Thunder: Multifaceted Phenomenon & Benchmark
- Thunder is a polysemous term defined as both the acoustic consequence of lightning and a multifaceted label used across scientific and technological domains.
- In meteorology, thunder is characterized by an initial shock-wave followed by ordinary acoustic radiation, with applications in lightning detection and short-range storm forecasting.
- Non-atmospheric uses of THUNDER include benchmark suites and frameworks in computational pathology, wireless communications, speech processing, robotics, astronomy, and arithmetic geometry.
Thunder denotes both a physical atmospheric phenomenon associated with lightning and a recurrent proper name or acronym in technical literature. In atmospheric science it refers to the acoustic consequence of lightning-generated shock and sound waves; in contemporary research nomenclature it labels benchmark suites, algorithms, educational systems, software frameworks, LLMs, and even an astrophysical object. The term therefore has no single disciplinary referent, and its meaning is determined by domain-specific context (Fineberg et al., 2022, Belagali et al., 24 Dec 2025).
1. Atmospheric meaning: thunder, lightning, and thunderstorms
In the atmospheric sense, thunder is the sound field produced by lightning discharge. One paper describes it as comprising an initial shock-wave phase near the rapidly heated lightning channel, followed by ordinary acoustic radiation after the shock attenuates; the cited spectrum spans roughly $1$ to $600$ Hz and includes infrasound below $20$ Hz (Tar, 2010). In operational meteorology, however, the practical target is often not thunder audibility itself but lightning occurrence as a proxy for thunderstorm activity, because convective thunderstorms are commonly accompanied by lightning (Schön et al., 2019).
This distinction matters because several papers use “thunder” metonymically for thunderstorm occurrence rather than for the acoustic wave. Deep-learning nowcasting work frames the task as “the prediction of thunderstorms and lightning,” but the implemented target is future lightning occurrence in the next 15 minutes (Schön et al., 2019). Likewise, recent convection-permitting post-processing studies define thunderstorm occurrence through nearby LINET lightning observations rather than through direct acoustic measurements (Yousefnia et al., 2023, Yousefnia et al., 2024).
A related misconception is that all “thunder” simulation work includes acoustics. “Thunderscapes” models mesoscale convective systems, cloud microphysics, hydrometeor electrification, and lightning flashes, but explicitly does not model thunder as sound; its scope is thunderstorm structure and visual electrodynamics rather than acoustic propagation (Hao, 2024).
2. Forecasting and simulation of thunderstorms
Short-range thunderstorm prediction in the cited literature is largely formulated as a data-driven inference problem over satellite imagery or numerical weather prediction output. “Make Thunderbolts Less Frightening -- Predicting Extreme Weather Using Deep Learning” uses SEVIRI satellite imagery and recent LINET lightning observations in a Residual UNet++ to predict lightning as a pixel-level binary classification task; its headline result is a probability of detection of more than for lightnings within the next 15 minutes (Schön et al., 2019). The same paper emphasizes the centrality of rare-event imbalance, reporting a positive rate of only .
Post-processing of convection-permitting forecasts has evolved from hand-crafted surrogate predictors toward direct inference from atmospheric structure. “A machine-learning approach to thunderstorm forecasting through post-processing of simulation data” introduces SALAMA, a feedforward network trained on 21 NWP-derived predictors and lightning-defined labels, and reports forecast skill superior to classification based only on NWP reflectivity for lead times up to at least 11 hours (Yousefnia et al., 2023). “Inferring Thunderstorm Occurrence from Vertical Profiles of Convection-Permitting Simulations” then replaces single-level predictors with full vertical profiles of 10 atmospheric variables in SALAMA 1D, obtaining a Brier skill score of $0.288(14)$ versus $0.239(12)$ for the single-level baseline and retaining superior skill to at least 11 hours (Yousefnia et al., 2024). The saliency analysis in that work ties model decisions to physically interpretable structures such as ice near the tropopause, lapse-rate-related temperature and pressure structure, low-level moisture, and the absence of cloud and graupel when storms are unlikely (Yousefnia et al., 2024).
Simulation work occupies a different point in the design space. “Thunderscapes: Simulating the Dynamics of Mesoscale Convective System” introduces a graphics-oriented framework that couples incompressible fluid dynamics, Grabowski-style bulk cloud microphysics, hydrometeor electrification, and dielectric-breakdown-model lightning growth. It represents the evolving mesoscale convective state as , partitions condensate and precipitation into warm and cold components through a temperature-dependent factor , and drives lightning when a threshold condition based on charge state and altitude-dependent breakdown is met (Hao, 2024). The paper is physically inspired rather than meteorologically predictive, but it is notable for coupling thunderstorm development and lightning activity in a single procedural framework.
3. Thunderstorms as high-energy laboratories
Thunderstorms also function as large-scale electrodynamic accelerators. “Observation of Variations in Cosmic Ray Single Count Rates During Thunderstorms and Implications for Large-Scale Electric Field Changes” uses the Telescope Array Surface Detector to measure low-threshold single-count-rate modulations over a area. The observed variations are typically at the $600$0 level, reach up to $600$1, have ground footprints from 6 to 24 km, and move with the storm. Simple CORSIKA simulations reproduce these signatures with potential differences of approximately $600$2–$600$3 across 2 km field layers, corresponding to about 100–200 kV/m (Abbasi et al., 2021). The paper’s methodological claim is that cosmic-ray secondaries provide a large-area probe of thunderstorm electric structure beyond what localized balloons or aircraft can sample.
A more local and more extreme perspective is given by “The prolonged gamma ray enhancement and the short radiation burst events observed in thunderstorms at Tien Shan,” which reports measurements at $600$4 altitude, sometimes within $600$5 m of the active discharge region (Shepetov et al., 2020). The paper distinguishes two emission classes. The first is a thunderstorm ground enhancement lasting about 1.5 minutes, with gamma-ray count-rate increases of about $600$6–$600$7, an electron flux estimate $600$8 above roughly 2 MeV, and abrupt cessation within $600$9 at the onset of a nearby discharge (Shepetov et al., 2020). The second is a sub-millisecond lightning-associated burst with gamma excess flux estimates of order $20$0 above $20$1 keV and a detectable $20$2 MeV electron component, interpreted as consistent with leader-associated acceleration in intense localized fields (Shepetov et al., 2020). The same study reports an indication of a $20$3 keV annihilation feature and argues that a neutron-monitor anomaly is more plausibly due to electric-field modulation of negative muons than to photonuclear neutron production (Shepetov et al., 2020).
These papers situate thunderstorm research at the interface of meteorology, cosmic-ray physics, and radiation transport. A plausible implication is that “thunder” in contemporary atmospheric literature often denotes not only a weather hazard but also a diagnostic window into cloud-scale electric-field organization.
4. THUNDER as an acronym or system name in computation and engineering
A large fraction of recent uses of THUNDER are acronymic rather than atmospheric. The term names unrelated artifacts across computational pathology, wireless communication, speech processing, computer vision, avatar generation, and language-model adaptation (Belagali et al., 24 Dec 2025, Ryu et al., 30 Jun 2026, Trachu et al., 2024, Zhou et al., 2022, Daněček et al., 18 Apr 2025, Kim et al., 18 Jun 2025).
| Use | Domain | Defining feature |
|---|---|---|
| THUNDER | Computational pathology | “Tile-level Histopathology image UNDERstanding benchmark” (Belagali et al., 24 Dec 2025) |
| THUNDER | Backscatter communications | “Transformer-Hypernetwork-Controlled Deep-Unfolded Phase-Aware Channel Estimation Refinement” (Ryu et al., 30 Jun 2026) |
| Thunder | Speech enhancement | Unified regression-diffusion enhancement with Brownian bridge and single-step capability (Trachu et al., 2024) |
| Thunder | Image denoising | “Thumbnail based Fast Lightweight Image Denoising Network” (Zhou et al., 2022) |
| THUNDER | Talking-head avatars | “Talking Heads Under Neural Differentiable Elocution Reconstruction” (Daněček et al., 18 Apr 2025) |
| Thunder-LLM | LLMs | Low-budget Korean adaptation of Llama 3.1 8B (Kim et al., 18 Jun 2025) |
In computational pathology, THUNDER is not a model but a benchmark suite. In the TICON study it is the main benchmark for tile classification when slide-level spatial context is unavailable; the evaluation covers 16 tile classification tasks in total, comprising 12 original THUNDER tasks plus 4 SPIDER tasks under the THUNDER-SPIDER subset, and uses average k-NN F1-score under a no-slide-context protocol (Belagali et al., 24 Dec 2025). Because slide context is absent, TICON is applied in isolated inference mode, $20$4, where each tile embedding is processed independently and the transformer effectively behaves like a deep MLP (Belagali et al., 24 Dec 2025).
In wireless communication, THUNDER is a model-driven estimator for phase-drifting monostatic SISO backscatter links. It initializes from PACE, then performs bounded deep-unfolded Gauss–Newton refinement on the exact phase-exponential observation model, with a transformer extracting pilot-wide phase context and a hypernetwork generating solver controls and pilot-reliability weights (Ryu et al., 30 Jun 2026). At $20$5 dB and $20$6, the paper reports an NMSE gain of 8.9 dB over the strongest learning-based baseline (Ryu et al., 30 Jun 2026).
In signal processing, Thunder for speech enhancement uses a Brownian bridge to unify regression and diffusion within a single model that predicts clean speech rather than the score function, enabling competitive performance even with a single reverse step (Trachu et al., 2024). Thunder for image denoising instead compresses the problem into an RGB thumbnail subspace using a wavelet-based Thumbnail Subspace Encoder and reconstructs details with a Subspace Projection based Refine Module; the reported model size is 2.68M parameters with 18.81 GFlops, achieving $20$7 on SIDD and $20$8 on DND (Zhou et al., 2022).
In audiovisual generation, THUNDER for talking heads introduces a differentiable analysis-by-audio-synthesis loop: a frozen mesh-to-speech model reconstructs speech representations from generated FLAME mouth motion, and the resulting mismatch is used as cross-modal supervision to improve lip-sync while retaining stochastic facial expressiveness (Daněček et al., 18 Apr 2025). In multilingual language modeling, Thunder-LLM adapts Llama 3.1 8B to Korean through tokenizer expansion, continual pretraining on 102B tokens, and SFT plus DPO post-training; Thunder-LLM-Ins attains the best reported average Korean score among the compared $20$9B open models while remaining competitive in English (Kim et al., 18 Jun 2025).
5. Infrastructure, education, planning, and named objects
Some uses of Thunder designate systems intended for training, planning, or cataloging rather than benchmark suites or neural architectures. “Thunder CTF: Learning Cloud Security on a Dime” defines Thunder CTF as a scaffolded, scenario-based capture-the-flag system for learning cloud security on Google Cloud Platform (Springer et al., 2021). Its initial deployment contains six levels covering topics such as open storage buckets, overprovisioned permissions, source-history key leakage, unsanitized error messages, IAM privilege escalation, and metadata credential compromise through SSRF (Springer et al., 2021). In a Portland State University course deployment with 48 students and 36 survey responses, mean ratings were 3.94 for understanding cloud security issues, 3.94 for cloud navigation skills, and 4.56 for helpfulness of the hint system on a 1–5 scale (Springer et al., 2021).
In robotics, “Experience-Based Planning with Sparse Roadmap Spanners” introduces THUNDER as an experience-based motion-planning framework for high-dimensional robots in varying environments (Coleman et al., 2014). It stores experience in a SPARS/SPARS2 sparse graph rather than in individual paths, runs retrieve-and-repair in parallel with planning from scratch, and was evaluated on a 30-DOF humanoid under stability, self-collision, and obstacle constraints. After 10,000 trials in varying environments, THUNDER was reported as 10.0 times faster than Lightning, 12.3 times faster than planning from scratch, and 98.8% more memory-efficient than Lightning, with a database of 235 kB versus 19,373 kB (Coleman et al., 2014).
In astronomy, Thunder is the proper name assigned to a newly identified bow-shock pulsar wind nebula, PWN J1631−4721, powered by PSR J1631−4722 and projected within the supernova remnant G336.7+0.5, dubbed Nimbus (Lazarević et al., 21 Jun 2026). Assuming a distance of 7 kpc, the radio nebula has an extent of 0 (1 pc), the X-ray counterpart extends 2 (3 pc), the radio spectral index is 4, the X-ray photon index is 5, and the inferred system age is approximately 30–45 kyr (Lazarević et al., 21 Jun 2026). Here “Thunder” is neither acronym nor atmospheric metaphor, but an astrophysical source name.
6. Thunder as a bibliographic surname in arithmetic geometry and Diophantine counting
In number theory and arithmetic geometry, “Thunder” can refer to the mathematician Jeff Thunder rather than to a phenomenon or acronym. “Counting rational points on a Grassmannian” is explicitly framed as a refinement of Thunder’s 1993 work on rational points of bounded twisted height on flag varieties and of his lattice-counting results over 6 (Kim, 2019). Kim’s contribution is to count all primitive rational points on the Grassmannian of bounded twisted height over 7, removing a structural restriction from Thunder’s earlier theorem while retaining the same leading secondary term in the asymptotic error (Kim, 2019).
A related example is Stewart’s “On the number of solutions of decomposable form inequalities,” which generalizes Thunder’s 2001 estimate from decomposable forms of finite type to forms of essentially finite type (Stewart, 2024). Thunder’s original theorem gave the uniform upper bound 8 for finite-type forms, while Stewart extends the framework and also proves an analogue of Thunder’s sharper asymptotic formula with main term 9 (Stewart, 2024). In this bibliographic context, the capitalized label THUNDER is absent; what remains is Thunder as an authorial reference point in the development of bounded-height counting theory.
Taken together, these uses show that THUNDER is not a unitary concept but a highly polysemous label spanning geophysics, atmospheric electricity, machine learning, communications, robotics, security pedagogy, astronomy, and mathematics. The commonality is lexical rather than technical: each literature supplies its own definition, objective, and evaluative regime.
