Mephisto: Multi-Modal Astrophysics & ML Framework

• Mephisto is a cutting-edge ecosystem comprising a multi-channel photometric survey telescope and LLM-based multi-agent systems for automated astrophysical data analysis.
• It employs simultaneous six-band imaging, advanced ML classification, and precise photometric calibrations to achieve high-cadence time-domain astronomy.
• The framework integrates automated scheduling, all-sky environmental monitoring, and reproducible crowdsourcing, advancing both astronomical discoveries and data annotation methods.

Mephisto refers to several advanced frameworks and instruments in astronomy and computer science—most notably a multi-channel photometric survey telescope optimized for time-domain astrophysics and cosmology, a family of LLM-driven multi-agent reasoning systems for automated interpretation of galaxy data, and an open-source ecosystem for portable, reproducible crowdsourcing. The Mephisto telescope is distinguished by its capability for true, wide-field, simultaneous imaging in up to six optical bands, enabling high-cadence, real-time color photometry for a variety of transient and variable phenomena. Simultaneously, the Mephisto agent framework and its variants enable interpretable, end-to-end SED analysis and hypothesis evaluation via LLM collaboration and knowledge distillation pipelines. Collectively, these systems represent state-of-the-art methodologies for astronomical time-domain science, survey operations, machine-assisted interpretation, and reproducible annotation workflows.

1. Mephisto: Multi-Channel Photometric Survey Telescope

The Mephisto (Multi-channel Photometric Survey Telescope) is a 1.6 m-class Ritchey–Chrétien reflector situated at Lijiang Observatory. Its design features a rear-located focal plane accommodating three independent camera channels, each equipped with exchangeable filters, yielding a 2° field of view per channel and total instantaneous coverage up to 6 deg². Currently, it is the only instrument capable of fully simultaneous, wide-field imaging in six optical bands (u, v, g, r, i, z) across three cameras (Wang et al., 2024, Yang et al., 2024, Cheng et al., 2024).

Each channel uses a high-pixel-count, low-noise CCD: u/v/g/r employ e2v CCD231-C6 sensors (6144×6160, 15 μm), while i/z utilize a custom CCD290-99 device (9216×9232, 10 μm). Read noise is ∼3–5 e⁻ rms, with system quantum efficiency ≥90% at peak, and dark current ≤0.01 e⁻/pixel/s (at –100 °C).

Mephisto’s filter set is specifically optimized, with broad blue–red leverage (Δλ from 40–150 nm), matching community standards while maximizing S/N on early transients and variable sources.

Key Engineering Attributes

Channel	CCD Sensor	Pixel Scale	Field of View
u/v/g/r	e2v CCD231-C6	0.429″/pixel	≈43′ × 43′
i/z	e2v CCD290-99	0.286″/pixel	≈43′ × 43′

The three-camera setup delivers simultaneous exposures every 20–300 s, with a minimum dead time (<10 s), and supports rapid slew (≤90 s to target). Unique to Mephisto is the ability to select filter triplets (u,g,i) or (v,r,z) on the fly, accommodating time-critical science and spectral feature tracking (Cheng et al., 2024, Yang et al., 2024).

2. Survey Strategies and Science Operations

Mephisto conducts wide-field time-domain surveys, optimized for both transient discovery and precise photometric follow-up (Lei et al., 2021, Wang et al., 2024, Yang et al., 2024). Its year-one operations focused on the northern sky (∼27,000 deg², –21°<δ<75°), with key campaigns targeting the SDSS Stripe 82 region, as well as rapid-response modes for gamma-ray burst and supernova afterglow capture (Cheng et al., 2024, Lunwei et al., 2024).

Survey modes utilize pairs of 20–300 s exposures, configured for maximum cadence (1–3 d revisit) while maintaining photometric depth—critical for identifying fast-evolving phenomena. For cosmological SN Ia work, simulation-driven strategies recommend 130 s exposures on a 3 d cadence, yielding ∼900 well-sampled SNe Ia (S/N>5, ≥15 points in ≥3 bands) per 200-night season (Wang et al., 2024).

Photometric calibration combines HST spectrophotometric standards, Gaia DR3 cross-matching, and real-time zero-point monitoring, yielding absolute zero-point uncertainties ≲0.01 mag (g/r/i/z) and <0.03 mag (u/v). Robust correction for Galactic and host extinction is standard.

The data reduction pipeline performs real-time bias/dark correction, flat-fielding, astrometric alignment (≲40 mas residuals), image registration, and PSF-matched subtraction for crowded fields, guaranteeing time-coincident multiband measurements (Cheng et al., 2024, Yang et al., 2024).

3. Data-Driven Classification and Feature Extraction

Mephisto’s multiband light curves underpin a suite of machine learning and deep learning classifiers for variable-source identification and rapid typing of astrophysical transients (Lei et al., 2021, Lunwei et al., 2024).

Variable Star and Quasar Identification

The Mephisto-W survey utilized Random Forest Classifiers (RFC) for RR Lyrae and quasar recognition. Training leveraged simulated uvgriz light curves (transformed from SDSS Stripe 82 data) with ∼150 extracted features per source:

• Statistical parameters (114): amplitude, standard deviation, flux-percentage ratios, Lomb–Scargle period, linear trends, etc., computed per band.
• Color indices (15): mean inter-band differences.
• Real-time colors (6): simultaneous colors, capturing rapid SED evolution.
• DRW parameters (12, quasar only): stochastic variability timescales τ, σ.

Performance (48/91-fold CV):

Type	Purity	Completeness
RR Lyrae	95.4%	96.9%
Quasar	91.4%	90.2%

Key discriminants for RR Lyrae were g-band amplitude and scatter; for quasars, color indices and DRW variability (Lei et al., 2021).

Early-Time Supernova Classification

Mesiri is a BiLSTM-based classifier exploiting true simultaneous (u,g,i) or (v,r,z) colors. Input features per epoch include photometry, uncertainties, and instantaneous colors. With only a single epoch (pre-peak), Mesiri reaches 96.75 ± 0.79% accuracy (AUC = 98.87 ± 0.53%). After three nights, accuracy is 99.48 ± 0.24%, enabling near real-time differentiation of SN Ia from core-collapse SNe, critical for triggering timely spectroscopic follow-up (Lunwei et al., 2024).

4. Automated, LLM-Based SED Interpretation (Mephisto Agent Framework)

Mephisto’s agentic framework is a modular, LLM-centered system for automated, interpretable analysis of multi-band galaxy photometry (Sun et al., 2024, Sun et al., 9 Oct 2025). Each agent specializes in a research workflow component:

• Reasoning agent: Generates model modifications/hypotheses based on the observed data, model state, fit results, memory, and a dynamic, externalized knowledge base.
• Execution agent: Runs the CIGALE code with proposed parameters, returning χ² and fit diagnostics.
• Evaluation agent: Assesses fit quality, tracks per-band residuals, and curates a tree of hypotheses.
• Memory and knowledge agents: Distill empirical insights from past searches, validate new “rules” (e.g., modifying AGN fraction when MIR is underestimated), and update knowledge for future efficiency.

Tree search (alternating DFS/BFS) enables systematic exploration via expansion, evaluation, and back-propagation of fit success. Knowledge accumulation occurs through self-play and rigorous validation on held-out data.

Empirical results on COSMOS2020 and JWST “Little Red Dot” datasets demonstrate that Mephisto can reconstruct physical parameters (stellar mass, A_V) and arrive at near-expert hypotheses (e.g., identifying distinct AGN/non-AGN solutions for LRDs) while evaluating only ∼1% as many models as exhaustive brute-force grid search, with <0.3 dex error on mass and fit success rates exceeding 90% (Sun et al., 9 Oct 2025). Cost-performance is competitive when leveraging high-quality LLMs (GPT-4o), with open models nearing similar efficacy at reduced computational expense.

5. Standard Candle and Stellar Population Calibrations in Mephisto Filters

Red clump (RC) stars serve as reliable standard candles for Galactic structure when robustly calibrated. Mephisto enables empirical RC absolute magnitude calibration in v, g, r, i bands using large (N=25,059) APOGEE-Gaia cross-matched samples (Yu et al., 25 Jun 2025).

Absolute magnitudes are fitted as second-order polynomials in [Fe/H] and T_eff/5000 K (denoted x, y), e.g.,

$$M_v = –76.789 – 4.767x + 33.766y + 0.115x^2 + 5.191xy – 12.009y^2\quad(σ_\mathrm{RMSE}=0.26\,\mathrm{mag})$$

• The v-band shows ΔM_v ≃ 2.0 mag across [Fe/H] = –1.0 … 0.5 dex and T_eff = 4500–5200 K.
• Scatter in M_g, M_r, M_i is 0.24 mag.
• Realistic measurement error is σ_M ≃ 0.07 mag (photometric + parallax), yielding ∼5% distance precision per RC star.

A Random Forest approach, using Mephisto colors as features, provides robust photometric metallicity estimates (σ_[Fe/H] ≃ 0.12 dex for σ_phot ≤ 0.01 mag) (Yu et al., 25 Jun 2025).

6. All-Sky Monitoring and Cloud-Mask Segmentation for Automated Scheduling

Mephisto integrates a real-time all-sky camera system for environmental awareness and observation scheduling (Cui et al., 22 Oct 2025). The system employs a CNN-based enhanced U-Net (encoder: EfficientNet-B4; decoder: SCSE attention blocks) for pixel-level segmentation of unobservable regions (clouds, moonlight) on 2568×2568-pixel images.

• Compound loss combines binary cross-entropy, Dice, and IoU terms (weights 0.4, 0.2, 0.4).
• On annotated test set, IoU=0.9212, F1=0.9537.
• Masks are mapped to celestial coordinates via Zenithal Equal-Area (ZEA) projection with 5th-order polynomial distortion correction (mean residual 0.95 px).

The resultant Boolean sky-grid feeds directly into the observation control system (OCS), allowing sub-50 ms latency for cloud-aware, autonomous scheduling. The system is readily generalizable to other sites after recalibration of projection and distortion coefficients (Cui et al., 22 Oct 2025).

7. Mephisto in Crowdsourcing and Data Annotation

Distinct from the astronomy-focused context, Mephisto is also the name of an open-source framework for reproducible, portable, and collaborative crowdsourcing in machine learning research (Urbanek et al., 2023).

• Core abstractions include: Blueprint (task logic/UI), Architect (hosting backend), CrowdProvider (worker pool interface), and MephistoDB (persistent storage).
• Supports declarative, versioned task specs, unified local/remote execution, multi-run/QA workflows, and Model-in-the-Loop development.
• Enables deterministic offline QA via MockProvider and reproducible pilot studies.
• Designed for extensibility and rigorous experiment annotation, addressing fields where tool standardization still lags behind modeling ecosystems (e.g., TensorFlow, PyTorch).

The data model encompasses Task→TaskRun→Assignment→Unit→Agent/Worker, enabling detailed audit trails and real-time collaboration.

• Variable source and quasar classification: (Lei et al., 2021)
• Cosmological strategies and SN Ia calibration: (Wang et al., 2024)
• Early multiband SN photometry and calibration: (Yang et al., 2024, Lunwei et al., 2024)
• GRB afterglow follow-up and six-channel operation: (Cheng et al., 2024)
• All-sky segmentation and observation control: (Cui et al., 22 Oct 2025)
• Red clump calibrations and Random Forest metallicities: (Yu et al., 25 Jun 2025)
• LLM-based SED analysis and multi-agent frameworks: (Sun et al., 9 Oct 2025, Sun et al., 2024)
• Portable crowdsourcing framework (separate from telescope): (Urbanek et al., 2023)

The Mephisto ecosystem thus spans state-of-the-art hardware, data-driven methodology, autonomous operation, and research software infrastructure, advancing both astronomical discovery and best practices for data-centric scientific research.