MIGA: Multi-domain Research Applications

• MIGA is an overloaded acronym used in varied disciplines, from precision gravity measurements to micro-gesture analysis, highlighting distinct methodological approaches.
• In gravitation, MIGA denotes a matter-wave interferometric antenna using cold rubidium atoms over a 150 m baseline for sub-Hz precision gravity and strain sensing.
• In computer vision, MiGA represents the Micro-Gesture Analysis challenge, focusing on fine-grained gesture recognition with 3D-CNNs and multimodal fusion techniques.

MIGA is an overloaded research acronym spanning experimental gravitation, computer vision, natural language interfaces, optimization, adversarial robustness, quantitative finance, and video generation. In physics, it denotes the Matter-wave laser Interferometric Gravitation Antenna, an underground long-baseline atom-interferometry facility at the Laboratoire Souterrain à Bas Bruit (LSBB) in Rustrel, France, designed for precision gravity measurements and as a demonstrator for future low-frequency gravitational-wave detection (Geiger et al., 2015, Canuel et al., 2016, Canuel et al., 2022). In computer vision, MiGA denotes the Micro-Gesture Analysis challenge series at IJCAI (Li et al., 2023, Liu et al., 8 Jun 2026). Other papers use MIGA for a unified multi-task generation framework for conversational text-to-SQL, a multiprocessing interface genetic algorithm, a mutual information-guided attack on denoising models, a mixture-of-experts framework for stock prediction, and an infinite-frame long-video generation method (Fu et al., 2022, Iliyas et al., 27 May 2025, Li et al., 10 Mar 2025, Yu et al., 2024, Feng et al., 18 May 2026).

1. Research uses of the acronym

The acronym appears in multiple, technically unrelated literatures. The table lists the principal uses represented in the cited corpus.

Usage	Domain	Representative paper
Matter-wave laser Interferometric Gravitation Antenna	Atom interferometry, gravitation, geophysics	(Canuel et al., 2016)
Micro-Gesture Analysis challenge series	Micro-gesture recognition	(Li et al., 2023)
MultI-task Generation frAmework	Conversational text-to-SQL	(Fu et al., 2022)
Multiprocessing Interface Genetic Algorithm	Hyperparameter optimization	(Iliyas et al., 27 May 2025)
Mutual Information-Guided Attack	Adversarial attacks on denoising models	(Li et al., 10 Mar 2025)
Mixture-of-Experts with Group Aggregation	Stock market prediction	(Yu et al., 2024)
MIGA	Train-free infinite-frame video generation	(Feng et al., 18 May 2026)

In bibliographic practice, disambiguation is therefore essential. Within gravitation and quantum sensing, “MIGA” almost always refers to the French underground matter-wave antenna. Within IJCAI challenge literature, “MiGA” refers to Micro-Gesture Analysis.

2. Matter-wave laser Interferometric Gravitation Antenna

MIGA, in its gravitation sense, is a large-scale gravity antenna based on matter-wave interferometry and an underground cold-atom instrument designed to measure gravity gradients and strains with long-baseline atom interferometry (Canuel et al., 2022, Beaufils et al., 2022). Its scientific goals are stated in two complementary forms across the project literature: precision gravity and gravity-gradient measurements for geoscience, hydrology, and underground monitoring; and a demonstrator for future gravitational-wave detectors based on atom interferometry, targeting the sub-Hz or infrasound regime and, in broader programmatic reviews, the intermediate band around $0.1\text{–}10$ Hz (Geiger et al., 2015, Canuel et al., 2017, Balaz et al., 27 Mar 2025).

Project descriptions evolved over time. Early papers presented a 300 m instrument with three $^{87}$Rb atom interferometers horizontally aligned and interrogated by a resonant optical cavity, together with two perpendicular 300 m galleries and future widenings every 50 m (Geiger et al., 2015, Canuel et al., 2016). Later construction and source papers describe a 150 m vacuum vessel, three cold rubidium atom interferometers regularly spaced along the antenna baseline, and two perpendicular 150 m galleries allowing a possible evolution toward a 2D instrument geometry (Canuel et al., 2022, Beaufils et al., 2022). A 2025 long-baseline atom-interferometry review places MIGA alongside MAGIS-100 as one of the key precursor detectors already under construction, and describes it as a two-arm horizontal detector at an underground laboratory in France (Balaz et al., 27 Mar 2025).

The site is LSBB in Rustrel, France, a low-noise underground laboratory in a karstic mountain. Across the project papers, LSBB is characterized as exceptionally favorable because of low seismic and magnetic background noise, passive thermal stability, distance from major anthropogenic disturbances, and direct relevance for hydro-geophysical studies (Geiger et al., 2015, Canuel et al., 2022, Canuel et al., 2017).

3. Measurement principle and instrument architecture

The basic operating principle is light-pulse atom interferometry with common optical interrogation of multiple spatially separated atom interferometers (Canuel et al., 2016, Canuel et al., 2017, Beaufils et al., 2022). In the later construction description, MIGA consists of a 150 m vacuum vessel housing in-cavity interrogation beams that interrogate and correlate an array of three cold rubidium atom interferometers, regularly spaced along the antenna baseline, plus two large vacuum tanks at the ends for optics, in-cavity mirrors, and cavity control systems (Canuel et al., 2022). In the cold-atom source description, each 150 m arm contains three horizontal Bragg atom interferometers separated by nearly 75 m and driven by two common, counter-propagating interrogation beams at two heights separated by 30.6 cm (Beaufils et al., 2022).

The atom species is rubidium, with $^{87}$Rb specified in the earlier project and laser-system papers (Canuel et al., 2016, Sabulsky et al., 2019). Atom sources are based on 2D–3D MOTs. One source paper reports a standard 2D–3D MOT configuration producing clouds at temperature $\approx 2~\mu\mathrm{K}$, Raman preparation to the pure hyperfine state $F=2,m_F=0$, and reduced effective temperature $\approx 150~\mathrm{nK}$ (Canuel et al., 2022). The dedicated cold-atom-source paper gives a more detailed sequence: a high-flux 2D MOT loaded into a 3D MOT, moving-molasses launch at $v_0=4.45~\mathrm{m\,s^{-1}}$, two-stage velocity-selective Raman preparation, and fluorescence detection after momentum-to-internal-state labeling (Beaufils et al., 2022).

The core AI relations are stated explicitly in the corpus. For constant acceleration $a$,

$\phi_a = k_{\rm eff}\, a\, T^2,$

and for a gravity gradient $\Gamma$ between two interferometers separated by baseline $^{87}$0,

$^{87}$1

These formulas recur across the MIGA literature and summarize the scaling of differential phase with effective wave vector, pulse separation time, and baseline (Canuel et al., 2022, Canuel et al., 2016, Beaufils et al., 2022).

The interrogation sequence is a Bragg $^{87}$2 sequence. A total interferometric time $^{87}$3 is repeatedly reported, corresponding to $^{87}$4 in the later source and antenna descriptions (Canuel et al., 2022, Canuel et al., 2017, Beaufils et al., 2022). Earlier architectural papers also described a three-pulse geometry near the ballistic apex of vertically launched atoms, with $^{87}$5 for first-order Bragg diffraction at $^{87}$6 (Canuel et al., 2016).

A defining feature is common interrogation. The same cavity field interrogates the different atom interferometers, correlating them and suppressing common-mode noise such as laser phase or frequency noise and platform vibrations to the extent that they couple identically to all interferometers (Canuel et al., 2022). Differential measurements between AIs then yield gravity-gradient information and establish the basis for gravitational-wave sensing via tidal effects along the baseline (Canuel et al., 2022, Canuel et al., 2017).

4. Subsystems, noise environment, and projected performance

The MIGA papers devote substantial attention to infrastructure, vacuum technology, cold-atom sources, and fibered lasers (Canuel et al., 2022, Sabulsky et al., 2019, Beaufils et al., 2022). The 150 m vacuum vessel is described as being assembled from mostly 6 m long, 5 mm thick SS 304 sections with 50 cm aperture, interconnected with 500 mm helicoflex metallic gaskets and 20 cm bellows to accommodate thermal elongation during bake-out (Canuel et al., 2022). Prototype validation on a 6.4 m atom gradiometer yielded a residual pressure of $^{87}$7 mbar after baking at $^{87}$8 for $^{87}$9 days, with outgassing rate $^{87}$0 mbar·l·s$^{87}$1·cm$^{87}$2; the paper states that a total pumping speed of $^{87}$3 l·s$^{87}$4 would be sufficient to obtain a residual pressure better than $^{87}$5 mbar across the full antenna if replicated (Canuel et al., 2022).

The fibered laser system produces and distributes all optical frequencies required for $^{87}$6Rb cooling, trapping, Raman manipulation, and fluorescence detection (Sabulsky et al., 2019). It is based on four frequency-agile C-band telecom diode lasers at 1560 nm, frequency-doubled to 780 nm after Erbium-doped fiber amplification. The reference laser is stabilized on a saturated-absorption signal, and three seed lasers are phase-locked through optical beat notes. The paper reports 14 polarization-maintaining fiber outputs, laser linewidths below 40 kHz at 1 s at 780 nm, maximum deviation of 64.5 kHz over 25 hours for the reference, and Raman-output powers of 443 mW and 445 mW per fiber (Sabulsky et al., 2019). Five such systems had been produced for MIGA (Sabulsky et al., 2019).

Noise characterization is central because the antenna aims at sub-Hz gravity sensing and possible future gravitational-wave applications. The project literature identifies seismic noise, magnetic noise, laser frequency and phase noise, vacuum-related backgrounds, atom shot noise, and technical noise as dominant considerations (Canuel et al., 2022, Canuel et al., 2016). Earlier analyses model Newtonian noise from both Rayleigh waves and atmospheric infrasound, and explicitly state that MIGA is intended to understand how geophysical fields couple into sub-Hz gravitational observables (Canuel et al., 2016, Junca et al., 2019).

Several quantitative performance statements appear in the literature. A 2017 design paper for the 200 m configuration gives a peak strain sensitivity of about $^{87}$7 at 2 Hz in the initial configuration, dominated by detection noise (Canuel et al., 2017). The 2016 simulation study for the earlier 300 m design states an initial strain sensitivity on the order of $^{87}$8 in amplitude spectral density and estimates that modeled seismic Newtonian noise limits the strain sensitivity to about $^{87}$9 at 0.1 Hz for $\approx 2~\mu\mathrm{K}$0 m and depth $\approx 2~\mu\mathrm{K}$1 m (Canuel et al., 2016). A later proceedings paper does not provide finalized strain curves, but reiterates the ambition to open the infrasound window for gravitational-wave observation (Canuel et al., 2022).

An important external constraint comes from the 2024 study of underground seismic correlations at LSBB and other sites (Janssens et al., 2024). For the LSBB GAS–RAM pair, the paper states that significant coherence is observed more than 90% of the time in the entire frequency range of 0.01–20 Hz, and that above 20 Hz at least 50% of the time significant coherence is observed. It further concludes that correlated Newtonian noise from underground body waves is a serious consideration for instruments such as MIGA and ELGAR, and recommends dedicated projections and mitigation strategies (Janssens et al., 2024). This does not negate the rationale for underground siting; rather, it sharpens the distinction between reducing uncorrelated environmental noise and handling correlated gravity perturbations.

5. MiGA as the Micro-Gesture Analysis challenge

In computer vision, MiGA denotes the Micro-Gesture Analysis challenge series at IJCAI, centered on recognizing subtle, short-duration, often spontaneous body movements linked to latent cognitive or affective states (Li et al., 2023, Gu et al., 11 Jul 2025, Liu et al., 8 Jun 2026). The challenge uses datasets including iMiGUE and SMG. The iMiGUE-based classification track is reported under slightly different label-count conventions across papers: one paper states “32 MGs plus one non-MG class,” another states “32 MG categories plus one non-MG class,” and later challenge papers describe a 32-way setting with “31 micro-gesture categories or a background/non-MG class” (Li et al., 2023, Xu et al., 15 Jun 2025, Liu et al., 8 Jun 2026, Gu et al., 11 Jul 2025). The common technical thread is fine-grained recognition of subtle motions from short clips or skeleton sequences under highly imbalanced class distributions.

The 2023 winning solution by team HFUT-VUT introduced a PoseC3D-style 3D-CNN pipeline over 3D heatmap volumes with a joint skeletal and semantic embedding loss (Li et al., 2023). The total loss is

$\approx 2~\mu\mathrm{K}$2

with

$\approx 2~\mu\mathrm{K}$3

On the iMiGUE test set under the competition protocol, that solution reached Top-1 accuracy $\approx 2~\mu\mathrm{K}$4, surpassing the runner-up by $\approx 2~\mu\mathrm{K}$5; on SMG, the Joint+Limb ensemble achieved Top-1 $\approx 2~\mu\mathrm{K}$6 and Top-5 $\approx 2~\mu\mathrm{K}$7 (Li et al., 2023).

Subsequent MiGA challenge papers show rapid performance progression. A 2025 skeleton-based solution built on PoseC3D, an extended 41-joint topology with facial landmarks, improved temporal resampling, and semantic label embeddings, and reported Top-1 $\approx 2~\mu\mathrm{K}$8, ranking third on the official leaderboard (Xu et al., 15 Jun 2025). Also in 2025, HFUT-VUT’s “MM-Gesture” introduced a six-modality fusion framework over joint, limb, RGB, Taylor-series video, optical-flow video, and depth video, with PoseConv3D and Video Swin Transformer branches. It achieved $\approx 2~\mu\mathrm{K}$9 Top-1 on iMiGUE and ranked first in the classification track (Gu et al., 11 Jul 2025). In the online-recognition track on SMG, HFUT-VUT reported an F1 score of $F=2,m_F=0$0, outperforming the previous state of the art by $F=2,m_F=0$1 (Liu et al., 13 Jul 2025).

The 2026 winner from XInsight Lab added a self-supervised RGB stream pretrained on 120K unlabeled in-domain clips via masked video modeling and fused it with supervised multi-stream branches (Liu et al., 8 Jun 2026). The self-supervised RGB baseline alone achieved $F=2,m_F=0$2 Top-1 on iMiGUE, while the final five-stream ensemble reached $F=2,m_F=0$3, surpassing the previous best, MiGA’25 winner at $F=2,m_F=0$4, by $F=2,m_F=0$5 percentage points (Liu et al., 8 Jun 2026). Across these papers, MiGA functions not as a single method but as a benchmark ecosystem for subtle-motion understanding, with shifting emphasis from skeleton-only recognition to multimodal and self-supervised learning.

6. Other computational expansions of MIGA

Outside atom interferometry and micro-gesture analysis, the acronym is used by several unrelated machine-learning papers (Fu et al., 2022, Iliyas et al., 27 May 2025, Li et al., 10 Mar 2025, Yu et al., 2024, Feng et al., 18 May 2026).

In conversational semantic parsing, MIGA stands for a two-stage unified MultI-task Generation frAmework for conversational text-to-SQL (Fu et al., 2022). It reformulates conversational text-to-SQL as a prompt-driven Seq2Seq problem over SQL Generation, Related Schema Prediction, Turn Switch Prediction, and Final Utterance Prediction. On development sets, the reported results are $F=2,m_F=0$6 QM/IM on SparC and $F=2,m_F=0$7 QM/IM on CoSQL, improving further with PICARD to $F=2,m_F=0$8 and $F=2,m_F=0$9, respectively (Fu et al., 2022).

In optimization for disease prediction, MIGA denotes a Multiprocessing Interface Genetic Algorithm that parallelizes fitness evaluation in genetic search for multilayer perceptron hyperparameters (Iliyas et al., 27 May 2025). The paper keeps population, selection, crossover, mutation, and elitism intact, and replaces the sequential fitness-evaluation bottleneck with concurrent execution. It reports approximately $\approx 150~\mathrm{nK}$0 tuning-time reduction relative to a standard genetic algorithm, with best accuracies of $\approx 150~\mathrm{nK}$1 for breast cancer, $\approx 150~\mathrm{nK}$2 for Parkinson’s disease, and $\approx 150~\mathrm{nK}$3 for chronic kidney disease (Iliyas et al., 27 May 2025).

In adversarial robustness, MIGA denotes Mutual Information-Guided Attack on denoising models for semantic manipulation (Li et al., 10 Mar 2025). The method minimizes task-relevant mutual information between clean and denoised images while maintaining perceptual plausibility through content and reconstruction losses. On ImageNet-10 with a ResNet-50 downstream classifier, the paper reports post-denoising accuracies reduced to $\approx 150~\mathrm{nK}$4 for Xformer, $\approx 150~\mathrm{nK}$5 for Restormer, $\approx 150~\mathrm{nK}$6 for PromptIR, and $\approx 150~\mathrm{nK}$7 for AFM, while keeping LPIPS$\approx 150~\mathrm{nK}$8 around $\approx 150~\mathrm{nK}$9 (Li et al., 10 Mar 2025).

In quantitative finance, MIGA denotes Mixture-of-Experts with Group Aggregation for stock market prediction (Yu et al., 2024). The framework combines sparse routing, linear experts, and inner-group attention. The abstract highlights that MIGA-Conv reaches $v_0=4.45~\mathrm{m\,s^{-1}}$0 excess annual return on the CSI300 benchmark; the paper also reports long-short AR/IR values such as $v_0=4.45~\mathrm{m\,s^{-1}}$1 for MIGA-Rec on CSI1000 (Yu et al., 2024).

In generative video modeling, MIGA is the name of a train-free infinite-frame long-video generation method that augments FIFO-style frame-level autoregressive diffusion with a two-stage alignment mechanism and a dual consistency enhancement mechanism (Feng et al., 18 May 2026). The paper does not expand the acronym. On VBench-Long with a VideoCrafter2 backbone at 128 frames, it reports overall score $v_0=4.45~\mathrm{m\,s^{-1}}$2 for MIGA versus $v_0=4.45~\mathrm{m\,s^{-1}}$3 for FIFO-Diffusion; on Wan2.1 backbones at 161 frames, it reports overall score $v_0=4.45~\mathrm{m\,s^{-1}}$4 versus $v_0=4.45~\mathrm{m\,s^{-1}}$5 (Feng et al., 18 May 2026).

Taken together, these usages show that “MIGA” is not a stable single concept across contemporary research literature. It is instead a high-collision acronym whose meaning is determined almost entirely by disciplinary context: matter-wave gravitation in experimental physics, Micro-Gesture Analysis in affective computing, and several independent method names in NLP, optimization, security, finance, and generative modeling.