Objective Hardware reservoir computing has garnered significant attention for its low power consumption and innate suitability for processing temporal signals. However, the performance of hardware reservoirs built on conventional neuromorphic devices is fundamentally constrained by their single, fixed relaxation dynamic. This limitation hinders the comprehensive capture of information from complex real-world signals that inherently possess multi-scale spatiotemporal features, leading to the loss of partial information and a ceiling on computational accuracy. To overcome this bottleneck, this work proposes a co-design strategy integrating novel device physics and reservoir architecture. The core objective is to develop a single photonic memristor device exhibiting dual-wavelength-sensitive, tunable short-term plasticity, and to leverage this "one-device-dual-mode" capability to construct a hybrid reservoir computing system. This system aims to intrinsically generate and fuse multiple temporal dynamics within a single physical reservoir, thereby enriching its state space and enhancing multi-scale feature extraction without increasing hardware scale or complexity.

Methods In this paper, a dual-wavelength hybrid storage cell based on dual-wavelength sensitive photoelectric memristor is reported, which achieves higher performance than the storage cell with single attenuation characteristics (Fig.1). The device has different attenuation characteristics under different wavelength pulse stimulation (Fig.2), and the principle comes from the different wavelength sensitivity of the absorption layer (Fig.3). Based on the attenuation of dual space-time characteristics, a dual-wavelength hybrid storage pool is designed, and good performance is achieved (Fig.4).

Results and Discussions This work successfully fabricated an organic photonic memristor (ITO/PEDOT:PSS/IDTBT/ZnO:PbS QDs/Al) with selective sensitivity to 808 nm and 365 nm light. The device exhibited fundamentally different short-term plasticity under these wavelengths. As shown in the transfer curves (Fig.1(c)), the device showed a stronger photocurrent under 365 nm light. The transient responses (Fig.2(a)-(f)) revealed distinct decay behaviors: a fast relaxation (τ~0.172) under 808 nm pulses and a slow, prolonged relaxation (τ~0.431) under 365 nm pulses. The underlying mechanism, illustrated by the energy band diagram and heatmaps in Fig.3, is attributed to different photogeneration and carrier trapping processes. Under 808 nm light, excitation is primarily confined to PbS quantum dots, leading to fast recombination. Under 365 nm light, the broader absorption across IDTBT, ZnO, and PbS generates a high density of carriers that are deeply trapped, resulting in long-term photo-memory. Based on this device, a hybrid reservoir computing system was constructed. The input signal was converted into parallel 808 nm and 365 nm optical pulse streams to co-stimulate the same physical node array. This enabled the single reservoir to concurrently produce two sets of transient states with complementary temporal scales—fast dynamics capturing fine details and slow dynamics integrating global trends—effectively enriching the high-dimensional state space. The system's performance was validated on the MNIST handwritten digit classification task. The final test accuracy, presented in Fig.4(c), was 84.75% for the dual-wavelength hybrid reservoir, significantly outperforming the single-mode reservoirs (81.70% for 808 nm-only and 82.25% for 365 nm-only). The confusion matrices in Fig.4(d)-(f) further visually confirmed that the hybrid reservoir made fewer classification errors. The training curves in Fig.4(b) also demonstrated that the hybrid reservoir achieved faster convergence and lower loss, highlighting the computational advantage of the fused dynamics.

Conclusions This work addresses the limitation of single relaxation dynamics in hardware reservoirs by developing a dual-wavelength-sensitive organic photonic memristor and a corresponding hybrid reservoir architecture. The key innovation lies in achieving "one-device-dual-mode" tunable dynamics (fast and slow decay) through wavelength-dependent photophysical processes. The proposed hybrid reservoir computing system leverages this to intrinsically fuse multiple spatiotemporal dynamics, significantly enriching the state space and enhancing multi-scale feature representation within a single, unchanged physical hardware footprint. The demonstrated performance improvement on a standard benchmark task validates this co-design strategy as an effective pathway to break the performance ceiling of existing hardware reservoirs. Moving beyond the conventional approach of scaling hardware nodes or adding circuit complexity, this research provides a new paradigm centered on engineering multi-modal dynamics at the fundamental device level. It paves the way for developing next-generation neuromorphic computing hardware with higher precision and efficiency, particularly for processing complex signals like video streams or biomedical sensor data. Future work will focus on expanding the multi-modal dimensions (e.g., more wavelengths, multi-physical field control) and exploring applications in more challenging spatiotemporal signal processing domains.