Hybrid neuromorphic optoelectronic Ising machine via dynamic steep-activation feedback
Zhentong Li, Xuelong Ma, Qi Chen, Zhoujuan Cui, Miaomiao Wei, Zezhou Tan, Yangrui You, Zhirong Fan, Ye Xiao, Ming Li · Photonics Research · 2026
Optoelectronic Ising machines have emerged as promising accelerators for combinatorial optimization, yet their performance is fundamentally hindered by intrinsic amplitude inhomogeneity, which distorts the effective Hamiltonian and traps the system in suboptimal local minima. Drawing inspiration from neural activation annealing, we propose and experimentally demonstrate the hybrid neuromorphic optoelectronic Ising machine (HNOIM), driven by a dynamic steep-activation feedback mechanism. This architecture orchestrates a synergistic temporal phase transition: initiating with a soft-activation analog phase to foster global exploration across energy barriers, followed by a rigorous digital locking phase that enforces discrete binary constraints. Benchmark evaluations on the G-set (G1–G21) demonstrate that the HNOIM consistently reaches the best known solutions (BKSs) on challenging instances, achieving a peak accuracy of 99.68% with a sub-millisecond time-to-solution ( TTS ≈0.44 ms ). This performance represents a three-orders-of-magnitude speedup over optimized simulated annealing ( SA ≈199.61 ms ) while maintaining superior solution fidelity. Crucially, in scalability stress tests on fully connected, 9-bit weighted graphs ( N =1601), the proposed method suppresses the residual energy gap from 2.93% to a negligible 0.068%. This 43-fold suppression of optimality error confirms that the time-multiplexed digital feedback effectively decouples solution fidelity from the cumulative noise of large-scale networks, establishing a robust and high-speed pathway for high-fidelity neuromorphic computing.