THOR: Cross-Domain Technical Systems

• THOR is a polyvalent acronym that denotes a variety of domain-specific projects, from high-resolution Galactic surveys in astronomy to advanced neuromorphic processors and robotics.
• THOR implementations emphasize architectural flexibility, modular design, and heterogeneity handling, enabling robust frameworks for energy estimation, data processing, and autonomous control.
• Practical applications of THOR include mapping the interstellar medium, optimizing energy-throughput in digital systems, and enabling efficient text-to-SQL and human–object interaction generation.

Searching arXiv for recent THOR papers across domains to ground the article with citations. THOR is a polyvalent acronym used across several technical domains for survey infrastructures, computational frameworks, and hardware systems rather than a single unified concept. In astronomy, it most prominently denotes The H I/OH/Recombination line survey of the inner Milky Way, a Karl G. Jansky VLA program targeting atomic, molecular, ionized, and polarized radio emission in the first Galactic quadrant (Beuther et al., 2016). In computer engineering and machine learning, the same name has been assigned to an all-digital neuromorphic processor optimized for Energy–Throughput efficiency (Senapati et al., 2022), a generic Gaussian-process-based estimator for on-device DNN training energy (Zhang et al., 27 Jan 2025), a text-to-SQL enterprise retrieval module (Shi et al., 13 Jul 2025), an inductive model for hyper-relational knowledge-graph link prediction (Yu et al., 5 Feb 2026), and a diffusion model for text-conditioned human–object interaction generation (Wu et al., 2024). Additional uses include a theorem-proving framework integrating LLMs with hammers (Jiang et al., 2022), a GPU-oriented planetary global circulation model (Mendonça et al., 2016), a lunar analog rover (Womack et al., 2014), a humanoid contact-rich control framework (Li et al., 30 Oct 2025), and an HST data-reduction pipeline for Roman precursor fields (Terry et al., 16 Jun 2026). This multiplicity suggests that “THOR” functions in the literature as a reusable project label whose meaning is entirely domain-specific.

1. Radio-astronomical survey of the Milky Way

In Galactic astronomy, THOR denotes The H I/OH/Recombination line survey of the inner Milky Way (Beuther et al., 2016). It is a VLA L-band survey designed to obtain a high-resolution, multi-line view of the interstellar medium by observing the H I 21 cm line, four OH 18 cm lines, hydrogen radio recombination lines, 1–2 GHz continuum, and full polarization (Beuther et al., 2016). The survey covers Galactic longitudes from 14.5° to 67.4° and latitudes $|b| \le 1.25^\circ$, with typical angular resolution around 20″, thereby filling a long-standing gap between coarse 21 cm surveys and higher-resolution infrared or submillimeter Galactic plane mapping (Beuther et al., 2016).

The survey’s scientific scope is explicitly multi-phase. H I traces the atomic medium, including warm neutral gas in emission and cold neutral gas via absorption and H I self-absorption; OH probes diffuse and translucent molecular gas as well as maser populations; radio recombination lines trace ionized gas in H II regions; and polarized continuum supports Faraday rotation and Galactic magnetic-field studies (Beuther et al., 2016). The observing strategy is a VLA C-array mosaic using WIDAR, with simultaneous spectral windows across the 1–2 GHz band and complementary combination with VGPS and Effelsberg data where recovery of large-scale structure is required (Beuther et al., 2016).

Several THOR derivative studies sharpen this general picture. The continuum-source catalog from the first half of the survey reports approximately 4400 sources at 10–25″ resolution and finds a double-peaked spectral-index distribution with maxima near $\alpha = -1$ and $\alpha = 0$, corresponding respectively to steep-spectrum synchrotron-dominated sources and flat-spectrum thermal emission (Bihr et al., 2016). This catalog uses THOR’s broad bandwidth to distinguish H II regions, SNRs, and extragalactic sources through radio spectral indices (Bihr et al., 2016). A dedicated OH absorption study characterizes THOR as the highest-resolution unbiased OH absorption survey of the first Galactic quadrant to date, uses continuum sources stronger than $F_{\rm cont}\ge 0.1$ Jy/beam, and shows that OH is a useful tracer of low-column molecular gas, including lines of sight with OH absorption but no ${}^{13}$CO emission (Rugel et al., 2018).

2. THOR and Galactic supernova remnants

One of THOR’s most visible astronomical applications is the identification of supernova remnants in the inner Milky Way. Using THOR 1.4 GHz continuum combined with VGPS and mid-infrared data from Spitzer GLIMPSE, MIPSGAL, and WISE, Anderson et al. identified 76 candidate SNRs in the THOR field by isolating radio structures with low MIR counterparts, after excluding objects in the WISE H II region catalog (Anderson et al., 2017). The relevant search region for that work is $67.4^\circ > l > 17.5^\circ$, $|b|<1.25^\circ$, and the candidates are, on average, smaller and fainter than previously cataloged remnants, with mean angular radius $6.4' \pm 4.7'$ versus $11.0' \pm 7.8'$ for known SNRs in the same footprint (Anderson et al., 2017).

The same work emphasizes the longstanding “missing SNR” problem: the Galactic census contains far fewer remnants than expected from supernova-rate arguments. The THOR-selected candidates would more than double the known SNR count within the survey area if confirmed, yet would still leave the total Galactic SNR population below the expected level by about a factor of two (Anderson et al., 2017). The survey therefore functions not merely as a radio map but as a systematic selection engine for low-surface-brightness or small-angular-size remnants that were difficult to isolate in earlier, lower-resolution Galactic plane datasets.

A subsequent confirmation study tests these THOR-selected candidates using spectral indices, morphology, and polarization diagnostics (Dokara et al., 2018). It confirms two candidates as genuine SNRs: G27.06+0.04 and G51.26+0.11 (the latter identified within the more complex candidate G51.21+0.11) (Dokara et al., 2018). In that analysis, THOR+VGPS morphology provides shell or partial-shell evidence, while non-thermal spectral indices are derived not from THOR itself but from a 150–1400 MHz TGSS–NVSS spectral-index map, chosen because THOR’s C-configuration spatial filtering can bias spectral indices for large shells (Dokara et al., 2018). The study explicitly concludes that spectral index plus morphology is a robust confirmation strategy, whereas fractional linear polarization cannot distinguish between SNRs and H II regions, likely because of contamination from diffuse Galactic synchrotron emission (Dokara et al., 2018). This establishes a precise methodological role for THOR in SNR work: it supplies the candidate list, shell morphology, and radio context, while complementary datasets supply decisive spectral evidence.

3. Neuromorphic processor for spiking neural networks

In neuromorphic hardware, THOR denotes an all-digital neuromorphic processor implemented in 28 nm FDSOI CMOS and designed to maximize Energy–Throughput (ET) efficiency rather than optimizing energy per synaptic operation or throughput density in isolation (Senapati et al., 2022). The processor is a single-core design operating at up to 400 MHz at 0.9 V, with 256 LIF neurons, approximately 65 k synapses in a 256 × 256 fully connected topology, on-chip online learning (SDSP), and a 0.77 mm² core area (Senapati et al., 2022).

The architectural novelty is centered on memory hierarchy and update scheduling. THOR employs a $P$-way parallel neuron engine with $\alpha = -1$0, two interleaved neuron memory banks, two interleaved synapse memory banks, and parallel neuron and synapse logic units that allow a neuron event of 256 SOPs to be completed in 9 cycles, versus 512 cycles in the ODIN-style sequential baseline (Senapati et al., 2022). The synapse memory uses SCM in the final implementation, despite SRAM being slightly more area-efficient in the chosen configuration, because SCM offers more flexible voltage–frequency scaling potential in future work (Senapati et al., 2022).

At 0.9 V, 400 MHz, THOR reports 1.40 pJ/SOP, 7.84 G SOP/s, and an ET efficiency of 7.29 G TSOP²/mm²·J·s, which the paper states is a 3× improvement over the previous state of the art among digital neuromorphic processors (Senapati et al., 2022). The system also includes multi-threaded schedulers for input and output spikes, decoupling spike handling from the main compute pipeline, and uses extensive clock and input gating to reduce dynamic power (Senapati et al., 2022). The emphasis on ET efficiency suggests an explicit repositioning of digital neuromorphic design away from the traditional trade-off in which ultra-low-energy operation is obtained only at extremely low throughput.

4. Machine-learning frameworks and AI systems

Several papers reuse THOR as the name of AI frameworks whose shared feature is structural generality rather than a common application domain.

THOR for on-device training energy estimation is a generic framework that models DNN training energy through a layer-wise energy additivity assumption and Gaussian Process regressors trained on small variant networks (Zhang et al., 27 Jan 2025). It profiles end-to-end energy on heterogeneous devices, infers per-layer energy using additive and subtractive decompositions, and then predicts full-model training energy by summing layer estimates (Zhang et al., 27 Jan 2025). Across smartphones, Jetson boards, and a server, the reported improvement is a reduction in MAPE from roughly 40% for a FLOP-based baseline to roughly 10%, with an absolute reduction of up to 30 percentage points (Zhang et al., 27 Jan 2025). The same framework is used to guide energy-aware pruning, reducing actual energy consumption to about 49.2% of the original while keeping within a 50% energy budget (Zhang et al., 27 Jan 2025).

THOR for text-to-SQL is “Transformer Heuristics for On-Demand Retrieval,” an enterprise module that converts natural-language questions into verified, read-only SQL analytics through a multi-agent architecture (Shi et al., 13 Jul 2025). Its components include a Supervisor Agent, dynamic schema retrieval, a SQL Generation Agent constrained to single-statement SELECT queries, a bounded Self-Correction & Rating loop with up to five regeneration attempts, and a Result Interpretation Agent that produces human-readable summaries (Shi et al., 13 Jul 2025). The smoke-test results are framed in operational terms rather than benchmark metrics: for example, delay-analysis reports in a logistics case study fall from roughly 1 hour/day of manual work to 2–3 minutes/query, while weekly performance review becomes effectively auto-generated (Shi et al., 13 Jul 2025).

THOR for hyper-relational knowledge graphs is an inductive link-prediction technique designed for the fully inductive case where both entity and relation vocabularies are unseen at inference time (Yu et al., 5 Feb 2026). It introduces relation foundation graphs and entity foundation graphs, each defined by interaction types that are agnostic to specific IDs, followed by two parallel graph encoders and a transformer decoder trained with masked prediction (Yu et al., 5 Feb 2026). On 12 datasets, the method reports improvements of 66.1%, 55.9%, and 20.4% over the best-performing rule-based, semi-inductive, and fully-inductive baselines, respectively (Yu et al., 5 Feb 2026). A plausible implication is that THOR’s central abstraction is not merely “induction” but structural invariance under entity and relation permutation.

THOR for text-to-human–object interaction generation is a diffusion model that generates temporally coherent 3D sequences of human and object motion from text plus object geometry (Wu et al., 2024). Its defining mechanism is relation intervention: at each denoising step, primitive human and object motions are produced, human-centric relative rotations and translations are computed, and a lightweight intervention network outputs residual corrections to object motion (Wu et al., 2024). The associated Text-BEHAVE dataset contains 2377 clips, 18 object models in 12 categories, and 440,840 frames (Wu et al., 2024). THOR improves over VAE- and MDM-based baselines on text-motion alignment and realism, including R-Precision Top-1 = 0.250 and FID = 1.983 on Text-BEHAVE (Wu et al., 2024).

THOR for theorem proving, styled “Thor” in the paper, is a hybrid framework that teaches a LLM to invoke ATP-backed hammers through a learned <hammer> action (Jiang et al., 2022). On the PISA dataset, it raises theorem-proving success from 39% for the LM baseline to 57%, while solving 8.2% of problems that neither the LM alone nor the hammer alone can solve (Jiang et al., 2022). On MiniF2F, it reaches performance comparable to expert-iteration methods with substantially smaller computational cost (Jiang et al., 2022). Here, the recurring THOR motif is orchestration: the system delegates a structurally hard subproblem—premise selection—to a specialized external engine.

5. Robotics, control, and autonomous systems

In robotics, THOR has been applied to both planetary surface systems and humanoid control, but with entirely different meanings.

The Telescope-deployment High-vacuum teleOperated Rover is a small proof-of-concept lunar rover built almost entirely from commercial off-the-shelf components and tested in the Lunar and Airless Bodies Simulator at the University of Colorado Boulder (Womack et al., 2014). It is intended both as a survivability demonstrator in lunar-relevant vacuum and thermal cycling and as a platform for deploying a Kapton-based radio telescope arm for the LUNAR program (Womack et al., 2014). The rover mass is 5.1 lb, total available torque is 13.8 in-lb, and the effective drive ratio is 134:1 (Womack et al., 2014). During testing it survived aggressive thermal cycling meant to emulate more than one simulated lunar year, though issues such as communication loss and battery expansion were observed (Womack et al., 2014). The work suggests that carefully engineered COTS hardware can persist in high-vacuum, thermally extreme environments for meaningful mission-like durations.

A much newer robotics use is Thor for humanoid whole-body reactions in intense contact-rich environments (Li et al., 30 Oct 2025). This framework targets forceful humanoid interaction tasks on the Unitree G1, a 29-DOF, 35 kg humanoid, using a decoupled RL architecture with separate policies for upper body, waist, and lower body, and a force-adaptive torso-tilt (FAT2) reward derived from a quasi-static force and torque analysis (Li et al., 30 Oct 2025). On real hardware, the robot achieves $\alpha = -1$1 N backward pulling and $\alpha = -1$2 N forward pulling, corresponding to 68.9% and 74.7% improvements over the best-performing baseline (Li et al., 30 Oct 2025). It also pulls a 130 N loaded rack and opens a 60 N fire door with one hand (Li et al., 30 Oct 2025). The shared design logic with other THOR systems is noteworthy: it uses an explicit structural decomposition to make a high-dimensional control problem tractable.

6. Scientific infrastructure, simulation, and Earth-observation models

THOR also appears in large-scale scientific software and observational pipelines where the name denotes a flexible infrastructure rather than a narrowly defined algorithm.

In planetary science, THOR is a non-hydrostatic, compressible global circulation model built from scratch to solve the full three-dimensional Euler equations on a rotating sphere without relying on the shallow-atmosphere or hydrostatic approximations (Mendonça et al., 2016). It uses an icosahedral grid to avoid the pole problem, spring dynamics to reduce grid imprinting, and a split-explicit plus horizontally explicit, vertically implicit integration strategy to manage acoustic time-step constraints (Mendonça et al., 2016). The model is GPU-oriented, part of the open-source Exoclimes Simulation Platform, and validated on both the Held–Suarez Earth benchmark and a hot Jupiter benchmark (Mendonça et al., 2016). This establishes THOR as a deliberately approximation-light atmospheric “planetary lab.”

In HST data processing, THOR denotes the Terry Hubble Observations of Roman reduction pipeline for the Roman Galactic Bulge precursor program GO-17776 (Terry et al., 16 Jun 2026). The pipeline combines hst1pass and thor_go to turn calibrated HST/ACS and HST/WFC3 imaging into calibrated crowded-field star catalogs and stacked reference images (Terry et al., 16 Jun 2026). An intermediate catalog contains roughly 22 million detected sources across 332 HST fields, with typical per-field counts of 70,000–100,000 stars for ACS and 40,000–70,000 for WFC3 (Terry et al., 16 Jun 2026). The companion HAMRR tool performs cone-search queries on THOR-derived catalogs and can return calibrated photometry, astrometry, image cutouts, color–magnitude diagrams, and luminosity functions (Terry et al., 16 Jun 2026).

In Earth observation, THOR is also used for a compute-adaptive foundation model that unifies Sentinel-1, Sentinel-2, and Sentinel-3 inputs across native resolutions from 10 m to 1000 m in a single architecture (Forgaard et al., 22 Jan 2026). The provided excerpt states that THOR uses randomized patch and image sizes during pre-training so that a single set of weights can operate with different patch sizes at inference, trading off resolution and computation without retraining (Forgaard et al., 22 Jan 2026). Because the excerpt is truncated, more detailed architectural claims would be speculative; nevertheless, the core characterization as a versatile multi-sensor EO foundation model is explicit (Forgaard et al., 22 Jan 2026).

7. Nomenclature, recurring design patterns, and scope of the acronym

Across these disparate uses, THOR almost never names a generic scientific principle. It names a project, platform, or framework whose technical identity is local to its field. The acronym expansions vary widely: The H I/OH/Recombination line survey of the Milky Way (Beuther et al., 2016), Transformer Heuristics for On-Demand Retrieval (Shi et al., 13 Jul 2025), inducTive link prediction for Hyper-relational knOwledge gRaphs (Yu et al., 5 Feb 2026), Telescope-deployment High-vacuum teleOperated Rover (Womack et al., 2014), and others. In some cases, the acronym is only partially or informally expanded; in others, such as the neuromorphic processor or humanoid-control framework, “THOR” functions chiefly as a project name (Senapati et al., 2022, Li et al., 30 Oct 2025).

Several recurring structural motifs nonetheless appear. Architectural flexibility is central in the GCM (Mendonça et al., 2016), the EO foundation model (Forgaard et al., 22 Jan 2026), and the HST reduction pipeline (Terry et al., 16 Jun 2026). Heterogeneity handling is explicit in the on-device energy estimator across platforms and models (Zhang et al., 27 Jan 2025), the text-to-SQL system across enterprise schemas (Shi et al., 13 Jul 2025), and the hyper-relational KG model across unseen entity and relation vocabularies (Yu et al., 5 Feb 2026). Delegation or modular decomposition appears in the theorem-proving framework that offloads premise selection to hammers (Jiang et al., 2022), the humanoid controller that decouples body regions (Li et al., 30 Oct 2025), and the neuromorphic processor that separates scheduling from the main neuron-update pipeline (Senapati et al., 2022). This suggests that the repeated selection of “THOR” may correlate less with domain semantics than with a design ethos emphasizing robustness, modularity, and extensibility.

A common misconception would be to treat THOR as a single cross-domain method. The literature does not support that view. “THOR” is instead a homonymous label attached to unrelated systems in astronomy, atmospheric modeling, AI, robotics, hardware, and data infrastructure. Any technical discussion therefore requires immediate disambiguation by field and citation context.