SISMA: Diverse Approaches in ML, Astrophysics & Seismology

Updated 12 July 2026

SISMA is an acronym representing three distinct systems: a diffusion-based model for semantic face synthesis, a stellar spectroscopic archive, and a geophysical seismic hazard framework.
In semantic face synthesis, SISMA employs a Mamba state-space model in latent space to generate photorealistic, mask-conditioned images with significantly improved inference speed.
The astrophysical and seismological usages of SISMA offer curated high-resolution datasets and integrated analytical tools to advance studies in stellar variability and time-dependent seismic hazard assessment.

Searching arXiv for papers on "SISMA" to ground the article in the specified literature. arXiv Search Query: all:SISMA

arXiv Search Query: ti:"SISMA" OR abs:"Spectroscopic Indicators in a SeisMic Archive" OR abs:"Semantic Face Image Synthesis with Mamba"

SISMA is an acronym used in arXiv-indexed research for three distinct entities with no direct disciplinary connection: an attention-free diffusion architecture for semantic face image synthesis based on Mamba; a VO-compliant archive of HARPS spectroscopy and CoRoT photometry for variable and active stars; and an ASI-funded prototype system for time-dependent seismic hazard assessment that integrates seismological prediction, GNSS/GPS, SAR/InSAR, and scenario earthquakes (Botti et al., 22 Sep 2025, Rainer et al., 2016, Panza et al., 2017). The common acronym therefore denotes, respectively, a generative model, a scientific database, and a geophysical monitoring framework.

1. Disambiguation and scope

On arXiv, the term “SISMA” refers to unrelated research artifacts in machine learning, stellar astrophysics, and seismology. The overlap is lexical rather than conceptual.

SISMA expansion	Domain	Core function
Semantic Face Image Synthesis with Mamba	Generative modeling	Semantic image synthesis of faces from masks
Spectroscopic Indicators in a SeisMic Archive	Asteroseismology and stellar spectroscopy	Public archive of HARPS spectra, indicators, and CoRoT data
The SISMA prototype system	Seismology and geodesy	Integration of CN, geodesy, and scenario earthquakes

The most recent usage is the face-synthesis architecture “Semantic Face Image Synthesis with Mamba,” published in 2025 (Botti et al., 22 Sep 2025). An earlier astrophysical usage, “Spectroscopic Indicators in a SeisMic Archive,” was published in 2016 and formalized SISMA as a public, VO-compliant database built within the FP7 SpaceInn project (Rainer et al., 2016). A seismological usage, discussed in 2017, presented SISMA as a prototype system funded by the Italian Space Agency for integrating geophysical modeling and Earth Observation in time-dependent seismic hazard assessment (Panza et al., 2017).

2. SISMA in semantic face image synthesis

In computer vision, SISMA denotes a diffusion-model-based architecture for semantic face image synthesis that replaces transformer attention with Mamba, a selective state-space model (Botti et al., 22 Sep 2025). The target task is semantic image synthesis of human faces: the input is a semantic mask $m$ in which each pixel is labeled with classes such as hair, skin, nose, eyes, and mouth, and the output is a photorealistic face image whose geometry follows that mask while appearance variables such as texture, color, shading, expression, and identity remain generative degrees of freedom. The paper positions this setting against diffusion and transformer backbones that are computationally expensive because self-attention over $L$ tokens has $O(L^2)$ memory and compute complexity.

SISMA operates in the latent space of a pretrained VAE. A frozen VAE encoder maps an image $I \in \mathbb{R}^{C \times H \times W}$ to a latent $\hat{I} \in \mathbb{R}^{c \times h \times w}$ with $c=4$ , $h=H/f$ , and $w=W/f$ . A patchify layer converts the latent to tokens

$x \in \mathbb{R}^{L \times D}, \qquad L = \frac{hw}{p^2},$

with hidden dimension $D=1024$ in the reported experiments. The semantic mask, provided in one-hot form for the 19 CelebAMask-HQ classes, is passed through a single $L$ 0 Conv2d to obtain a per-pixel embedding in $L$ 1, then patchified into

$L$ 2

The core backbone consists of 24 stacked SISMA blocks, each comprising a Self-Mamba module, a Cross-Mamba module, and an MLP, with AdaLN conditioning on the diffusion timestep $L$ 3. Self-Mamba models intra-sequence dependencies analogously to self-attention but through state-space recurrence. Cross-Mamba performs semantic conditioning by letting the state-space parameters depend partly on the mask embedding: $L$ 4 This differs from standard input-dependent Mamba parameterization, in which $L$ 5, $L$ 6, and $L$ 7 are all functions of $L$ 8. In the SISMA design, the mask controls shape by modulating the input-to-state dynamics and discretization step sizes, while the content tokens govern the output mapping.

Training uses Flow Matching rather than an adversarial objective. The forward interpolation is

$L$ 9

and the network $O(L^2)$ 0 is optimized with

$O(L^2)$ 1

No adversarial loss, perceptual loss, or explicit reconstruction loss is described; only the SISMA blocks and mask encoder are trained, while the VAE encoder and decoder remain frozen. This yields a conditional velocity predictor in latent space for a rectified-flow or diffusion-like generative process.

The model is trained on CelebAMask-HQ, using images center-cropped and resized to $O(L^2)$ 2, batch size 8, learning rate 0.004, AdamW with default parameters, no weight decay, EMA decay 0.9999, and 150k training iterations (Botti et al., 22 Sep 2025). The paper emphasizes that the design is intended to preserve diffusion-level quality while reducing the computational overhead associated with attention and semantic-aware normalization schemes such as SPADE and CLADE.

The reported empirical results on CelebAMask-HQ place SISMA among mask-conditioned face synthesis methods evaluated by FID and LPIPS (Botti et al., 22 Sep 2025). The paper reports FID $O(L^2)$ 3 and LPIPS $O(L^2)$ 4 for SISMA, compared with FID $O(L^2)$ 5 and LPIPS $O(L^2)$ 6 for SDM. Earlier GAN-based methods are listed with FID values of 38.5 for Pix2PixHD, 29.2 for SPADE, 29.1 for DAGAN, 20.8 for SCGAN, 30.6 for CLADE, 25.9 for GroupDNet, and 21.5 for INADE; LPIPS is reported as $O(L^2)$ 7 for Pix2PixHD, SPADE, DAGAN, SCGAN, and CLADE, $O(L^2)$ 8 for GroupDNet, and $O(L^2)$ 9 for INADE. In the paper’s interpretation, LPIPS $I \in \mathbb{R}^{C \times H \times W}$ 0 corresponds to effectively deterministic outputs for a given mask.

The computational comparison with SDM is central. For single-image generation with 200 diffusion or flow steps on an RTX 4090, SDM is reported at 653M parameters and 22.79 s, whereas SISMA is reported at 799M parameters and 8.51 s. The paper summarizes this as an approximately $I \in \mathbb{R}^{C \times H \times W}$ 1 or “3× faster” inference speed despite the slightly larger parameter count. The claimed source of the speedup is the attention-free state-space backbone together with simpler conditioning.

Relative to related work, the paper contrasts SISMA with GAN-based systems such as SPADE, SEAN, CLADE, INADE, and SemanticStyleGAN, and with diffusion-based systems such as SDM, ControlNet, T2I-Adapter, Composer, and SCA-DM. Its specific distinctions are that it is diversity-driven, does not require a reference style image, does not use SPADE or CLADE normalization, and replaces attention with Cross-Mamba. The model is specialized for semantic mask-based face synthesis rather than generic controllable diffusion.

The stated limitations are equally specific. Validation is restricted to CelebAMask-HQ; the paper notes that extension to more complex scene datasets such as COCO-Stuff may require more sophisticated mask injection strategies. Explicit style-image conditioning is not supported, text conditioning is absent, and no detailed robustness or fairness analysis is presented. The paper also does not include extensive ablations on the number of SISMA blocks, removal of Self-Mamba or Cross-Mamba, or alternative derivations of state-space matrices from masks and content. These omissions matter because the design choice $I \in \mathbb{R}^{C \times H \times W}$ 2 and $I \in \mathbb{R}^{C \times H \times W}$ 3 is presented as the key mechanism for shape control.

4. SISMA as a spectroscopic archive for stellar variability and asteroseismology

In stellar astrophysics, SISMA stands for “Spectroscopic Indicators in a SeisMic Archive,” a VO-compliant public database created within the FP7 SpaceInn project to preserve and exploit the ground-based spectroscopic follow-up of CoRoT asteroseismic targets and stars with similar variability classes (Rainer et al., 2016). After quality filtering, the archive contains 7103 high-resolution HARPS spectra of 261 stars; the initial set was 7135 spectra, of which 32 were discarded. It stores the spectra themselves together with a rich layer of derived indicators, including mean line profiles, indices of differential rotation, activity and emission lines, $I \in \mathbb{R}^{C \times H \times W}$ 4, radial velocities, and homogeneous atmospheric parameters $I \in \mathbb{R}^{C \times H \times W}$ 5, $I \in \mathbb{R}^{C \times H \times W}$ 6, and $I \in \mathbb{R}^{C \times H \times W}$ 7.

The stellar sample includes 72 main CoRoT asteroseismic targets and auxiliary targets from the same variability classes. The main-sequence sample comprises OB (8), Be (41), SPB (23), SPB-bCep (2), bCep (6), $I \in \mathbb{R}^{C \times H \times W}$ 8 CVn (4), HgMn (2), BAF (13), $I \in \mathbb{R}^{C \times H \times W}$ 9 Sct (97), $\hat{I} \in \mathbb{R}^{c \times h \times w}$ 0 Sct- $\hat{I} \in \mathbb{R}^{c \times h \times w}$ 1 Dor hybrids (1), $\hat{I} \in \mathbb{R}^{c \times h \times w}$ 2 Dor (22), TTau (1), and solar-like oscillators (14). The evolved-star sample comprises $\hat{I} \in \mathbb{R}^{c \times h \times w}$ 3 Cygni (1), S Doradus (2), red giants (23), and RR Lyrae (1). For 72 CoRoT targets, SISMA also stores CoRoT light curves in RAW, BAR, and BARREG versions, the last being BAR data with missing or invalid points inpainted following Pires et al. 2015.

All spectra were obtained with HARPS on the ESO 3.6-m telescope at La Silla under Large Programmes LP182.D-0356 and LP185.D-0056, spanning December 2008 to January 2013. The spectral coverage is 3780–6910 Å. Two observing modes are represented: EGGS, with resolving power $\hat{I} \in \mathbb{R}^{c \times h \times w}$ 4 and typical S/N $\hat{I} \in \mathbb{R}^{c \times h \times w}$ 5 at 5800 Å, and HAM, with $\hat{I} \in \mathbb{R}^{c \times h \times w}$ 6 and typical S/N $\hat{I} \in \mathbb{R}^{c \times h \times w}$ 7. Although HARPS provides official DRS products, the SISMA team reprocessed the e2ds spectra at Brera Observatory because they found the default blaze removal and merging unsatisfactory for detailed variability work.

The reduction pipeline removes the blaze function order by order, computes per-pixel S/N, applies wavelength calibration and barycentric correction, rebins spectra to 0.1 Å, performs iterative order-by-order normalization, and merges overlapping orders with S/N weighting. It produces five-column unmerged spectra (_full.fits: wavelength, flux, normalized flux, S/N, order number) and two-column merged spectra (_nor.fits: wavelength, normalized flux). The paper explicitly warns that normalization is automated and that particular regions, especially Balmer lines and blue metal lines, may require user inspection or renormalization.

SISMA’s scientific utility derives from its precomputed diagnostics. Mean line profiles are obtained through Least-Squares Deconvolution using VALD line masks over the wavelength regions 4415–4805 Å, 4915–5285 Å, and 5365–6505 Å, excluding the blue end, the red telluric-dominated end, and Balmer regions. These profiles are stored in _mean.fits and visualized in OBJECT_profmed.pdf. The same profiles support Fourier-transform estimation of $\hat{I} \in \mathbb{R}^{c \times h \times w}$ 8 and the differential-rotation indicator $\hat{I} \in \mathbb{R}^{c \times h \times w}$ 9, with thresholds $c=4$ 0 for solar-like differential rotation, $c=4$ 1 for rigid rotation, and $c=4$ 2 for anti-solar differential rotation. Radial velocities come from the HARPS DRS in HAM mode and from Gaussian centroids of LSD profiles in EGGS mode.

For cool stars, SISMA defines a relative Ca II H&K activity index,

$c=4$ 3

which is not calibrated to the Mount Wilson $c=4$ 4 index and is intended as a comparative tracer within a time series or across similar spectral types. For H $c=4$ 5-emitting objects, mostly Be stars, the archive stores the violet-to-red peak ratio

$c=4$ 6

Atmospheric parameters are derived homogeneously with SME version 3.3 and VALD in the 5160–5190 Å region, using spectra shifted to the rest frame and averaged when multiple observations exist. External uncertainties are reported as approximately 240 K in $c=4$ 7, 0.32 dex in $c=4$ 8, and 0.33 dex in $c=4$ 9 for cool main-sequence stars, rising to approximately 400 K, 0.50 dex, and 0.40 dex for early-type stars.

The archive is accessible through http://sisma.brera.inaf.it/ and through the SpaceInn Seismic Plus portal at http://voparis-spaceinn.obspm.fr/seismic-plus/. Its backend uses PostgreSQL, with a Java/Servlet/JSP web interface on Apache Tomcat, RGraph for plotting, and STIL for reading FITS files. A common misconception is to treat SISMA as a raw spectral repository only; in fact, its distinguishing feature is the coupling of high-resolution time-series spectra with precomputed indicators tuned to variability and asteroseismic analysis.

5. SISMA as a prototype for integrating seismology and geodesy

In seismology, SISMA denotes a multidisciplinary prototype project funded by the Italian Space Agency whose aim is to integrate seismological prediction tools—primarily the CN algorithm and scenario earthquakes—with geodetic methods and Earth Observation, including continuous GNSS/GPS monitoring and SAR/InSAR (Panza et al., 2017). The paper is explicit that earthquakes cannot be predicted with precision; the project instead addresses intermediate-term, middle-range forecasting of main shocks above a pre-assigned threshold. The full name cited from Panza et al. 2011 is “The SISMA prototype system: integrating Geophysical Modeling and Earth Observation for time-dependent seismic hazard assessment.”

The seismological core is the CN algorithm, developed by Keilis-Borok and collaborators and used operationally in Italy since 1998. CN targets strong earthquakes above region-specific magnitude thresholds in large tectonic regions with linear dimensions of hundreds of kilometers. Its alarms typically last months to several years. The paper summarizes the Italian experiment as follows: 9 strong earthquakes occurred in monitored territory, and 7 were predicted in real time. For the Northern region, the baseline annual probability of a target event is reported as approximately 15% in normal conditions and approximately 48% during alarms; for the Central region, approximately 12% in normal conditions and approximately 45% during alarms. The probability of a target event when no alarm is reported as approximately 2–3%.

The geodetic contribution is organized around a transect-based analysis rather than a standard 2D strain-rate map. In the standard approach, strain-rate components are derived from a velocity field $h=H/f$ 0 through

$h=H/f$ 1

SISMA instead defines 50 km-wide transects, typically 180–200 km long, oriented according to the tectonic setting, projects GNSS velocities onto the transect direction,

$h=H/f$ 2

and examines spatially localized velocity gradients and differential motion across opposite sides of the transect. If the mean relative velocity is $h=H/f$ 3 and the recurrence time of target earthquakes is $h=H/f$ 4, the accumulated displacement is estimated as

$h=H/f$ 5

The principal case study is the Central Italy seismic crisis beginning in August 2016. CN had an ongoing alarm for the Central region since November 1, 2012. In the Amatrice transect, defined along azimuth $h=H/f$ 6 and passing through Amatrice, GNSS data from approximately January 2005 to mid-August 2016 show a clear eastward increase in velocity from the Tyrrhenian side toward Amatrice, with a peak gradient in the Rieti–Amatrice segment. Using spline interpolation, the mean velocity difference between south and north is reported as $h=H/f$ 7 mm/y; the West–East component is $h=H/f$ 8 mm/y. With a recurrence time $h=H/f$ 9 years, this gives $w=W/f$ 0 cm. Sentinel-1 InSAR for the August 24, 2016 Amatrice event shows a coseismic West–East displacement field with peak displacements of about 6 cm westward near Amatrice, matching the GNSS-based estimate. The paper interprets this as evidence that GNSS captured long-term strain accumulation and SAR captured its coseismic release.

Control transects at Camerino, Raiano, Mantova, Finale Emilia, and Imola show the expected regional tectonic regimes—extension in the Central Apennines and shortening toward the Po Valley—but not a spatially localized acceleration peak comparable to the Rieti–Amatrice segment. On that basis, the paper argues that combined CN + GNSS + SAR analysis can highlight smaller critical seismogenic areas within broad CN-alarmed regions and significantly reduce the size of the areas to which prevention and mitigation efforts should be prioritized. A recurrent misunderstanding is to equate this with deterministic short-term prediction; the framework instead combines intermediate-term regional alarms with geodetic localization of strain accumulation.

6. Comparative significance of the three SISMAs

The three SISMAs differ not only by field but by epistemic role. The face-synthesis SISMA is a trainable generative architecture for conditional image formation in latent space (Botti et al., 22 Sep 2025). The stellar SISMA is a curated observational archive whose value lies in homogeneous reduction, derived indicators, and interoperability (Rainer et al., 2016). The seismic SISMA is a prototype operational framework for combining heterogeneous evidence streams—seismicity patterns, GNSS, InSAR, and scenario earthquakes—in time-dependent hazard assessment (Panza et al., 2017).

These differences matter because the acronym can mislead cross-disciplinary readers. In machine learning, SISMA denotes a Mamba-based substitute for attention-heavy semantic image synthesis. In astrophysics, it denotes a long-term archive supporting joint photometric–spectroscopic studies of oscillations, rotation, activity, and circumstellar phenomena. In seismology, it denotes a system-level integration strategy intended to refine broad regional alarms into narrower priority zones for mitigation. The only shared feature is an emphasis on structured conditioning: semantic masks in image synthesis, spectroscopic and photometric metadata in stellar variability studies, and tectonically informed geodetic constraints within seismological alarms. This suggests that the acronym has repeatedly been adopted for systems designed to make high-dimensional data more operationally usable, but the underlying methods, objectives, and validation regimes remain entirely domain-specific.