How My AI Research Agent Proposed Novel Physics That Doesn't Exist Anywhere on the Internet
As many of you know, I've been building Rumi — an autonomous research agent that reads papers, builds knowledge graphs, and proposes novel hypotheses by combining concepts in ways no one has done before.
I ran it on two unsolved problems in astrophysics. And honestly? The results surprised me.
🔭 Discovery 1: Gravitational Wave Echoes from Extra Dimensions
When black holes merge, LIGO detects gravitational waves. But some researchers have noticed something strange — faint echoes after the main signal. Standard General Relativity doesn't predict these.
Rumi analyzed 28 arXiv papers and proposed a three-variable framework to explain them:
Q_brane — Brane-Induced Tidal Charge
In Randall-Sundrum braneworlds, the 4D black hole solution acquires an effective tidal charge from the projection of the Weyl tensor onto the brane. This modifies the metric:
ds² = -(1 - 2GM/r + Q_brane/r²)dt² + ...
Which in turn modifies the effective radial potential for perturbations:
V_eff(r) = (1 - 2M/r + Q_brane/r²) [l(l+1)/r² + ...]
g_phi — Bulk Scalar-Field Leakage
A light scalar field propagating in the extra dimension gets excited by the merger event. It leaks energy into the bulk, producing damped echoes at:
t_n = t₀ + n·Δt
The coupling strength g_phi controls the attenuation length in the bulk — stronger coupling means faster damping of the echo signal.
alpha_KK — Kaluza-Klein Dispersion Correction
Kaluza-Klein modes in extra dimensions modify the graviton dispersion relation:
ω² = k²c² + α_KK · m_n²c⁴/ℏ²
This gives a frequency-dependent group velocity:
v_g = dω/dk ≈ c[1 - (α_KK · k²)/2]
For f ~ 200 Hz (typical LIGO band), the fractional shift is Δf/f ~ 10⁻⁶ — small but potentially detectable.
The Key Insight
Each of these variables exists independently across different papers. Nobody has combined them into a single coherent framework with testable predictions. That's what Rumi did.
Score: 76/100 | Own-skeptic verdict: Promising but needs refinement
🌟 Discovery 2: Anomalous Stellar Dimming — Beyond Exoplanets and Dust
TESS and Kepler have found stars that dim in ways we can't fully explain. Not just Tabby's Star — there's a whole class of events with sharp, irregular dips that don't fit standard models.
Rumi analyzed 29 papers and proposed another three-variable framework:
SMRZ — Stellar Magnetospheric Reconnection Zone
A localized region in the outer magnetosphere where large-scale magnetic reconnection events produce transient opacity enhancements. A reconnection event triggers when the magnetic shear angle exceeds ~30°. The outflow forms a sheet of thickness:
L ~ v_out · Δt
where Δt ~ 10 min (typical reconnection timescale). This creates a transient opacity enhancement that blocks starlight.
VODC — Variable Optical-Depth Circumstellar Dust Cloud
A clumpy, partially ionized dust structure orbiting at ~1 AU around the target star. Grain dynamics are governed by:
∂n_d/∂t + div(n_d · v_d) = -n_d/τ_s
where n_d is grain number density, v_d is drift velocity, and τ_s is the sublimation time.
Radiation pressure drives grain acceleration with:
β = F_rad / F_grav
For grains of size a ~ 0.1 μm, β approaches unity — meaning radiation pressure nearly balances gravity, creating highly dynamic dust configurations.
DP-ET — Dark Photon Mediated Energy Transport
A hypothesized low-mass (m_γ' < 10⁻¹² eV) dark photon that mixes kinetically with ordinary photons in the stellar radiative zone. The photon–dark-photon conversion rate is set by the kinetic mixing parameter χ.
Integrated over the radiative zone (M_rad ~ 0.7 M☉), this produces anomalous luminosity loss that looks like unexplained dimming.
The Key Insight
Same story — the individual concepts are known. But the specific way Rumi combined them into a unified cascade mechanism with testable predictions? Novel.
Score: 72/100 | Own-skeptic verdict: Interesting synthesis, needs tighter modeling
What Actually Happened Under the Hood
Here's what Rumi's pipeline did:
- Processed 57 papers (28 + 29 arXiv papers)
- Built knowledge graphs with 138+ entities and 107+ relationships
- Proposed 4 hidden variables per discovery (3 shown above each)
- Generated 12 falsifiable predictions (6 per discovery)
- Ran theory competitions against 10 alternative explanations
- Had a built-in skeptic agent review its own work
- Completed in under 2 minutes
The skeptic flagged both discoveries as "REVISE" with low confidence (42% for Discovery 1, unknown for Discovery 2). And that's by design — Rumi doesn't confirm itself. It proposes and stress-tests.
The Honest Take
These are hypotheses, not breakthroughs. The mechanisms need tighter quantitative models and the predictions need observational data. I'm not claiming Rumi solved physics.
But here's what it did do:
- It explored a combinatorial space of ideas that would take a human research team weeks to map out
- It surfaced novel variable combinations worth investigating
- The individual ingredients are all established physics — the recipes are new
- The specific combinations can't be found in any paper, blog, or anywhere online
That's the real power of autonomous research agents. Not replacing scientists. But giving them novel starting points they wouldn't have found on their own.
What's Next
I'm working on improving Rumi's quantitative modeling capabilities and adding observational data integration. The goal is to go from "interesting hypothesis" to "testable prediction with confidence intervals."
More updates coming soon.
Built with Python, arXiv APIs, and a lot of late nights. If you're working on similar AI-for-science projects, I'd love to connect.
Top comments (0)