STELLA: Multi-Domain Scientific Systems

• STELLA is a multi-domain designation encompassing advanced physical instrumentation, computational models, and secure analysis tools across fields like nuclear astrophysics, spectroscopy, and plasma physics.
• It integrates precise experimental designs and numerical methods to achieve high sensitivity, improved timing, and scalability, as demonstrated in sub-barrier fusion measurements and supernova light curve modeling.
• The frameworks leverage domain-specific languages and optimized architectures to ensure reproducibility, modular extensibility, and efficient simulation across diverse scientific and technical domains.

STELLA is a designation for a wide variety of scientific instruments, computational frameworks, datasets, model architectures, and algorithms across physical sciences, computational modeling, information retrieval, and machine learning. This article provides a rigorous survey of key STELLA-related systems, their foundational principles, mathematical formulations, implementation specifics, representative applications, and benchmarked effects on their respective scientific and technical domains.

1. Experimental Nuclear Astrophysics: The STELLA Particle-Gamma Coincidence Station

The STELLA (STELlar LAboratory) station at the Institut de Physique Nucléaire, Orsay, is designed for direct measurement of heavy-ion fusion cross sections in light systems (e.g., ¹²C+¹²C) down to regimes of tens of picobarn, critical for constraining astrophysical reaction rates under stellar conditions (Heine et al., 2018).

• Physics Motivation: ¹²C+¹²C and ¹²C+¹⁶O fusion processes govern carbon burning in massive stars; their rates at the relevant Gamow window (E_c.m. ≈ 1–3 MeV) are undetermined at the required cross section sensitivities (σ ≲ 1 nb).
• Detection Architecture:
  • Rotating target wheel (3 rotating + 7 stationary slots, φ=6.3 cm, 1000 rpm) sustains ≥10 μA beam on 20–50 μg/cm² foils, where spot diameters of 2 mm yield T_max ≈ 920°C for 2 W deposited power, below damage thresholds.
  • Charged particle array: three annular DSSSDs (2 upstream, 1 downstream), ΔE/E ≈ 0.5% at 5.154 MeV, covering 30% of 4π.
  • Gamma-ray detection: 36 × 1.5″×2″ LaBr₃(Ce) detectors (UK FATIMA), ε_sing ≈ 3.5% (1 MeV), 3% FWHM @ 1333 keV, 23% 4π coverage.
  • Cryopumped vacuum at 10⁻⁸ mbar prevents carbon build-up during weeks of continuous running.
• DAQ and Timing:
  • Gamma: CAEN V1751 (1 GHz, 10-bit, 1 ns timestamping).
  • Particle: ABACO μTCA FMC112 (125 MHz, 8 ns timestamping).
  • System-wide synchronization to a common 10 MHz GLIB clock; drift < ±2 ns.
• Coincidence Methodology:
  • Fusion cross section derived from the rate:

  $$\sigma(E) = \frac{R_{\mathrm{coinc}}}{\Phi N_t \epsilon_\gamma \epsilon_{\mathrm{ch}}}$$

  where R_coinc is observed coincidence rate, Φ is beam flux, N_t number of target nuclei, εγ and ε_ch are detection efficiencies. - Coincidence timing (Δt = tγ − t_ch) and energy discrimination resolves proton, α, and γ–channel events with negligible contamination.

• Performance:

  • Sensitivity reaches σ ~ 10 pb in weeks, an order of magnitude beyond prior setups (e.g., Gammasphere at Argonne, ~1 nb/week).
  • Timing: LaBr₃ intrinsic jitter ≲ 0.5 ns; system γ–ch < 3 ns.
  • Massively suppressed backgrounds: >10⁴ reduction for proton singles.
  • Integrity of diamond foils confirmed for cumulative charges ~100 C; beam and detection parameters stable across multi-week runs.

STELLA thus establishes a new regime for direct sub-barrier fusion cross-section measurements in nuclear astrophysics, through a combination of high current, beam rotation, advanced calorimetry, and nanosecond-resolved coincidence techniques.

2. Astrophysical Radiation Hydrodynamics: The STELLA Supernova Code

The STELLA code is a one-dimensional, multi-group radiation-hydrodynamics solver used for modeling the light curves of core-collapse and thermonuclear supernovae (Tsang et al., 2020, Kozyreva et al., 2020). Its architecture and physics approximations strike a balance between efficiency and physical fidelity.

• Equations:
  • Lagrangian hydrodynamics for density ρ, velocity u, coupled to angular moment equations for the frequency-dependent radiation field: energy density $E_\nu$, flux $F_\nu$, pressure tensor $P_\nu$.
  • Radiative transfer governed by moments:

  $$\begin{aligned} & \frac{\partial E}{\partial t} + \frac{1}{r^2} \frac{\partial (r^2 F)}{\partial r} = \rho \kappa (4\pi B - cE) + Q, \ & \frac{\partial F}{\partial t} + \frac{c^2}{r^2}\frac{\partial (r^2P)}{\partial r} + \frac{c^2}{r}(E-P) = -\rho (\kappa+\sigma) c F, \end{aligned}$$

  with opacities κ, σ and Planck function B. - Frequency is discretized (N_freq ≈ 40 logarithmic bins, 1–50,000 Å). - Closure via variable Eddington factor $f = P/E$ from short characteristics across each zone.

• Opacity Treatment:

  • Line opacity is handled using a Sobolev–LTE approach.
  • Thermalisation parameter ε interpolates between absorption and scattering:

  $S_\nu = (1-\epsilon) J_\nu + \epsilon B_\nu$

  where ε ≈ 0.8–1.0 yields empirical bolometric light curves and color evolution in agreement with Monte Carlo codes (ARTIS, Sedona) and real SNe. The recommended default is ε=0.9.

• Numerical Scheme:

  • Opacities from coarse precomputed tables (14×14 in $(\log \rho, \log T)$, with spatial interpolation; denser grids (56×56) remove artifacts without significant light curve change (<3%).
  • Implicit time advancement allows large time steps, with typical grids of ~400 mass zones.
  • Bound–bound transitions and gamma-ray deposition handled by approximate (gray) schemes.
  • Run time: days on multicore CPUs for a full II-P light curve; error in plateau observables < 5% versus Sedona.
• Model-Observation Agreement:
  • Plateau-phase bolometric light curves for Type II-P SNe agree with Sedona within 5%, with late-time tails diverging by ~3% owing to gamma-ray transport schemes.
  • STELLA reproduces empirical bolometric and multiband light curves, plateau durations, and photospheric evolution in SNe II-P/L/Ia with high accuracy.

This structured, versatile radiation-hydrodynamic platform allows rapid, parameter-ready surveys in stellar explosion modeling, and provides practical guidelines for the optimal use of microphysical opacities and radiative parameters (Tsang et al., 2020, Kozyreva et al., 2020).

3. High-Resolution Spectroscopy: The STELLA Observatory and SES Instrument Suite

The STELLA observatory at Izaña (Tenerife) operates a 1.2 m robotic telescope equipped with fibre-fed Echelle spectrographs for high-cadence stellar spectroscopy (Strassmeier et al., 2020, Weber et al., 2020). The SES system is being modernized with specialized instruments.

• SES System:
  • Current Echelle covers 390–870 nm at R≈55,000 (Δλ ~ 0.12 Å at 650 nm), S/N 80–700 per pixel, and order-wise radial velocity stability ≈6 m/s, overall rms ≈30 m/s.
  • Future split:
  • UV (SES-H&K): 380–470 nm, R4 grating, R≳55,000, <5 m fibre, optimized for Ca H&K.
  • Visual (SES-VIS): 470–690 nm, vacuum-stabilized, Fabry–Pérot calibration, m/s-class RV precision.
  • NIR (SES-NIR): 690–1050 nm, deep-depletion CCD, R≈55,000.
• Science and Astrophysical Parameterization:
  • Cross-correlation with synthetic templates for radial velocity (order-by-order weighted mean), precision as fine as 30 m/s.
  • Full atmospheric analysis (Teff, log g, [M/H], vsini) via χ² minimization with ATLAS-9 models.
• Empirical Results:
  • Over 10 years, >9000 SES spectra for Capella yield period, eccentricity, and mass measurements at <0.3% precision.
  • Long-term pulsational monitoring, SB1/SB2 orbital solutions, and spot-induced RV signals for diverse targets.
• Instrumental Advances:
  • Dedicated UV optimization yields ×3 higher S/N at 390 nm.
  • Visual band RV stability improves from 30–40 m/s to ≤5 m/s.
  • Real-time calibration through simultaneous Fabry–Pérot, vacuum and thermal stabilization.

The STELLA + SES platform enables precision astrophysical observatory science from a fully robotic facility, covering stellar evolution, exoplanet searches, and time-resolved variable star phenomena (Weber et al., 2020, Strassmeier et al., 2020).

4. Domain-Specific Frameworks: STELLA in Sparse Taint Analysis and Scientific DSLs

STELLA also refers to two advanced computational infrastructures in security and high-performance computing with distinct implementations.

4.1. Sparse Taint Analysis for Enclave Leakage Detection

The STELLA (Sparse Taint Analysis) framework statically analyzes Intel SGX enclave code by value-flow propagation over the LLVM IR, identifying implicit pointer-based leakage paths between secure and insecure memory (ECALL/OCALL [in]/[out] interfaces) (Chen et al., 2022).

• Program Representation: Value-flow graph (VFG) N,E and call-graph (CG) for interprocedural analysis.
• Leak Patterns: Five pointer-misuse patterns (P1–P5) involving ECALL out/user_check/OCALL in/return/unchecked NULL pointers are explicitly targeted.
• Sparse Taint Algorithm:
  • Forward: BFS taint propagation over the VFG using formal rules for assignment, store/load, pointer arithmetic (GEP), bitcasting.
  • Backward: Trace each identified taint sink to its sensitive data allocation, pruning at encrypt/seal APIs to avoid false alarms.
• Empirical Effectiveness: 78 previously unknown leaks in 13 open-source projects (up to 527 kLoC), analysis time <10 min/project, confirmed by domain experts (CVE assignment for TaLoS P2 leak). The method is robust to value-flow complexity, but limited on path sensitivity and exotic pointer manipulation.

4.2. The STELLA Stencil Language for HPC in PDE Solvers

The STEncil Loop LAnguage (STELLA) is a domain-specific embedded language (DSL) for stencil-based computation in C++ (Arteaga et al., 2014). Its primary use-case is implementing time–space parallel PDE integrators:

• Core Abstraction: Multi-dimensional Field objects and Stencil functors encode pointwise computations; backends for OpenMP (CPU) or CUDA (GPU) are interchangeable.
• Time-parallelism with Parareal: MPI-layered time-slicing yields strong scaling up to 128 time slices; performance follows theoretical Parareal bounds (efficiency E ≳ 32% up to N_p = 16), energy-to-solution scales as predicted.
• Practitioner Recommendations: Select the cheapest coarse propagator and match the temporal and spatial discretizations for optimal resource use. The approach is best suited to structured-grid, memory-resident problems.

5. Computational Plasma Physics: δf-Gyrokinetic stella

The δf-gyrokinetic code stella models turbulence-driven heat transport in magnetically confined plasma, with an emphasis on density gradient effects across stellarators and tokamaks (Thienpondt et al., 2024).

• Mathematical System:
  • Evolves non-adiabatic part of the perturbed distribution function $g_s$ in (x, y, z, v_∥, μ, t).
  • Central equation (normalized):

  $$\partial_t g_{s,k} + v_{∥} b \cdot \nabla_z g_{s,k} + i\omega_{*n,s}(k_y) J_0 \phi_k F_{0,s} + \ldots = \hat{C}[g_{s,k}]$$

  with electrostatic potential, FLR effects (through Bessel function J₀), nonlinear E×B brackets, and (optional) linearized collision operator.

• Physics Regimes:

  • Microinstability identification via velocity-space structure of $h_e$: strong cone confinement ↔ TEM dominance; isotropy ↔ passing electron universal instability.
• Key Findings:
  • In stellarators (NCSX, W7-X), increasing density gradient suppresses ITG modes (ion heat flux drops an order of magnitude for $0.5 \leq a/L_n \leq 2.5$), unlike in tokamaks, thus offering a prospective path for turbulence mitigation via active profile control.
  • Trapped-electron-modes dominate in most devices except W7-X, where universal instability also plays a significant role.
• Modeling Parameters: 3D VMEC geometries, fully kinetic ions/electrons, β ≈ 0 (electrostatic), grid resolutions up to (N_x,N_y,N_z,N_{v∥},N_{μ}) = (61,61,XX,48,12).

6. Large-Scale ML: STELLA in Benchmark Construction and Model Distillation

6.1. Aerospace IR: The STELLA Benchmark

The STELLA framework, designed for the aerospace domain, constructs a publicly released information retrieval benchmark from NASA NTRS documents (Kim, 7 Jan 2026).

• Pipeline Steps:

1. Document layout detection via DocLayout-YOLO
2. Passage chunking with recursive token chunker
3. Construction of a domain-specific terminology dictionary
4. Candidate passage selection and intent classification using GPT-5 (five intent categories)
5. Synthetic query generation: Dual query type—Terminology Concordant (TCQ, requiring term presence) vs Terminology Agnostic (TAQ, term ban, using paraphrased definitions); queries generated with a Chain-of-Density/self-reflection loop for controllable lexical/semantic contrast
6. Cross-lingual hybrid extension, preserving English technical terms in non-English queries in the TCQ format

• Evaluation: Models' nDCG@10 reported for seven IR models; results show large decoder-based embedding models provide strongest semantic retrieval, but BM25 remains very strong for lexical queries.

6.2. SOTA Embedding Distillation: Stella and Jasper

Stella_en_1.5B_v5 is a 1.5B-parameter text Transformer encoder, released as a strong "teacher" model for dense retrieval (MTEB score 71.19/100 across 56 datasets) (Zhang et al., 2024). In a multi-step distillation pipeline, its pooled 4096-dim embeddings serve as an unmodified supervision signal for student models (e.g., Jasper, a 2B-parameter multimodal embedding).

• Distillation Losses: Cosine distance, similarity matrix MSE, and triplet loss (mined from teacher combinations) over concatenated teacher vectors (Stella + NV-Embed, up to 12,288-dim target), followed by MRL-inspired compression to downstream embedding sizes (e.g., 512).
• Empirical Insight: Stella embeddings, while not achieving top rank in isolation (surpassed slightly by NV-Embed), offer complementary signals in multi-teacher pipelines, underlining the importance of diverse teacher architectures in state-of-the-art distillation.

7. Theoretical Programming Languages: Stella Educational Language

Stella is an educational statically typed programming language with a minimal simply-typed functional core and a modular set of classical extensions (algebraic data types, references, exceptions, subtyping, recursive types, universal polymorphism, Hindley–Milner reconstruction) (Abounegm et al., 2024).

• Core Syntax and Typing:
  • Core: PCF-style calculus with booleans, naturals, λ-abstraction, application, and conditionals.
  • Extensions: Enabled via top-level pragma (extend with #feature1, #feature2), each with formal syntax and inference rules (e.g., products, sums, references, exceptions, structural subtyping, recursion, System F, Algorithm W for type inference).
  • Pedagogical BNF grammar and classical inference-style typing rules provided for all constructs; sample code and derivations for key language idioms facilitate incremental learning and implementation.

Stella is structured to make teaching and implementing increasing type system complexity tractable, providing explicit modularity for students to master typechecking and inference for each language feature.

---

STELLA, in all its instantiations, represents the convergence of precise experimental design, mathematical rigor in computational simulation, principled domain-specific language architecture, and advanced machine learning evaluation. Across nuclear astrophysics, radiation transport, observational astrophysics, secure computation, HPC, plasma physics, retrieval benchmarks, embedding learning, and programming language instruction, STELLA frameworks consistently emphasize reproducibility, modular extensibility, and quantitative performance assessment.