Three blocks of convolution are sufficient to detect strong gravity lenses, with a AUC of 0.99

LUMA combines physics-inspired pretreatment, which brings out bows and tenus rings, with a compact network — and achieves 96% accuracy on simulated data.

Strong gravitational lensing—configurations where the mass of an object in the foreground distorts the light from a more distant source into arcs or rings—are rare and laborious to spot in large image surveys. Giovanni Vincenzo Donatiello, Achille A. Nucita, Francesco De Paolis, Antonio Franco and their co-author propose LUMA, an automatic detection chain that relies on sobriety rather than model size.

The device combines two elements. First, a pre-processing step motivated by the physics of the phenomenon, responsible for enhancing the most tenuous arcs and rings before the network sees them. Second, a compact three-block convolution network, trained with reweighting of classes to compensate for their imbalance and a modern ordering of the learning rate.

On simulated data, the model achieves a test accuracy of about 96% and an area under the ROC curve close to 0.99 for non-trivial classes, with the analysis of the confusion matrix showing high completeness and purity on the candidates. The authors draw a measured conclusion: a lightweight convolutional architecture is a competitive benchmark for this type of research, which justifies extending the work to real survey images and transformer-based models.