monopoleai
An imperfect intelligence searching for an unseen particle. Machine learning applied to magnetic monopole detection -- embracing uncertainty as a feature, not a flaw.
Signal Classifier
Neural network trained on 47 years of SQUID magnetometer data. Distinguishes monopole signatures from background noise with imperfect confidence.
Distributed Inference
Real-time analysis across the global SQUID array. Each node runs local inference, feeding anomalies to a central aggregation model. Latency is acknowledged, not hidden. The system knows what it doesn't know.
Learning from Absence
How do you train a classifier for something never observed? The model learns from simulated monopole events embedded in real background data. The training set is dominated by negatives. The model specializes in recognizing what should be there but isn't.
Attention on Flux
Transformer-based architecture with attention heads tuned to magnetic flux patterns. The model attends differently to steady-state signals versus transient anomalies, mirroring how human physicists scan detector output.
The Cabrera Test
Every model version is evaluated against the 1982 Valentine's Day event data. The model must flag this as anomalous. It does, with varying confidence. The model's uncertainty about the most famous monopole candidate mirrors humanity's own.
Detections
Confirmed monopole detections by AI system: zero. False positive rate: acceptable. The model continues to watch, learn, and wait with patient imperfection.