280
J. Eur. Opt. Society-Rapid Publ. 22, 27( 2026)
Fig
. 5. Simulation results of digit‘ 5’ with intensity detection and true class distribution.
Fig
. 6. Magnetic domain patterns under random initial state, for flip sizes( a) 1 1,( b) 2 2,( c) 3 3,( d) 4 4,( e) 5 5 and( f) 10 10 lm 2.
reduced learning stability. At a flip size of 10 10 lm 2,( panel f) the resolution collapses into enlarged domains, demonstrating instability driven by excessively large parameter updates. This agrees with the accuracy and loss curves, which likewise exhibit degraded performance at larger flip sizes.
Physically, this behavior originates from the diffractionlimited optical coupling condition, which governs light propagation between adjacent layers through its dependence on the diffraction angle( b) and the interlayer spacing. In a fully connected neural network, each neuron in one layer interacts with all neurons in the next. In an MO-D 2 NN, this connectivity is realized optically through the first-order diffracted light.
For a neuron of diameter d n, the1 st order diffraction angle b is given by: b ¼ sin �1 ðk = 2d n Þ
From this angle, the interlayer distance D l needed to maintain full optical connectivity followed from the geometrical separation between the farthest neuron in adjacent layers: