machine, the system is reported to have 4,810 total nodes and 211,640 total CPU cores; the GPU nodes used in the benchmark employ Intel E5-2699v4 Broadwell processors at 2.8 GHz, two sockets per node, useable memory per node of 247 GB for the GPU configuration row, NVIDIA P100 PCIe accelerators with 16 GB GPU memory each, and a Cray Aries interconnect. The system is described as liquid-cooled.

The workload is a material segmentation model originally implemented in Caffe and converted to PyTorch using Microsoft Research’s MMdnn toolkit. The network is GoogLeNet, and the dataset is MINC-2500, a 57,500-image subset consisting of 2,500 samples per category across 23 material classes, with each image a $\Omega$6 patch. The benchmark measures inference rather than training: images are scaled to three resolutions, cropped into windows, passed through the model to produce 23 probability maps, and timed at the per-image and per-directory level. Dense CRF post-processing is explicitly excluded from the benchmark because the study’s initial focus is on GPU performance.

Performance characterization emphasizes stability rather than absolute peak speed. Onyx was benchmarked across 5 runs on GPU nodes. The summary states that Onyx’s model times are consistent across benchmarks and that its average model times cluster tightly across runs. The per-image inference times are on the order of milliseconds; the study remarks that, in most applications, the difference between $\Omega$7 and $\Omega$8 is negligible. Vulcanite’s V100-based nodes are faster for the mean, median, and 5th percentile, but the average 95th percentile model times for Vulcanite start to become slower than the model times on Onyx at the same percentile.

The interpretation offered is infrastructural. Vulcanite is air-cooled and appears more susceptible to environmental factors and possible thermal throttling, whereas Onyx is liquid-cooled and therefore less susceptible to ambient temperature variation. The result is a familiar HPC trade-off: Onyx provides lower raw accelerator performance than newer V100-based hardware, but better tail behavior and greater run-to-run predictability for this inference workload.

5. Onyx as a liquid embolic agent in radiosurgical dosimetry

In the radiosurgical literature summarized here, Onyx is a commercially available, non-adhesive liquid embolic agent used in the management of cerebral arteriovenous malformations, and it is considered alongside cyanoacrylate and polyvinyl alcohol in the context of Gamma Knife dose attenuation (Dagli et al., 2021). The study abstract states that dose accumulations by these three embolization materials were analysed using a Monte-Carlo simulation implemented in Geant4, with an 8 mm collimator helmet size, and that statistically significant differences in the calculated dosimetries were observed.

The supplied technical summary characterizes Onyx as an ethylene-vinyl alcohol copolymer system using dimethyl sulfoxide as solvent and tantalum powder for radiopacity. In that description, tantalum is the clinically relevant feature for dosimetry because it introduces a high-$\Omega$9 component absent from the lower-$\Omega$0 comparison materials. This suggests that the central physical question is not simply whether embolization alters the target geometry, but whether the embolic composition itself perturbs the attenuation of the $\Omega$1 beam.

The relevant dosimetric concern is straightforward. Gamma Knife radiosurgery relies on multiple focused cobalt-60 photon beams converging on an isocenter. A radiodense embolic cast can attenuate or perturb the beam before it reaches parts of the nidus. The cited abstract confirms only that statistically significant material-dependent differences were found; it does not provide the numerical magnitude of those differences in the accessible summary. Even so, the larger literature summary supplied with the entry treats Onyx as the embolic among the three with the most consequential attenuation potential because of its tantalum content, whereas cyanoacrylate and PVA are described as closer to tissue-equivalent in MeV photon fields.

A common oversimplification is to treat embolic materials as dosimetrically interchangeable. The study’s framing rejects that assumption: Onyx, cyanoacrylate, and PVA are not merely alternative occlusion materials but also different radiation transport media. The precise clinical significance depends on geometry, thickness, and planning context, but the category distinction itself is explicit.

6. Onyx in enterprise retrieval-augmented generation research

In enterprise AI, Onyx is the organization releasing EnterpriseRAG-Bench, a benchmark for retrieval-augmented generation over company-internal knowledge (Sun et al., 5 May 2026). The dataset consists of approximately 500,000 documents; the detailed count given in the appendix is 511,962 documents spanning nine source types: Slack, Gmail, Linear, Google Drive, HubSpot, Fireflies, GitHub, Jira, and Confluence. The corpus simulates a single AI software company, Redwood Inference, and the benchmark includes 500 questions across 10 categories: Basic, Semantic, Intra-Doc Reasoning, Project Related, Constrained, Conflicting Info, Completeness, Miscellaneous, High Level, and Info Not Found.

Its central methodological contribution is to model enterprise knowledge as cross-document coherent, heterogeneous, and noisy rather than as a clean web corpus. The generation framework constructs top-level scaffolding artifacts, initiatives, employee directories, source hierarchies, and agents.md files; decomposes initiatives into projects designed to yield around 100 documents each; and mixes a high-fidelity project-centric generation path with a cheaper topic-scaffolded high-volume path. Realistic noise is then injected by randomly relocating about 5% of documents within a source type, moving an additional 3% through an LLM-based structurally plausible but wrong relocation procedure, creating near-duplicates with conflicting facts, and adding miscellaneous off-topic directories.

The evaluation protocol separates answer quality from retrieval quality. Correctness is binary and LLM-judged; completeness is fact-level and defined as

$\Omega$2

document recall is measured as Recall@10 against required gold documents; and invalid extra documents are counted as an absolute number rather than a ratio. A correction-aware gold-set pipeline then revises required versus valid versus invalid document labels by pooling gold and candidate documents and adjudicating them with three independent LLM judges.

The baseline results are notable because they invert common expectations derived from public web benchmarks. Using the same generator and judge model, BM25 with OpenSearch achieved 68.8% correctness, 56.0% completeness, 68.4% Document Recall@10, and 9.0 invalid extra documents per question on average. Vector search using text-embedding-3-large with a Qdrant cosine index achieved 51.4% correctness, 42.9% completeness, 46.0% recall, and 9.3 invalid extras. A Bash agent using shell tools achieved 60.6% correctness, 61.1% completeness, 55.8% recall, and only 2.0 invalid extras. BM25 therefore wins on correctness and recall, whereas the agent yields the best completeness and the lowest retrieval noise.

This result is important because it shows that enterprise retrieval behavior differs materially from web-oriented benchmark intuition. Internal jargon, structured SaaS artifacts, and densely overlapping topics can favor lexical signals over public-web embedding spaces. The benchmark’s cluster analysis reinforces that interpretation: local top-10 KNN cosine similarity is reported as 0.83 for both EnterpriseRAG-Bench and Onyx’s own internal corpus, versus 0.69 for BrowseComp-Plus. A plausible implication is that enterprise corpora behave like “harder” retrieval environments per document because semantically similar distractors are abundant.

7. Onyx as a disk-oblivious ANN system

In systems and security research, Onyx is a cost-efficient disk-oblivious approximate nearest neighbor search system for TEEs that co-designs a new ANN engine, Onyx-ANNS, with a new ORAM layer, Onyx-ORAM (Rathee et al., 22 Apr 2026). The motivating problem is that cost-efficient ANN deployments in confidential computing environments cannot keep large vector indices entirely in enclave memory and must therefore use external SSDs, whose access patterns are visible to the host. Encrypting stored vectors is insufficient because the physical sequence of reads and writes can leak query information and dataset structure.

The system’s key design claim is an inversion of the usual ANN/ORAM division of labor. Prior designs such as Compass minimize access count in the ANN layer and bandwidth in the ORAM layer. Onyx instead minimizes bandwidth in the ANN layer and access count in the ORAM layer. The argument is architectural: ANN search can exploit approximation to avoid unnecessary vector transfers, whereas ORAM is better positioned to restructure tree depth and locality so as to reduce I/O operations.

Onyx-ANNS modifies DiskANN by decoupling traversal from refinement. Traversal blocks store neighbor lists together with compact pruning hints, while full-precision vectors are moved into a separate refinement ORAM. Search proceeds in three phases: graph traversal using small in-memory hints, pruning of the visited candidate set using more precise disk-resident pruning hints, and refinement of only the top $\Omega$3 candidates using full vectors. The resulting logical bandwidth changes from

$\Omega$4

to

$\Omega$5

The reported effect is a $\Omega$6-$\Omega$7 reduction in ANN-side bandwidth with only a 5%–15% increase in access count at 90% top-10 recall, depending on dataset.

Onyx-ORAM extends RingORAM through three linked changes: local bucket metadata inside the TEE, full-bucket reads during evictions, and a locality-aware shallow $\Omega$8-ary tree. The headline comparison given in the paper is that RingORAM has approximately $\Omega$9 access count and $\Omega$0 bandwidth overhead, whereas Onyx-ORAM has approximately $\Omega$1 access count and $\Omega$2 bandwidth. The chosen sweet spot is $\Omega$3, which reduces depth by a factor of about 3 while keeping bandwidth close to that of bandwidth-efficient binary RingORAM.

The security objective is disk-access privacy: logical access patterns to each ORAM instance are independent of queries and index content, up to the public operation type and system parameters. The paper states a theorem that Onyx is disk-access private when instantiated with a secure authenticated encryption scheme. Protected information includes query vectors, dataset vectors, encrypted index contents, and logical disk access patterns; out-of-scope channels include CPU microarchitectural side channels, page-fault leakage inside the TEE, rollback, and denial of service.

The end-to-end performance results position Onyx as a practical rather than merely theoretical private vector-search design. Compared with Compass-in-TEE at 90% top-10 recall, Onyx achieves $\Omega$4-$\Omega$5 lower latency and $\Omega$6-$\Omega$7 higher throughput on the evaluated datasets. Compared with the best baseline for a given dataset, it is $\Omega$8-$\Omega$9 faster. In cost terms, it yields $\Omega$0-$\Omega$1 higher queries per dollar than Compass-in-TEE. A concrete deployment point reported in the paper is that a 64 GB WIKI index can run on a 1-vCPU, 4-GB, 1-SSD resource unit at 70 QPS, 12 ms latency, and 90% top-10 recall, serving more than 8 million queries per dollar.

Taken together, these uses show that Onyx is not a single object of knowledge but a recurrent label attached to technically distinct artifacts. In current research discourse it can denote a SEM GUI, an engineered composite filament, an HPC installation, a radiopaque embolic, an enterprise-AI benchmark context, or a confidential-computing ANN system. Precision therefore requires domain qualification: “Onyx” alone is underspecified, whereas “$\Omega$2nyx,” “Onyx filament,” “Onyx supercomputer,” “Onyx embolic,” “Onyx EnterpriseRAG-Bench,” and “Onyx disk-oblivious ANN” are materially different referents with different methodological and theoretical significance.