---
title: 'AION-1: Cross-Domain Research Overview'
url: https://www.emergentmind.com/topics/aion-1
type: topic
---

# AION-1: Cross-Domain Research Overview

AION-1 is a research designation used for several distinct systems across atom interferometry, multimodal astronomical machine learning, Android malware detection, and online database-isolation checking. In the gravitational-wave literature, it most prominently denotes an atom-interferometer stage within the AION programme, intended to probe the mid-frequency band between LIGO/Virgo and LISA [2003.13480]. In astronomy, it denotes an omnimodal foundation-model family pretrained across large sky surveys [2510.17960]. In cybersecurity and database systems, it denotes unrelated architectures for active-learning-based Android repackaged-malware detection and online snapshot-isolation checking, respectively [1808.01186] [2504.01477]. The term therefore has no single cross-domain referent; its meaning is determined by the surrounding research context.

## 1. Nomenclature and domain-specific referents

The label is used heterogeneously in the literature.

| Research domain | Meaning of AION-1 | Source |
|---|---|---|
| Atom interferometry | A stage of the AION gravitational-wave programme | [2003.13480] |
| Astronomy ML | Omnimodal foundation model for astronomical sciences | [2510.17960] |
| Android security | Active-learning malware-detection architecture | [1808.01186] |
| Database verification | Online snapshot-isolation checker | [2504.01477] |

A further related but not identical usage appears in optimization theory, where AION denotes the “Analytic, Infinitely-Optimisable Network” architecture introduced together with the Method of Infinite Descent; that paper does not use the AION-1 designation [2510.05489].

This multiplicity is not merely nominal. Different papers attach the same string to objects with different physical scales, computational goals, and mathematical formalisms. Within atom interferometry alone, the label is not uniform across publications. This suggests that any technical reading of “AION-1” must begin with source-specific disambiguation.

## 2. AION-1 in atom-interferometric gravitational-wave detection

In the AION programme described in “Probes of Gravitational Waves with Atom Interferometers,” AION-1 is the first 10 m–scale terrestrial demonstrator, designed to bridge the mid-frequency gap between LIGO/Virgo at frequencies $\gtrsim 10\,\mathrm{Hz}$ and LISA at frequencies $\lesssim 0.1\,\mathrm{Hz}$ [2003.13480]. The instrument is a vertical atomic fountain with baseline length $L=10\,\mathrm{m}$, using ultracold $^{87}\mathrm{Sr}$ selected for the narrow ${}^1S_0$–${}^3P_0$ clock transition at $\lambda=698\,\mathrm{nm}$. Its laser system employs two counter-propagating ultra-stable $698\,\mathrm{nm}$ beams frequency-stabilized by a high-finesse cavity, and the interferometric geometry is a Mach–Zehnder light-pulse sequence $\pi/2-\pi-\pi/2$. Large-momentum-transfer techniques raise the effective momentum transfer to $O(100\times 2\pi/\lambda)$, while the pulse separation is chosen as $T=O(0.3\,\mathrm{s})$, limited by the free-fall height in the 10 m tube [2003.13480].

The basic gravitational-wave response is written in the frequency domain as
$$
\Delta\phi(f)=k_{\mathrm{eff}}\,L\,h(f)\,T(f),
$$
with transfer function
$$
T(f)=\sin^2(\pi fT)e^{i\pi fT}.
$$
An equivalent time-domain expression is
$$
\Delta\phi(f)=k_{\mathrm{eff}}\,L\int dt\,h(t)\,W(t;T),
$$
where $W(t;T)$ is the response kernel of the $\pi/2$–$\pi$–$\pi/2$ sequence [2003.13480]. The dominant noise is assumed to be atom-shot-noise-limited, with phase uncertainty $\delta\phi_n\approx 10^{-3}\,\mathrm{rad}/\sqrt{\mathrm{Hz}}$ for $N_{\mathrm{atoms}}\approx 10^8$, yielding
$$
h_n(f)=\delta\phi_n/[k_{\mathrm{eff}}L|T(f)|].
$$
With $k_{\mathrm{eff}}\approx 100\times(2\pi/698\,\mathrm{nm})\approx 9\times 10^7\,\mathrm{m}^{-1}$, $L=10\,\mathrm{m}$, and $T=0.3\,\mathrm{s}$, the quoted single-sided strain-noise amplitudes are $3\times 10^{-19}$ at $0.1\,\mathrm{Hz}$, $1\times 10^{-19}$ at $0.3\,\mathrm{Hz}$, $5\times 10^{-20}$ at $1\,\mathrm{Hz}$, $1\times 10^{-19}$ at $3\,\mathrm{Hz}$, and $5\times 10^{-19}$ at $10\,\mathrm{Hz}$, all in units of $1/\sqrt{\mathrm{Hz}}$ [2003.13480]. In characteristic-strain form, the mid-band minimum is near $f\approx 1\,\mathrm{Hz}$ with $h_c\approx 5\times 10^{-20}$.

The same study treats AION-1 as a probe of modified gravitational-wave propagation. For a massive graviton with dispersion relation $E^2=p^2+m_g^2$, a Fisher-matrix forecast for a GW150914-like binary with redshifted chirp mass $\mathcal{M}_z\approx 30\,M_\odot$ at luminosity distance $D_L=420\,\mathrm{Mpc}$, observed over $\sim 1$ day, yields a $90\%$ confidence-level bound
$$
m_g<2\times 10^{-22}\,\mathrm{eV}.
$$
For Lorentz-violating modified dispersion relations of the form $E^2=p^2+A\,p^\alpha$, AION-1 improves on LIGO by factors $10$–$100$ in the regime $\alpha\le 1$, excluding $|A|$ down to $\sim 10^{-35}\,\mathrm{GeV}^{2-\alpha}$ for $\alpha=1/2$ or $1$ [2003.13480].

The paper also positions AION-1 within a staged roadmap. Scaling from $L=10\,\mathrm{m}$ to AION-100 m reduces $h_n(f)$ by $10\times$ and tightens graviton-mass bounds to $\sim 2\times 10^{-23}\,\mathrm{eV}$, while the full AION-1 km stage reaches $h_n(1\,\mathrm{Hz})\sim 5\times 10^{-21}/\sqrt{\mathrm{Hz}}$ and $m_g<1\times 10^{-24}\,\mathrm{eV}$ at $90\%$ confidence level, exceeding LIGO’s current limit by $\sim 40\times$ in a stand-alone inspiral measurement [2003.13480].

## 3. Nonuniform atom-interferometer usage of the label

Later papers use the same designation for related but not identical atom-interferometric concepts. In “Testing Supersymmetric Hidden Sectors with Long-Baseline Atom Interferometers,” AION-1 is treated as a laboratory-scale probe of ultralight moduli, dilatons, or hidden scalars with masses in the range $10^{-17}$–$10^{-10}\,\mathrm{eV}$, detectable through coherent oscillations in atomic energy levels [2606.09943]. The microscopic coupling dictionary is expressed through
$$
g_{\mathrm{eff}}=K_\alpha g_e+K_{m_e}g_{m_e}+K_g g_g+\dots,
$$
and the interferometric phase shift is approximated by
$$
\Delta\Phi_\phi \simeq \omega_A T_{\mathrm{eff}} g_{\mathrm{eff}}\phi_0\cos(m_\phi t+\beta),
$$
with amplitude
$$
A_\phi=\omega_A T_{\mathrm{eff}} g_{\mathrm{eff}}\sqrt{2F_{\mathrm{DM}}\rho_{\mathrm{DM}}}/m_\phi
$$
if the scalar contributes a fraction $F_{\mathrm{DM}}$ of the local dark-matter density [2606.09943]. The projected oscillatory-phase sensitivity is stated as $A_{\mathrm{lim}}\sim 10^{-8}\,\mathrm{rad}$ per Fourier bin in the $m_\phi/2\pi\simeq 0.01$–$1\,\mathrm{Hz}$ band, leading to the benchmark scaling
$$
\epsilon \lesssim 5.3\times 10^{-9}
\times (A_{\mathrm{lim}}/10^{-8}\,\mathrm{rad})
\times (m_\phi/10^{-15}\,\mathrm{eV})
\times [1/(d_{\mathrm{eff}}^{(0)}\sqrt{F_{\mathrm{DM}}})]
\times (1\,\mathrm{s}/T_{\mathrm{eff}}),
$$
with examples $\epsilon\lesssim 5\times 10^{-9}$ at $m_\phi=10^{-15}\,\mathrm{eV}$ and $\epsilon\lesssim 5\times 10^{-11}$ at $m_\phi=10^{-17}\,\mathrm{eV}$ [2606.09943]. The assumed “AION-1-quality instrument” in that projection has baseline $L\approx 100\,\mathrm{m}$, a strontium clock atom with $\nu_A\approx 4.3\times 10^{14}\,\mathrm{Hz}$, $k_{\mathrm{eff}}\sim 10^{10}\,\mathrm{m}^{-1}$, pulse spacing $T\approx 1\,\mathrm{s}$, and total run-time $O(\mathrm{year})$.

A different usage appears in “Probing supermassive black hole seed scenarios with gravitational wave measurements,” where AION-1, also referred to as AION-1 km, is described as a space-borne atom-interferometer concept with a $1\,\mathrm{km}$ baseline, operating with ultracold strontium atoms in two drag-free spacecraft separated by $L=1\,\mathrm{km}$ [2312.02983]. In that treatment, the detector spans roughly $0.01\,\mathrm{Hz}$ to a few hertz, with a low-noise-model minimum $\sqrt{S_n}\sim 10^{-22}\,\mathrm{Hz}^{-1/2}$ at $f\sim 0.3\,\mathrm{Hz}$. The inspiral matched-filtering analysis uses
$$
\mathrm{SNR}^2 = 4\int_0^\infty \frac{|\tilde h(f)|^2}{S_n(f)}\,df,
$$
and gives a detection horizon out to redshifts $z\sim 3$–$7$ for binaries with chirp masses $\mathcal{M}_c\sim 10^2$–$10^4\,M_\odot$ in the last minutes before merger [2312.02983]. One-year event yields are quoted as $\mathcal{O}(10)$ in low-mass-seed scenarios and below unity in high-mass-seed scenarios, while Fisher-matrix estimates give $\Delta m_i/m_i\sim 10$–$30\%$ and $\Delta z/z\sim 20$–$50\%$ for a typical $(500+300)\,M_\odot$ binary at $z\lesssim 1$, improving to a few percent for the heaviest events seen into merger [2312.02983].

These descriptions are not numerically interchangeable. One paper defines AION-1 as a 10 m terrestrial demonstrator, another uses AION-1-quality assumptions at approximately 100 m, and a third uses AION-1 or AION-1 km for a $1\,\mathrm{km}$ space-borne concept [2003.13480] [2606.09943] [2312.02983]. A plausible implication is that the label functions as a stage or concept marker rather than a universally fixed instrument specification.

## 4. AION-1 as an omnimodal astronomical foundation model

In astronomy, AION-1 denotes a family of multimodal foundation models intended for joint modeling of heterogeneous imaging, spectroscopic, and scalar data [2510.17960]. The architecture has two stages. Stage 1 comprises modality-specific tokenizers: an imaging tokenizer built from a ResNet autoencoder with FSQ quantization, a spectra tokenizer using ConvNeXt-V2 with LFQ quantization, and a parameter-free FSQ scalar tokenizer based on empirical CDF estimation, Gaussianization, and inverse mapping through standard-normal quantiles [2510.17960]. Stage 2 is a multimodal encoder–decoder transformer that consumes sequences of discrete tokens from any combination of modalities, with embeddings of the form
$$
e_t^{(\mathrm{enc})}=\mathrm{Embed}_i(\mathrm{token})+m_i+p_t,
$$
where $m_i$ is a learnable modality-plus-instrument embedding and $p_t$ a shared positional encoding [2510.17960].

The pretraining corpus spans five surveys: Legacy Survey, Hyper Suprime-Cam, Sloan Digital Sky Survey, Dark Energy Spectroscopic Instrument, and Gaia. These are stated to cover more than 200 million observations of stars, galaxies, and quasars [2510.17960]. Optimization uses AdamW with $(\beta_1,\beta_2)=(0.9,0.95)$, weight decay $0.05$, global batch size $8096$, linear warmup to $2\times 10^{-4}$ followed by cosine decay over $205\,\mathrm{k}$ steps, mixed precision in bfloat16, and FSDP ZeRO-2 [2510.17960]. Three transformer variants are reported: AION-B with 300 M parameters, AION-L with 800 M parameters, and AION-XL with 3.1 B parameters [2510.17960].

The model is evaluated with a frozen encoder and lightweight downstream heads. In galaxy property estimation on PROVABGS targets, AION-B improves from $R^2(z)=0.75$ and $R^2(M_\star)=0.72$ using photometry alone to $R^2(z)=0.93$ and $R^2(M_\star)=0.89$ with photometry plus imaging, and to $R^2(z)=1.00$ and $R^2(M_\star)=0.96$ with photometry, imaging, and spectra [2510.17960]. In Galaxy Zoo 10 morphology classification, AION-L reaches $87.2\%$ accuracy, compared with $80.0\%$ for EfficientNet-B3 from scratch and $71.4\%$ for DINOv2, while ZooBot trained on $300\,\mathrm{k}$ labels achieves $89.6\%$ [2510.17960]. In Galaxy Zoo 3D segmentation, AION-B attains IoU $0.60$ for spiral arms and $0.31$ for bars, compared with $0.55$ and $0.28$ for a U-Net trained from scratch [2510.17960]. For rare-object retrieval, AION-XL reaches nDCG@10 values of $0.62$ for spirals, $0.38$ for mergers, and $0.015$ for strong lenses, exceeding the AstroCLIP results of $0.60$, $0.25$, and $0.006$ [2510.17960].

The low-data behavior is emphasized as well. Across galaxy-property, morphology, and stellar-parameter tasks, AION-B reaches approximately $R^2\sim 0.5$ or accuracy $\ge 80\%$ with only $10^2$–$10^3$ labels, whereas supervised baselines remain near zero until approximately $10^4$–$10^5$ labels [2510.17960]. The release includes code, tokenizers, pretrained weights for the 300 M, 800 M, and 3 B models, and a lightweight evaluation suite under an MIT license [2510.17960].

## 5. Uncertainty quantification on frozen AION-1 embeddings

A subsequent study examines uncertainty quantification for galaxy-property regression using frozen AION-1 embeddings [2606.07771]. In that setup, the pretrained encoder produces a pooled $768$-dimensional representation
$$
\phi(x)\in\mathbb{R}^{768},
$$
which is passed to a separate three-layer MLP for each target: redshift $z$, stellar mass $\log M_\star$, mass-weighted stellar age $t_{\rm age}$, gas-phase metallicity $\log Z$, and specific star-formation rate $\log \mathrm{sSFR}$ [2606.07771]. The MLP architecture is $768\rightarrow 256\rightarrow 256\rightarrow 1$ with GELU nonlinearities and MSE training under Adam with learning rate $10^{-3}$, weight decay $10^{-5}$, and cosine annealing over $500$ epochs. On a held-out test set of $11\,362$ galaxies, the point-estimate $R^2$ scores are
$$
\{0.98,\;0.89,\;0.85,\;0.63,\;0.86\}
$$
for $\{z,\log M_\star,t_{\rm age},\log Z,\log \mathrm{sSFR}\}$ [2606.07771].

The UQ comparison includes Deep Ensembles, MC Dropout, VanillaSplit conformal prediction, MADSplit, Conformalized Quantile Regression, and Locally Valid and Discriminative variants using either MLP hidden activations or raw AION-1 embeddings [2606.07771]. At nominal level $1-\alpha=0.90$, all conformal methods achieve marginal coverage within approximately $1$ percentage point of $90\%$ across the five targets, while the non-conformal baselines do not calibrate reliably: Deep Ensembles under-cover severely, and MC Dropout over-covers [2606.07771]. In the worst-predicted $10\%$ bin, CQR gives the highest coverage, approximately $44$–$66\%$, whereas VanillaSplit collapses to approximately $0\%$ and MADSplit is inconsistent at approximately $7$–$41\%$ [2606.07771]. Among well-calibrated methods, the LVD variants yield the tightest intervals.

The principal methodological claim concerns local validity. The LVD framework defines weights
$$
w(x,x_i)=\exp\!\left(-\|g(\phi(x))-g(\phi(x_i))\|^2\right)
$$
and constructs a weighted local conformal quantile $q_{\rm LVD}(x)$, producing intervals that adapt to local prediction difficulty [2606.07771]. The paper states that only LVD—particularly when operating on AION-1 embeddings—provides finite-sample local validity rather than only marginal coverage. This makes the AION-1 embedding space part of the uncertainty model itself, not merely a fixed feature extractor.

## 6. AION-1 in cybersecurity and online database checking

In Android malware research, Aion or AION-1 denotes an architecture that couples app stimulation and malware detection through active learning [1808.01186]. Its workflow contains five core modules in a feedback loop: App Stimulation Engine, Monitoring and Trace Collection, Behavior Representation, Classifier and Active Learner, and Feedback Manager. APKs are installed on an Android Virtual Device and stimulated for $60\,\mathrm{s}$ with Droidutan, a UI exerciser that emulates taps, text input, and manifest-declared broadcast-intent injection [1808.01186]. Static features extracted through Androguard include metadata, permissions, and counts in 27 sensitive API categories, forming a vector of approximately $37$ dimensions. Dynamic features from droidmon intercept 71 key methods and are collapsed into a 37-dimensional vector. The hybrid representation is
$$
x=[x^{(\mathrm{stat})}\|\;x^{(\mathrm{dyn})}]\in\mathbb{R}^{74}.
$$
The active-learning loop retrains on $\mathcal{D}_t$, identifies the misclassified query set
$$
Q_t=\{x_i\in\mathcal{D}_t\mid f_t(x_i)\neq y_i\},
$$
re-stimulates those apps, and iterates until $|\mathrm{F1}_t-\mathrm{F1}_{t-1}|\le 1\%$ or $t\ge 10$ [1808.01186]. The ensemble combines KNN, Random Forest, and linear SVM classifiers by majority vote. On the Piggybacking dataset, the best reported test result is approximately $\mathrm{F1}\approx 0.72$ with specificity approximately $0.70$ using Random Forest with 100 trees and hybrid features, after up to a $+7\%$ F1 lift relative to single-run dynamics and convergence at iteration approximately $5$ on average [1808.01186].

In database systems, AION is an online checker for snapshot isolation, derived from the offline checker CHRONOS [2504.01477]. It is designed for streaming histories with out-of-order timestamps, bounded memory, and no need for fence transactions or database pauses. The core versioned structures are `frontier[$\tau$]`, which stores the last committed value of each key at or before timestamp $\tau$, and `ongoing[$\tau$][k]`, which tracks transactions whose start has occurred by $\tau$, whose commit is after $\tau$, and which write key $k$ [2504.01477]. The online algorithm processes a new transaction in three phases: initial checks on session order, start–commit order, and internal/external read consistency; re-checking write–write conflicts for overlapping transactions; and re-checking external reads of later transactions whose snapshots might have included the new commit [2504.01477]. Its axioms include
$$
T.start\le T.commit,\qquad
T_1 <_{\mathrm{vis}} T_2 \iff T_1.commit\le T_2.start,
$$
together with session-order, internal-consistency, external-consistency, and no-write-write-conflict conditions [2504.01477]. The amortized per-transaction complexity is stated as
$$
O(\log N + M + K),
$$
where $N$ is the number of transactions seen so far, $M$ the number of operations in the new transaction, and $K$ the number of events in the relevant timestamp interval [2504.01477]. Experimentally, AION and AION-SER sustain approximately $12\,\mathrm{K}$ transactions per second; CDC extraction imposes only approximately $5\%$ overhead on TiDB and YugabyteDB; and a $5\,\mathrm{s}$ timeout virtually eliminates spurious violation reports caused by temporary out-of-order arrival effects [2504.01477].

These two systems share neither domain nor method. One is an active-learning malware-analysis pipeline over behavioral traces; the other is an online consistency checker over timestamp-indexed transactional histories. Their coexistence under the same name is a case of terminological collision rather than conceptual relatedness.

## 7. Related AION nomenclature in optimization theory

A distinct but adjacent usage is the AION architecture introduced in “The Method of Infinite Descent” [2510.05489]. Here AION means “Analytic, Infinitely-Optimisable Network,” a separable model family designed to satisfy the algebraic-closure conditions required by the Infinite Descent optimizer. For input $x=(x_1,\dots,x_d)$ and rank $r$, the model represents
$$
f_r^\infty(x;\Theta)=\sum_{j=1}^r\prod_{i=1}^d \psi_i^{(j)}(x_i;\theta_i^{(j)}),
$$
with each atom given by a finite linear combination of exponential–trigonometric terms [2510.05489]. The crucial property is algebraic closure under arbitrary differentiation orders and multiplicative composition, enabling exact resummation of the parameter-shift Taylor series:
$$
\nabla\Phi(\Theta+\Delta)=\exp(\Delta\cdot \nabla_\Theta)\,[\nabla\Phi(\Theta)].
$$
The resulting update is posed as an exact nonlinear algebraic root problem in $\Delta$ rather than as an approximation through truncated gradients [2510.05489].

The demonstration problem is $g(x,y)=\cos(\pi(x-y))$ on the unit square, sampled on a $25\times 25$ grid and fitted with least squares. With rank $r=2$ and $P=1$ per atom, the paper states that the model can represent the target exactly as a sum of two separable trigonometric products [2510.05489]. Reported optimization outcomes are: Steepest Descent, 1,000 iterations to loss approximately $6\times 10^{-6}$ in $6\,123\,\mathrm{ms}$; Newton–CG, 28 iterations to loss approximately $5\times 10^{-13}$ in $383\,\mathrm{ms}$; and Infinite Descent, 1 outer iteration to loss approximately $9\times 10^{-18}$ in $102\,\mathrm{ms}$ [2510.05489]. The paper further defines the “Infinity Class” as the class of separable neural architectures whose univariate bases are algebraically closed under differentiation and multiplication, so that the Taylor expansion in parameter shifts can be resummed in closed form [2510.05489].

Although this work does not name its model AION-1, it illustrates how the AION stem has proliferated into yet another technically specific meaning. Across the corpus, “AION-1” therefore functions less as a unique identifier than as a reused label spanning atom interferometers, astronomical foundation models, cybersecurity pipelines, and online verification algorithms.

Source: https://www.emergentmind.com/topics/aion-1