Scalable All-Optical Deep Neural Networks Enabled by MoS₂ Integration

Photonic circuits have achieved a 10-picosecond response time for nonlinear activation, solving the persistent bottleneck of broadband signal processing in light-based AI. This speed was previously difficult to reach because conventional waveguides lacked the material properties to handle diverse wavelengths without significant signal loss.

Overcoming Constraints in All-Optical Deep Neural Networks

Electronic neural networks consume vast energy due to resistance and heat. Optical alternatives promise efficiency but often fail at the 'activation function'—the gate that determines signal transmission. Most current solutions are either too slow or limited to narrow frequency bands, preventing large-scale application. Researchers integrated vertically grown molybdenum disulfide (MoS₂) onto silicon chips. This material exhibits saturable absorption, allowing it to act as a high-speed switch across both the O-band and C-band telecommunication windows. The design remains compatible with standard wafer-scale manufacturing, unlike more complex experimental setups.

• Achieves response speeds of ~10 ps.
• Operates across O-band and C-band frequencies.
• Compatible with industrial fabrication processes.

The measured performance suggests that these systems could soon outperform digital processors in specific inference tasks. However, the study does not solve the long-term structural stability of MoS₂ under sustained, high-intensity laser exposure. Future implementation may require robust thermal management to maintain these performance gains.