Maintenance measures for tangential flow filtration system
05,11,2026 157 views
【InstrumentOnline R&D News】Combination optimization problems are widely present in fields such as finance, logistics, and machine learning. The traditional "von Neumann" architecture faces significant bottlenecks when dealing with increasingly large-scale problems. The optoelectronic Ising machine utilizes the dynamic evolution of physical systems for intrinsic parallel solving, which is an important technical route for achieving large-scale combinatorial optimization hardware acceleration. However, there is an essential difference between the continuous simulation characteristics of its underlying physical carrier and the discrete binary constraints of the Ising model, which makes the system evolution process susceptible to factors such as uneven hardware amplitude, noise accumulation, and performance drift, resulting in decreased accuracy and limited scalability of the solution.
Recently, Li Ming, a research team from the National Key Laboratory of Optoelectronic Materials and Devices, Institute of Semiconductor, Chinese Academy of Sciences, jointly with Beijing Ising Intelligent Technology Co., Ltd., published the latest research progress in Photonics Research: the research team proposed a hybrid neuromorphic optoelectronic Ising machine (HNOIM) based on the dynamic steep activation feedback mechanism. Through the cooperation of simulation exploration and digital locking, it corrected the calculation formula deformation problem (Hamiltonian distortion) caused by amplitude inhomogeneity, and realized the effective combination of continuous dynamic evolution and binary stable convergence. G-set (G1-G21) is widely recognized and highly recognized in the industryStandardIn the case library test, HNOIM achieved a solution accuracy of 99.68%. Multiple sets of complex cases stably output known Z-optimal solutions (BKS), with a single solution time (TTS) of only 0.44 milliseconds, which is three orders of magnitude higher than the traditional simulated annealing algorithm (199.61 milliseconds). The research team conducted feasibility tests on ultra large scale complex scenarios. In a fully connected ultra large scale optimization network with 1601 nodes and 9 high-precision weights for each connection, the HNOIM hybrid architecture reduced the residual energy error (the proportion of the difference between the actual computational energy of the device and the theoretical Z-optimal Z-low energy of the problem) to 0.068%, and improved the Z-optimal energy approximation ability by 43 times.
This study provides a new technical solution to improve the accuracy and scalability of optoelectronic Ising machines, and provides important support for the development of optoelectronic Ising calculations and their application in large-scale combinatorial optimization.
As the main body of academic innovation, the Semiconductor Institute proposed the "Dynamic Steep Activation Feedback Mechanism" by a team of doctoral students, and was fully responsible for the architecture transformation, full process experimental physics verification, data analysis, and academic paper writing and publication of the optoelectronic Ising machine system; Beijing Yixin Intelligent Technology Co., Ltd. relied on its industrialization advantages to provide a highly comprehensive experimental environment and basic hardware platform for the research. It also dispatched an FPGA engineering team to work with semiconductor research personnel to jointly develop the software and hardware of the underlying core computing logic, laying the engineering foundation for high-precision and high-performance case solving.
The research findings, titled "Hybrid neuromorphic optoelectric Ising machine via dynamic steam activation feedback," were published in Photonics Research. Li Zhentong, a doctoral student at the Institute of Semiconductors, is the first author, Ma Xuelong, a doctoral student at the Institute of Semiconductors, is the co first author, and Li Ming, a researcher at the Institute of Semiconductors, is the corresponding author. This work was funded by the Beijing Science and Technology Commission's Beijing Science and Technology Plan Project "Research on Key Technologies of Optoelectronic Ising Machine Intelligent Computing" (Z241100004224001).
