Obelix: A Cross-Domain Research Label

• Obelix is a polysemous research label defined by its capitalization and contextual usage, covering areas such as materials informatics, information retrieval, detector instrumentation, plasma kinetics, computer security, and hadron spectroscopy.
• In materials informatics, OBELiX denotes a curated dataset of 599 lithium solid-state electrolytes that serves as both a data resource and a benchmark exposing current limitations of crystal-based ML models.
• In detector instrumentation and information retrieval, Obelix has driven advances in re-ranking scientific search results at CERN and in developing radiation-tolerant pixel sensors for the Belle II vertex-detector upgrade.

In contemporary research literature, Obelix, OBELIX, and OBELiX denote several unrelated scientific and technical entities rather than a single object. The name has been used for a personalized re-ranking system for scientific search, a curated dataset for lithium solid-state electrolytes, a depleted monolithic active pixel sensor for the Belle II vertex-detector upgrade, a collisional-radiative electron-kinetics model for HiPIMS discharges, a compiler-based obfuscation framework for trusted execution environments, and an antiproton-annihilation experiment cited in hybrid-meson phenomenology (Carlsen, 2015, Therrien et al., 20 Feb 2025, Schwickardi et al., 2023, Onuki et al., 1 Jul 2026, Rudolph et al., 2021, Wichelmann et al., 23 Sep 2025, Meyer et al., 2015). The shared label is therefore best understood as a cross-domain naming coincidence whose meaning is determined entirely by disciplinary context.

1. Taxonomy of usages

The research uses of the name span information retrieval, materials informatics, detector instrumentation, plasma modeling, computer security, and hadron spectroscopy.

Form	Domain	Denotation
Obelix	Information retrieval	Recommendation and personalization system for re-ranking CDS search results
OBELiX	Materials informatics	Curated dataset of lithium solid-state electrolytes and room-temperature ionic conductivities
OBELIX	Belle II instrumentation	DMAPS chip for the VTX upgrade
OBELIX	Plasma physics	Orsay Boltzmann equation for ELectrons coupled with Ionization and eXcited states kinetics
Obelix	Computer security	Dynamic obfuscation framework for TEEs
Obelix / OBELIX	Hadron spectroscopy	Experiment cited for $\bar{p}p \rightarrow 4\pi$ evidence related to $\pi_1(1400)$

A notable source of ambiguity is capitalization. In the materials paper, OBELiX stands for Open solid Battery Electrolytes with Li: an eXperimental dataset and refers to an open benchmark dataset (Therrien et al., 20 Feb 2025). In the Belle II papers, OBELIX names a silicon sensor platform derived from TJ-Monopix2 and adapted to Belle II trigger and readout constraints (Schwickardi et al., 2023, Onuki et al., 1 Jul 2026). In the plasma paper, OBELIX is an acronym for Orsay Boltzmann equation for ELectrons coupled with Ionization and eXcited states kinetics (Rudolph et al., 2021). In the security paper, Obelix is a software tool rather than a physical device or dataset (Wichelmann et al., 23 Sep 2025).

This distribution suggests that the term has no stable transdisciplinary referent. A plausible implication is that precise citation and capitalization are unusually important whenever the name appears in scholarly writing.

2. OBELiX in materials informatics

In materials science, OBELiX is a domain-expert-curated database of 599 synthesized solid electrolyte materials with experimentally measured room-temperature ionic conductivity $(\sigma_{\text{RT}})$ for lithium solid-state electrolytes (Therrien et al., 20 Feb 2025). Each entry records, as available, chemical composition, space group, lattice parameters $(a,b,c,\alpha,\beta,\gamma)$, room-temperature ionic conductivity, total formula-unit count $Z$, bulk ionic conductivity when reported, and full crystal structure in CIF format for a subset. The dataset contains 321 entries with CIF crystal structures, represents 55 space groups, and includes 48 repeated composition/space-group sets, covering 122 entries.

The curation workflow is literature-driven. OBELiX was built primarily from the Liverpool Ionics Dataset and the Laskowski dataset; the authors selected materials for which they could obtain experimental room-temperature ionic conductivity, space group, and lattice parameters, manually retrieved missing values from original papers when necessary, removed exact duplicates, checked near-duplicates, and added 15 entries from other sources (Therrien et al., 20 Feb 2025). For structural data, they obtained 234 exact matches from ICSD, 27 manually retrieved crystal structures, and 60 additional close matches from ICSD or Materials Project based on the same space group, composition within $\pm 0.05$, and lattice parameters within $\pm 3\%$. Because ICSD is not openly redistributable, they openly provide 292 CIF files with small random perturbations to fractional coordinates,

$\epsilon \sim N(0,0.01),$

while preserving crystal symmetry.

A defining design feature is inclusion of partial occupancy, disorder, and experimentally realistic fractional compositions such as $K_{0.1}Li_{0.9}SbO_3$ (Therrien et al., 20 Feb 2025). The paper emphasizes that these features are often crucial for ion transport, but they also create a modeling mismatch because many geometric ML models cannot directly handle partial occupancy; occupancies therefore had to be rounded to the nearest integers before those models were applied. This makes OBELiX simultaneously a data resource and a benchmark exposing current methodological limits in crystal-based ML for disordered ionic conductors.

The dataset also formalizes leakage-aware evaluation. The train/test split avoids placing entries from the same paper or with the same composition across both partitions. A Monte Carlo method was used to optimize similarity of the log ionic conductivity distribution and of the CIF-bearing subset distributions, subject to a test-set size between 20% and 30%. The resulting split places 20.2% of the full dataset and 20.9% of the CIF subset in the test set (Therrien et al., 20 Feb 2025). The repeated-material analysis reports a root mean squared deviation from set averages of 0.63 log(S/cm) and a mean absolute deviation from set medians of 0.41 log(S/cm), which the paper uses as an estimate of experimental variability.

OBELiX further serves as a benchmark for seven existing ML models: Random Forest (RF), Multilayer Perceptron (MLP), PaiNN, SchNet, M3GNet, SO3Net, and CGCNN, together with pretrained variants p-PaiNN, p-SchNet, p-M3GNet, and p-CGCNN (Therrien et al., 20 Feb 2025). Evaluation uses 5-fold cross-validation, grid search, and mean absolute error (MAE). The reported test MAEs show RF at 1.54 (1.82) and MLP at 1.64 (2.17), outperforming the geometric models, whose test MAEs range from 2.52 to 3.04 for the pretrained variants and from 2.74 to 2.89 for the non-pretrained 3D models. The paper attributes this in part to the larger effective training set available to RF and MLP and in part to the difficulty geometric models have with widespread partial occupancy. In that sense, OBELiX functions not only as a dataset but as a benchmark structured to reveal when ostensibly richer structural models underperform simpler composition-and-symmetry baselines.

3. Obelix in personalized scientific search

In information retrieval, Obelix is a recommendation and personalization system developed at CERN to re-rank search results in the CERN Document Server (CDS) built on Invenio (Carlsen, 2015). It is not a standalone search engine. Rather, it operates as a second-stage component on top of an existing IR stack such as Solr, ElasticSearch, or Invenio’s own search system, and it re-ranks only the results already returned by the base search engine. It therefore changes ordering but does not add or remove documents from the candidate set.

Its data model is a graph of users and items, with nodes representing users and items and edges representing item use by a user (Carlsen, 2015). The recommendation logic traverses this graph outward from the current user, so that items used by users closer to the current user are considered more relevant. The prototype scoring rule is described in terms of $n$, the number of users who have used the item, and $\pi_1(1400)$0, the sum of distances from the current user to all users who have used that item, combining a popularity component and a personal-proximity component. The paper further describes the traversal algorithm as a breadth-first search (BFS) that initializes a queue with the starting user, maintains distances $\pi_1(1400)$1, and returns a distance array $\pi_1(1400)$2 and traversal tree $\pi_1(1400)$3.

The system is based on implicit feedback rather than explicit ratings. Its primary signals are item views and item downloads, while click position in search results and the relation between a search result and a later item usage are logged mainly for evaluation and tuning (Carlsen, 2015). The underlying CDS logs span about 10 years, with more than 500 million log entries, more than 1.4 million items, and more than 40,000 unique users. The main analysis uses only logged-in users, because anonymous users cannot be reliably identified across sessions.

Obelix was engineered as a standalone service with a REST API, but the production architecture adopted Redis-based queueing and caching to avoid REST bottlenecks at scale (Carlsen, 2015). The architecture includes a graph database, event queue, recommendation queue, key/value store, REST API, and an integration layer with the IR system. The paper argues for Neo4j as the default graph database and uses Redis for caching and queueing in CDS. This asynchronous design separates event ingestion from recommendation computation: user events are inserted into the graph database through one queue, and updated recommendations are computed and cached through another.

In CDS, the base search engine first returns a ranked set, typically latest first, after which Obelix retrieves user-specific recommendations from Redis, intersects them with the original results, boosts the matching items, and returns a reranked list (Carlsen, 2015). The online experiment used an importance factor with 0 meaning no use of recommendations, 1 meaning full override of the original ranking, and intermediate values mixing the base ranking and Obelix’s score. During the experiment, the factor varied among

$\pi_1(1400)$4

every ten minutes.

Offline evaluation tested four parameter dimensions: time frame $\pi_1(1400)$5, number of usages included $\pi_1(1400)$6, graph traversal depth $\pi_1(1400)$7, and item usage weighting $\pi_1(1400)$8 (Carlsen, 2015). The experiment covered about 6 million combinations, around 90 million predictions, each prediction taking about 15 ms, for a total runtime of about 14 days. The main findings were that about one year of history was surprisingly effective, that the system could predict about 45% of user interactions with one year’s worth of data, and that best graph depth was around 2 or 3, with performance flattening after depth 5. Online evaluation over about two months reported average click positions of 12.12 for latest first, 9.34 for word similarity, and 5.64 for Obelix in global search. The paper is explicit, however, that Obelix was not uniformly best across all collections: latest-first was preferred for CERN Yellow Reports, and word similarity worked best for books. The system is therefore presented as a collection-sensitive personalization layer whose clearest benefit appears in heterogeneous global search.

4. OBELIX as a Belle II vertex-detector sensor

In Belle II instrumentation, OBELIX is the depleted monolithic active pixel sensor under development for the vertex detector upgrade, or VTX, intended to replace the present mixed DEPFET pixel plus silicon-strip detector with a fully pixelated system (Schwickardi et al., 2023, Onuki et al., 1 Jul 2026). The motivation is the harsher background environment of SuperKEKB and the altered interaction-region geometry associated with the upgraded final-focus system. The detector must operate under conditions including hit rates up to 120 MHz/cm$\pi_1(1400)$9 in the innermost layer, non-ionizing fluence of about $(\sigma_{\text{RT}})$0 per year, and stringent material-budget constraints.

The 2023 Belle II paper presents OBELIX as the Belle II-specific successor to TJ-Monopix2: the pixel matrix design is essentially carried over from TJ-Monopix2, while the periphery is adapted for Belle II readout and trigger requirements (Schwickardi et al., 2023). The 2026 project paper describes the sensor more explicitly as a chip derived from TJ-Monopix2 but with a largely redesigned digital periphery for Belle II compatibility (Onuki et al., 1 Jul 2026). OBELIX uses a pixel matrix of

$(\sigma_{\text{RT}})$1

with a 33 $(\sigma_{\text{RT}})$2 pixel pitch, and it is designed to withstand

$(\sigma_{\text{RT}})$3

and a total ionizing dose of

$(\sigma_{\text{RT}})$4

The trigger and readout adaptations are central. OBELIX incorporates a trigger unit, a dual-stage memory buffer, support for the external Belle II trigger, hit-rate capability up to

$(\sigma_{\text{RT}})$5

trigger latency of

$(\sigma_{\text{RT}})$6

and trigger rate of

$(\sigma_{\text{RT}})$7

(Onuki et al., 1 Jul 2026). It also includes a 7-bit Time-over-Threshold (ToT), time stamping of 50–100 ns, optional fine time stamping of about 5 ns for hit rates below $(\sigma_{\text{RT}})$8, optional trigger output with about 10 ns time resolution, and a 3-bit per-pixel tuning circuit to compensate threshold dispersion. Two Belle II-specific low-latency features, TTT (Track Trigger Transmission) and PTD (Periphery Time-to-Digital), are foreseen for the outer VTX only.

The detector geometry is cylindrical and consists of five layers of OBELIX sensors around the beam pipe, spanning radii from 14 mm to 140 mm (Onuki et al., 1 Jul 2026). The layer radii are

$(\sigma_{\text{RT}})$9

with L1 at $(a,b,c,\alpha,\beta,\gamma)$0 mm and 6 ladders, L2 at $(a,b,c,\alpha,\beta,\gamma)$1 mm and 10 ladders, L3a/b at $(a,b,c,\alpha,\beta,\gamma)$2 mm and 36 ladders, L4a/b at $(a,b,c,\alpha,\beta,\gamma)$3 mm and 48 ladders, and L5a/b at $(a,b,c,\alpha,\beta,\gamma)$4 mm and 48 ladders. The inner vertex detector (iVTX) uses self-supported all-silicon ladders formed from four consecutive OBELIX sensors interconnected using a redistributed layer (RDL) process, whereas the outer vertex detector (oVTX) uses discrete sensors mounted on omega-shaped carbon-fiber support structures with readout flex circuits attached above the chips.

Material minimization is a first-order design constraint. For the iVTX, the target per-layer material budget is

$(a,b,c,\alpha,\beta,\gamma)$5

with assemblies thinned to

$(a,b,c,\alpha,\beta,\gamma)$6

and passive cooling via a thin thermal pyrolytic graphite (TPG) layer with thermal conductivity around

$(a,b,c,\alpha,\beta,\gamma)$7

(Onuki et al., 1 Jul 2026). For the oVTX, the per-layer target is

$(a,b,c,\alpha,\beta,\gamma)$8

and the readout uses lpGBT and VTRx+ optical links. A major engineering focus is low-mass flex technology using aluminum-conductor flexible circuits, which the paper states can reduce material budget by as much as

$(a,b,c,\alpha,\beta,\gamma)$9

in the oVTX context.

The experimental basis for the OBELIX concept is the predecessor TJ-Monopix2. That device, fabricated in TowerJazz 180 nm CMOS, has a $Z$0 matrix with 33.04 $Z$1m pitch and was characterized using laboratory charge-injection tests and DESY beam tests (Schwickardi et al., 2023). Reported results include an example threshold of 283 $Z$2, threshold dispersion of 17 $Z$3 standard deviation, noise of 8 $Z$4, cluster charge MPV

$Z$5

independently estimated deposited charge

$Z$6

depletion depth

$Z$7

hit efficiency

$Z$8

for a threshold of about 500 $Z$9 at 3 V bias, and cluster position resolution of about

$\pm 0.05$0

The Belle II papers use these results as validation that the OBELIX sensor class can provide low threshold, low noise, high efficiency, and adequate spatial resolution while supporting finer granularity than strip detectors, especially for low-$\pm 0.05$1 tracking in a high-background environment.

5. OBELIX in plasma kinetics

In plasma physics, OBELIX denotes the Orsay Boltzmann equation for ELectrons coupled with Ionization and eXcited states kinetics, used in the study of the electron energy distribution function in a high power impulse magnetron sputtering (HiPIMS) discharge (Rudolph et al., 2021). It is described as a volume-averaged collisional-radiative model in which the electron energy distribution function (EEDF) is computed self-consistently from an isotropic Boltzmann equation, rather than being assumed a priori.

The paper writes the time-dependent isotropic Boltzmann equation for the EEDF as

$\pm 0.05$2

where the terms represent elastic electron-neutral energy flux, electron-electron Coulomb energy flux, Ohmic heating, excitation reactions, ionization reactions, a source of injected secondary electrons, and a loss term associated with quasi-neutrality and diffusion (Rudolph et al., 2021). The EEDF is discretized using the Rockwood formalism with non-uniform energy bins satisfying

$\pm 0.05$3

The model is explicitly contrasted with the Ionization Region Model (IRM). In the IRM, the discharge is represented by cold bulk electrons with a Maxwellian EEDF in the few-eV range and hot secondary electrons as a separate Maxwellian-like high-energy tail (Rudolph et al., 2021). OBELIX instead solves the distribution dynamically and derives rate coefficients from the computed EEDF. In the merged IRM + OBELIX approach, the IRM provides discharge current and voltage constraints, $\pm 0.05$4, sheath voltage $\pm 0.05$5, particle fluxes, and Ohmic heating power, while OBELIX computes the self-consistent EEDF and detailed excited-state kinetics. The discharge voltage is split as

$\pm 0.05$6

The secondary-electron source in OBELIX is represented as

$\pm 0.05$7

with

$\pm 0.05$8

at the energy corresponding to $\pm 0.05$9, and zero elsewhere, so the secondaries are injected essentially monoenergetically at the sheath voltage (Rudolph et al., 2021). The paper also defines the electron energy probability function as

$\pm 3\%$0

The principal result is that the temporal evolution of particle densities agrees very well between IRM and OBELIX, and that the bi-Maxwellian EEDF assumed by the IRM agrees very well with the self-consistent EEDF calculated by OBELIX (Rudolph et al., 2021). OBELIX nonetheless reveals finer structure: early in the pulse the low-energy part is more Druyvesteyn-like, later it becomes more Maxwellian-like, a high-energy tail persists, a strong peak appears near the sheath-injected energy around 460 eV, and a second peak appears about 12 eV below the injection energy. The high-energy peak is small in absolute density, about $\pm 3\%$1 to $\pm 3\%$2 of the cold bulk. For argon, the cost of ionization (COI) computed by OBELIX approaches about 37.9 eV later in the pulse, while the IRM uses about 33 eV for most of the pulse. Even so, the study concludes that assuming a bi-Maxwellian electron energy distribution, composed of a cold bulk population and a hot secondary-electron population, is a good approximation for modeling the HiPIMS discharge. OBELIX thereby functions as a validation-grade kinetic model against which a much faster semi-empirical discharge model can be checked.

6. Obelix in computer security and hadron spectroscopy

In computer security, Obelix is a compiler-based dynamic obfuscation framework for trusted execution environments (TEEs) designed to protect both secret data and the executed code itself against side channels (Wichelmann et al., 23 Sep 2025). Its threat model is strong: the attacker behaves like a malicious hypervisor or operating system, can read encrypted enclave or VM memory, can single-step execution deterministically with instruction granularity, and can measure instruction counts, instruction latency, and cache usage. The framework combines uniform code blocks with oblivious RAM (ORAM) storage for both code and data, so that an attacker cannot identify which code block is executing or which data block is being accessed.

The runtime uses a code controller and a data controller. The code controller fetches the next code block from code ORAM into a code scratchpad; if a block needs memory access, execution jumps to the data controller, which fetches or writes the corresponding data block from data ORAM into a data scratchpad, after which execution returns to the block and then unconditionally back to the code controller (Wichelmann et al., 23 Sep 2025). The prototype chooses optimized Linear ORAM for both code and data, although the paper also analyzes Path ORAM. For $\pm 3\%$3 blocks, Linear ORAM has complexity

$\pm 3\%$4

whereas Path ORAM offers

$\pm 3\%$5

asymptotically but incurs high constant-time controller overheads; with bucket size $\pm 3\%$6, the paper estimates the break-even point around $\pm 3\%$7.

The central protection mechanism is the use of uniform code blocks resistant to single-stepping. Obelix empirically classifies instructions into indistinguishability classes, notably class1 for standard arithmetic, shifts, address computation, and register or immediate moves, and class2 for division instructions (Wichelmann et al., 23 Sep 2025). Blocks are divided into fixed-size instruction slots; in the implementation, each slot is 8 bytes, padded with multi-byte NOPs if necessary. A typical block may already need about 10 slots = 80 bytes, and the paper gives 160 bytes as a reasonable upper bound for code blocks. The system further uses a cost-based optimization over candidate block patterns, with basic-block cost

$\pm 3\%$8

and a genetic algorithm configured with $\pm 3\%$9, $\epsilon \sim N(0,0.01),$0, and a cap of 10,000 generations or 10 seconds.

Obelix also addresses ciphertext side channels in deterministically encrypted memory by enforcing ciphertext freshness through masking, interleaving, or address rotation (Wichelmann et al., 23 Sep 2025). For code, it uses address rotation of the code scratchpad; in the benchmarks, the code scratchpad pool uses ten memory pages, i.e. 40,960 bytes, allowing 160 possible 256-byte aligned locations for 160-byte blocks. For data, the prototype uses rotation and interleaving with 16-byte data blocks split into two 8-byte halves with 8-byte counters interleaved. The framework is implemented as a drop-in compiler extension in LLVM 17 with about 9710 lines added, and the workflow allows developers to mark sensitive functions while the compiler propagates protection through the call tree, initializes data ORAM, computes a block pattern, and rewrites machine code. The cost of this protection is substantial: reported overheads include about 660x–720x for matmul, 163x–231x for modexp, 1424x–3010x for aes, 4205x–4806x for cc20, and up to 102,910x for ecdh. The paper therefore presents Obelix as a broad-coverage software defense whose applicability depends heavily on workload size and performance tolerance.

In hadron spectroscopy, by contrast, Obelix refers to an experiment cited in the review literature on hybrid mesons rather than to a computational framework or device (Meyer et al., 2015). The review discusses Obelix in connection with the reaction

$\epsilon \sim N(0,0.01),$1

and specifically with evidence for the exotic state $\epsilon \sim N(0,0.01),$2 decaying to

$\epsilon \sim N(0,0.01),$3

The quoted resonance parameters are

$\epsilon \sim N(0,0.01),$4

The review’s interpretation is explicitly cautious: it notes that the $\epsilon \sim N(0,0.01),$5 signal in comparable antiproton data arises from a $\epsilon \sim N(0,0.01),$6-wave initial state, whereas the $\epsilon \sim N(0,0.01),$7 signal comes from the $\epsilon \sim N(0,0.01),$8 initial state, and concludes that it is unlikely that the exotic state seen in $\epsilon \sim N(0,0.01),$9 and that seen in $K_{0.1}Li_{0.9}SbO_3$0 are the same (Meyer et al., 2015). It further suggests that the observed structure may arise from re-scattering rather than a genuine exotic resonance. In this usage, then, Obelix is not the subject of the review but one experimental input in a broader controversy over the interpretation of low-mass $K_{0.1}Li_{0.9}SbO_3$1 signals.

Across these domains, the recurrent name designates markedly different epistemic objects: curated data, recommender systems, radiation-tolerant sensors, kinetic plasma models, obfuscation compilers, and experimental collaborations. The only consistent encyclopedia-level characterization is therefore contextual: Obelix is a polysemous research label whose meaning is fixed by capitalization, field, and citation rather than by any common underlying concept.