MINERVA: Multifaceted Research Platforms

• MINERVA is a versatile label applied to distinct research infrastructures, from Fermilab’s neutrino experiments to exoplanet arrays and advanced video reasoning benchmarks.
• In neutrino physics, MINERvA employs a fine-grained scintillator detector to measure neutrino–nucleus interactions, providing critical data for oscillation studies and model validation.
• Applied computing implementations of Minerva enhance cybersecurity, contact-center AI, and decentralized query processing with significant performance and efficiency gains.

MINERVA is a reused research name rather than a single project. In the arXiv literature it denotes, among other things, the Fermilab neutrino–nucleus experiment usually styled MINERvA, the Miniature Exoplanet Radial Velocity Array and its Australian counterpart, the “Medium-band Imaging with NIRCam to Explore ReVolutionary Astrophysics” JWST treasury survey, “Multimodal INterpretablE Reasoning Video Annotations” for video reasoning, and several computing systems in cybersecurity, contact-center AI, and decentralized data processing (Perdue, 2011).

1. Nomenclature and disambiguation

A common misconception is to treat MINERVA as a single acronym with a stable expansion. The literature instead uses the name for unrelated facilities, datasets, and software systems across particle physics, astronomy, machine learning, and distributed systems (Swift et al., 2014).

Representative usages include the following (Horner et al., 6 May 2026, Nagrani et al., 1 May 2025, Muzzin et al., 25 Jul 2025):

Usage	Expansion or definition	Research area
MINERvA	Main INjector ExpeRiment for $\nu$-A	Neutrino scattering
MINERVA	Miniature Exoplanet Radial Velocity Array	Exoplanet detection
MINERVA-Australis	Southern-hemisphere counterpart to MINERVA	Exoplanet follow-up
MINERVA	Multimodal INterpretablE Reasoning Video Annotations	Video reasoning benchmark
MINERVA	Medium-band Imaging with NIRCam to Explore ReVolutionary Astrophysics	JWST extragalactic survey
Minerva	File-based ransomware detector	Cybersecurity
Minerva CQ	Real-time voice-based agent assist product	Contact-center AI
Minerva	Decentralized collaborative query processing over IPFS	Distributed data systems

The name therefore functions as an overloaded label whose meaning is determined entirely by domain context. In particle physics, capitalization itself is diagnostic: MINERvA denotes the neutrino experiment, whereas several later uses adopt MINERVA or Minerva without relation to Fermilab.

2. MINERvA in neutrino physics

In neutrino physics, MINERvA is a dedicated Fermilab experiment for neutrino–nucleus interactions in the few-GeV regime, operating in the NuMI beamline and using both Low-Energy and Medium-Energy beam configurations that peak at about $3$ GeV and $6$ GeV, respectively (Lu et al., 2021). The experiment began taking data in the Fall of 2009, detector installation was completed in March 2010, and the full program recorded neutrino and antineutrino scattering data from 2009 to 2019 (Perdue, 2011).

The detector is a fine-grained scintillator instrument with electromagnetic and hadronic calorimetry, an active plastic-scintillator tracking region, and passive nuclear targets. Across the supplied papers these targets include carbon, iron, lead, helium, water, and scintillator, with the target architecture motivated by direct study of $\nu$-A dependence and nuclear-medium effects in a common beam (Osmanov, 2011). Muons from charged-current interactions are measured with the downstream magnetized MINOS near detector, which improves momentum and charge determination for exiting tracks (Lu et al., 2021).

The scientific role of MINERvA is twofold. First, it is a cross-section experiment: it measures inclusive and exclusive channels, event kinematics, nuclear effects, and parton-distribution-sensitive observables in a beam-energy regime central to long-baseline oscillation programs (Perdue, 2011). Second, it is a nuclear-effects experiment: because oscillation analyses infer $E_\nu$ from visible final-state particles rather than direct neutrino measurement, uncertainties in binding energy, Fermi motion, final-state interactions, multinucleon dynamics, and detector response feed directly into oscillation systematics (Lu et al., 2021).

The scale of the dataset is also distinctive. The preservation study reports that between 2013 and 2019 MINERvA collected Low Energy and Medium Energy samples containing over 4 million charged-current $\nu_\mu$ interactions in the active $CH$ detector region, roughly half as many $\bar{\nu}_\mu$ interactions, and about $\mathcal{O}(1\text{ million})$ additional interactions in passive targets including He, $H_2O$, C, Fe, and Pb (Fine et al., 2020). The same paper stresses that these are the only currently available data at intermediate and high momentum transfers for multiple nuclear targets in the same beam, making MINERvA unusually relevant to DUNE-era interaction modeling.

3. Measurements, analysis methods, and data preservation in MINERvA

MINERvA’s published measurements span quasielastic, CCQE-like, inclusive charged-current, pion, kaon, coherent, and nuclear-ratio observables. One foundational result is the first direct neutrino-scattering measurement of nuclear dependence in inclusive charged-current cross sections, reported as per-nucleon ratios for C, Fe, and Pb relative to scintillator. Those data show depletion at low Bjorken $3$0 and enhancement at high $3$1, with both effects increasing with nuclear mass number, and the observed trends are not reproduced by GENIE or by the alternative nuclear-modification models considered (Tice, 2014).

A separate landmark measurement is the first MINERvA measurement of an exclusive few-GeV electron-neutrino charged-current quasielastic-like process on a hydrocarbon target. The differential cross sections were reported versus electron energy, electron angle, and $3$2, and the ratio to a corresponding MINERvA $3$3 result in $3$4 was used to test the assumption that flavor dependence is dominated by the charged-lepton mass. Within the quoted uncertainties, the data were consistent with GENIE within $3$5 in all cases (Wolcott, 2015).

Methodologically, MINERvA introduced and exploited transverse-imbalance observables to isolate nuclear dynamics. In the single-transverse kinematic imbalance study on CH, the imbalance variable

$3$6

was used to probe Fermi motion, final-state interactions, multinucleon effects, and intranuclear rescattering. The analysis developed an elastically scattering contained proton selection based on the last 6 measurement points and a transverse-momentum-scale correction using Cauchy fits; the reconstructed $3$7 distribution was found to be clearly sensitive to FSIs, unlike the one-dimensional proton momentum spectrum (Lu et al., 2016).

Flux determination is another major technical contribution. MINERvA used neutrino–electron elastic scattering as an in situ standard candle, identifying 135 LE and 810 ME events. According to the 2021 review, these measurements reduced the uncertainty at the flux peak from $3$8 to $3$9 in LE and from $6$0 to $6$1 in ME (Lu et al., 2021). This is integral to absolute cross-section extraction, which in the preservation paper is summarized schematically as

$6$2

with signal, background, efficiency, and flux propagated through $6$3 systematic “universes” using modified ROOT histogram classes such as MnvH1D (Fine et al., 2020).

The preservation program formalizes this analysis model into a long-term data product. Its three principal components are a single ROOT tuple with low-level and high-level reconstructed objects, the MINERvA Analysis Toolkit (MAT) for systematic-uncertainty calculations on event tuples, and a MAT-based software package that can reproduce published results and serve as a template for new analyses (Fine et al., 2020). The preserved dataset is expected to be about $6$4 TB in total, and a one-dimensional analysis with about $6$5 histograms can run over the full FHC ME dataset in about 1 hour using Fermilab’s default 2 GB batch allocations. The collaboration’s stated goal is to make all aspects of the preserved data product publicly available, but the same paper states candidly that full independence from present collaborators has not yet been achieved by modern neutrino experiments.

4. MINERVA in exoplanet instrumentation

In exoplanet science, MINERVA denotes the Miniature Exoplanet Radial Velocity Array, a dedicated ground-based observatory built for radial-velocity discovery and characterization of nearby exoplanets around bright stars (Swift et al., 2014). The original US-based facility is organized as a robotic array of four PlaneWave CDK-700 telescopes, each of $6$6 m aperture, whose combined collecting area is used with a stable, bench-mounted, fiber-fed spectrograph. The primary science program is a dedicated 3-year RV survey, while the secondary program is high-precision transit photometry. The spectrograph is specified at $6$7 over $6$8–$6$9 nm, and commissioning demonstrated sub-mmag photometric precision on 3–5 minute timescales as well as transit observations of WASP-52b (Swift et al., 2014).

Minerva-Australis extends the same architectural idea to southern-sky TESS follow-up. The 2019 commissioning paper describes it as the Southern Hemisphere counterpart to the original MINERVA array, located at the University of Southern Queensland’s Mount Kent Observatory and designed to comprise up to six independently operated $\nu$0 m PlaneWave CDK-700 telescopes feeding a single Kiwispec R4-100 spectrograph (Addison et al., 2019). Its science goals include follow-up, confirmation, characterization, and mass measurement of bright transiting planets discovered by TESS, along with Rossiter–McLaughlin measurements, non-transiting RV searches, transit timing variation studies, and ephemeris refinement.

A later decade review presents MINERVA-Australis as Australia’s only professional dedicated exoplanet discovery and follow-up facility and describes the operational array as four PlaneWave Instruments CDK-700 telescopes, purchased off the shelf in 2017, with a collecting area equivalent to a $\nu$1 m diameter telescope at roughly an order of magnitude lower cost than a bespoke single-aperture instrument (Horner et al., 6 May 2026). In RV mode the array feeds a high-resolution spectrograph with $\nu$2, housed in an environmentally controlled class 100,000 cleanroom and stabilized to $\nu$3 K. After first light in 2018, the facility typically achieved about $\nu$4 m/s RV precision for TESS targets of $\nu$5–11 and photometric precision of a few millimagnitudes.

The scientific impact is already measurable. The decade review states that MINERVA-Australis has contributed to the discovery of 40 new exoplanets and has also continued the legacy of the Anglo-Australian Planet Search through an AAPS-Legacy survey of long-term radial-velocity targets (Horner et al., 6 May 2026). In this domain, MINERVA names a family of dedicated, robotic, multi-telescope infrastructures optimized for high-cadence RV and transit work rather than a single observatory.

5. Survey and benchmark infrastructures bearing the name MINERVA

Two recent uses of MINERVA denote research infrastructures for data-rich astronomy and multimodal evaluation rather than detectors or telescopes.

The JWST program MINERVA is “Medium-band Imaging with NIRCam to Explore ReVolutionary Astrophysics.” It is an approved Cycle-4 treasury survey with $\nu$6 prime hours and $\nu$7 parallel hours, targeting UDS, COSMOS, AEGIS, and GOODS-N with eight NIRCam medium bands and two MIRI filters (Muzzin et al., 25 Jul 2025). The survey reaches a $\nu$8 depth of $\nu$9 mag in F300M, covers $E_\nu$0 arcmin$E_\nu$1 with NIRCam and $E_\nu$2 arcmin$E_\nu$3 with MIRI, and increases existing JWST medium-band coverage in at least eight bands by $E_\nu$4. Its stated science goals include uncovering the physics of enigmatic sources hiding in broadband catalogs, improving systematics on stellar mass functions and number densities by factors of $E_\nu$5, and enabling resolved mapping of stellar mass and star formation at $E_\nu$6 (Muzzin et al., 25 Jul 2025). The same paper reports simulated photometric-redshift improvement from a PRIMER-like broadband-only $E_\nu$7 of $E_\nu$8 to $E_\nu$9 for a MINERVA-like filter set.

In machine learning, MINERVA is “Multimodal INterpretablE Reasoning Video Annotations,” a benchmark for complex video reasoning (Nagrani et al., 1 May 2025). The dataset contains 1,515 questions over 223 videos, each question is 5-way multiple choice, and each example includes a hand-crafted reasoning trace with timestamps, object/action evidence, and explicit logical steps. Video lengths range from under 2 minutes to 100 minutes, the mean length is 12 minutes, average reasoning-trace length is 92 words, and 99.6% of reasoning traces include timestamps. Benchmarking shows a large gap between models and humans: human performance is 92.54%, while the best reported model, Gemini 2.5 Pro Thinking, reaches 66.20% (Nagrani et al., 1 May 2025).

The benchmark’s analytic contribution is its error taxonomy and scoring rubric. It organizes failures into perceptual correctness, temporal localization, logical reasoning, and completeness, and introduces MiRA with 0–2 Likert scoring on those four axes (Nagrani et al., 1 May 2025). The main empirical finding is that model failures are primarily related to temporal localization, followed by visual perception errors, rather than logical or completeness errors. This challenges the adequacy of outcome-only video QA evaluation and makes the MINERVA name, in this instance, associated with interpretable supervision rather than only final-answer accuracy.

6. Applied computing systems named Minerva

In cybersecurity, Minerva is a file-based ransomware detector built around the principle of monitoring file effects rather than process behavior (Hitaj et al., 2023). The paper argues that process-based behavioral detectors are vulnerable to evasion through altered API sequences, benign-padding actions, delayed encryption, multi-process orchestration, or other changes in execution trace, whereas file-derived evidence is causally downstream of the ransomware’s destructive objective. Minerva is therefore described as robust by design against evasion attacks, and the reported headline result is that over 99% of detected ransomware are identified within 0.52 sec of activity, enabling data loss prevention with near-zero overhead (Hitaj et al., 2023).

Minerva CQ is a separate, real-time, voice-based agent-assist system for contact centers (Agrawal et al., 16 Sep 2025). The paper defines its organizing logic as an agentic loop, “observe → understand → decide → act → assist → learn,” combining real-time transcription, intent and sentiment detection, entity recognition, contextual retrieval, dynamic customer profiling, partial conversational summaries, and proactive workflow triggering. In a live production A/B study with two cohorts of 50 agents each and approximately 40,000 production voice interactions, the system is reported to achieve a 38% reduction in average handling time, from 4m 43s to 2m 55s, a 33% uplift in Lead-to-Enquiry conversion, and a 4.8% uplift in booking conversion (Agrawal et al., 16 Sep 2025). The same paper reports that, in logs from 10,000 calls, about 7,000 queries were answered from a validated FAQ cache rather than RAG, saving about 6 seconds per avoided retrieval and totaling roughly 11.7 hours of cumulative latency saved.

In distributed systems, Minerva is a decentralized collaborative query engine over IPFS (Yao et al., 2023). It places Apache Drill above IPFS, uses DHT lookup to identify content providers, introduces a fat Merkle tree to reduce flattening delay, and adds metadata and provider caching through MinervaCache. The system supports read and write query plans with decentralized workers, and the paper reports up to $\nu_\mu$0 acceleration compared to original IPFS data query, average latency of $\nu_\mu$1 second in Internet-like environments, and up to $\nu_\mu$2 performance acceleration over centralized query with raw data shipment (Yao et al., 2023).

Taken together, these computing uses show that MINERVA frequently labels systems designed around structural robustness: file-level invariants instead of mutable process traces, stateful agentic workflows instead of prompt-only assistance, and compute-near-data execution instead of raw-data centralization. This suggests that, outside physics and astronomy, the name has become associated less with a specific acronym than with purpose-built infrastructure for difficult operational constraints.