The ^G Infrared Search for Extraterrestrial Civilizations with Large Energy Supplies. V. When Galaxies Glow with Industry

The paper presents the most robust stellar population synthesis (SPS)-based search for galaxy-spanning technological waste heat (Dyson spheres) to date. The authors incorporate the AGENT Dyson sphere formalism into the Flexible Stellar Population Synthesis (FSPS) code at the stellar population level, so that nebular and dust emission respond self-consistently to Dyson sphere reprocessing. They pair this forward model with the Prospector Bayesian inference framework to fit 24-band photometry from the Brown et al. (2014) galaxy atlas, which contains 129 nearby galaxies spanning a wide range of spectral energy distribution (SED) types, including ultraluminous IR galaxies and MIR-luminous active galactic nuclei (AGN).

The authors perform a suite of 1,419 injection recovery tests across a range of covering fractions, α, where they successfully recover the injected covering fractions (best-fit slope m = 0.92) and detect them through Bayesian model selection down to α ∼ 4–5% in quiescent galaxies. None of the 129 galaxies prefer a Dyson sphere component, and they place the first per-galaxy 95% upper limits on warm (TBB ≳ 100K) swarms, reaching a median α < 0.3% across quiescent hosts without a dominant AGN. Their injection-calibrated detection rates convert these zero detections into a population bound of < 2.6% of galaxies hosting α = 25% swarms (95% confidence).

Because survey colors cannot separate waste heat from starbursts and AGN, the authors develop a scaffold for future searches, running from inexpensive archival screens such as the Balmer decrement and the stellar-to-dynamical-mass offset a swarm leaves behind, through resolved fitting with nuclear excision, to PRIMA FIR photometry that makes targeted JWST imaging decisive. They find that the outskirts of quiescent galaxies are the best hunting grounds for future technosignature searches.

Key results include:

• Detectability is easiest in the outskirts of quiescent galaxies, with Bayesian model selection recovering covering fractions as low as ∼4–5% in old, quiescent galaxies, whereas dusty star-forming systems require α ≳ 20%.

• Restricting the fits to wide-field survey photometry raises these thresholds by a further factor of ∼3–4 for quiescent galaxies.

• When waste heat is present but unmodeled, the fitter inflates the inferred AGN fraction by up to three orders of magnitude and the recent star formation rates of quiescent galaxies by 1.4–1.8 dex.

• The dominant systematic is the degeneracy between warm AGN dust and waste heat, and spatially resolved fitting with nuclear excision breaks it. For M77, excising the inner 34′′ lowers the recovered covering fraction from α = 0.182 to α = 0.100+0.137−0.054, consistent with zero at 1.9σ.

• Future observations extend these limits: PRIMAger photometry pins the FIR continua of the waste-heat and AGN hypotheses jointly, concentrating their remaining disagreement in the 8–13 µm region that JWST/MIRI can test directly.

Improvements to AI Systems:

1. Self-Consistent Multi-Component Forward Modeling

• Integrate domain-specific physical models (e.g., Dyson sphere reprocessing) directly into the generative pipeline, so that all derived observables (nebular emission, dust continuum, photometry) respond coherently to a single set of parameters.

• Improved capability: An AI that can simulate and fit complex astrophysical systems with coupled feedback loops (e.g., energy recycling, radiative transfer) without ad hoc corrections, reducing systematic biases in parameter inference.

2. Bayesian Model Selection with Injection-Calibrated Detection Thresholds

• Embed injection-recovery testing as a standard calibration step for any detection claim, producing empirical sensitivity curves (e.g., covering fraction vs. detectability) rather than relying on theoretical noise estimates.

• Improved capability: An AI that can report not just best-fit parameters but also calibrated false-positive/negative rates and population-level upper limits (e.g., <2.6% of galaxies host a 25% swarm) with rigorous confidence intervals.

3. Degeneracy-Aware Posterior Analysis

• Implement explicit modeling of known degeneracies (e.g., warm AGN dust vs. waste heat) by running multiple hypotheses in parallel and comparing their evidence, rather than fitting a single model.

• Improved capability: An AI that can flag when two physically distinct scenarios produce nearly identical observables, and then recommend the most informative next observation (e.g., nuclear excision or FIR photometry) to break the tie.

4. Spatially Resolved Fitting with Region Excision

• Extend the inference framework to operate on multi-resolution spatial data, allowing the AI to mask or down-weight contaminated regions (e.g., galactic nuclei) and fit the outskirts separately.

• Improved capability: An AI that can isolate faint signals (e.g., technosignatures) from dominant foregrounds by learning which spatial scales and wavelengths maximize signal-to-noise, as demonstrated by the M77 case where excision reduced false covering fraction from 0.182 to 0.100.

5. Multi-Wavelength Photometric Joint Fitting with Future Instrument Priors

• Incorporate synthetic photometry from upcoming instruments (e.g., PRIMA, JWST/MIRI) as virtual observables during training, so the AI can predict which additional data points will most reduce posterior uncertainty.

• Improved capability: An AI that acts as an adaptive observation planner, suggesting specific filters, exposure times, or spatial cuts that will most efficiently discriminate between competing hypotheses (e.g., 8–13 µm region for waste heat vs. AGN).

6. Systematic Error Propagation from Unmodeled Physics

• Train the AI to detect when its own model is incomplete by monitoring for inflated nuisance parameters (e.g., AGN fraction increasing by 3 orders of magnitude) as a diagnostic of unmodeled components.

• Improved capability: An AI that self-audits its fits and alerts the user when a parameter is being used as a garbage collector for missing physics, prompting a model revision rather than a false positive detection.

7. Population-Level Inference from Zero Detections

• Use Bayesian hierarchical modeling to convert individual upper limits into population constraints, even when no positive detections occur.

• Improved capability: An AI that can produce statistically rigorous statements like fewer than 2.6% of galaxies host α=25% swarms from a null result, enabling meaningful technosignature searches without requiring a single hit.