In the era of fifth-generation (5G) mobile communication [1–3], optical fiber continues to play an essential role due to superfast transmission speed, tremendous bandwidth, and multiplexing capability in terms of wavelength [4]. With the rapid development of high-speed bandwidth services, higher requirements have been put forward for the optical fiber and transmission network’s functions, flexibility, and expansibility [5–7]. It requires the intelligent transformation in optical communication while realizing the large-capacity and ultra-high-speed transmission [8]. Therefore, we try to entrust optical fiber novel functions, storage, and computing, to serve intelligent optical communication more competently in addition to what we know about optical transmission function.

The common approaches to achieving light storage rely on either the bistability of artificial optical resonances [9–15] or the inherent bistable characteristics of devices stemming from their material properties [16–20]. These optical memories mainly rely on the master–slave configuration [9–12], the feedback loop scheme [13–15], or the injection-locking techniques [16–20], which have high integration density, short memory access time, and low-energy consumption. However, these devices are complex and volatile (extra energy is required to maintain the state of memory). Phase-change materials (PCMs) are ideal candidates for nonvolatile memories, which have large contrast in electrical/optical properties between the crystalline state and the amorphous state. They have been extensively studied and applied for electronic memories [21–26]. However, the “electronic rate bottleneck” of electronic memories limits modern computing technology development. All-optical memory based on PCMs can mitigate the von Neumann bottleneck with high speed and low power consumption in data transmission. The researchers use ${\mathrm{In}}_3{\mathrm{SbTe}}_2$ [27], ${\text{CsPbIBr}}_2$ [28], ${\mathrm{Ge}}_2{\mathrm{Sb}}_2{\mathrm{Te}}_5$ (GST) [29–32], ${\mathrm{Ge}}_2{\mathrm{Sb}}_2{\mathrm{Se}}_4\mathrm{Te}$ [33], ${\mathrm{Ag}}_5{\mathrm{In}}_5{\mathrm{Sb}}_{60}{\mathrm{Te}}_{30}$ [34,35], and other PCMs to achieve optical memories; in particular, the GST has the best performance in terms of speed and stability. Reference [30] demonstrates a fast nonvolatile GST-based PCM memory with a capacity of 3 bits, occupying an area of only $0.4\mathrm{\mu m}\times 0.4\mathrm{\mu m}$, and the speed was close to 1 GHz. The recent GST memory realized the 12-level all-optical storage and realized logic operations (logic “OR” and “NAND”) by using the pulse-width modulation method [36]. However, current chip-to-fiber coupling relies on edge or grating couplers with limitations in alignment tolerances, efficiency, and bandwidth, respectively [37], so the waveguide-based device is difficult to make compatible with optical fiber communication systems. We design a PCMs-based all-fiber memory, which can couple light to and from optical fiber communication network with large efficiency.

Herein, we demonstrate all-fiber memory by combining the ordinary single-mode optical fiber (Corning SMF-28), GST, and indium-tin-oxide (ITO) (as the capping layer). Since GST has a significant contrast in optical reflectivity between the amorphous state (logic state 0, low-reflectivity state) and the crystalline state (logic state 1, high-reflectivity state), information is encoded as the reflectivity of the GST. It is stored in the phase state of GST. We use pump pulse (${\lambda }_1=1550\mathrm{nm}$) with a fixed pulse width to switch the state of GST (the low-energy pulse converts GST to the crystal state, and the high-energy pulse converts GST to the amorphous state), and the switching energy of the all-fiber memory is as low as about 50 nJ. We use continuous light (${\lambda }_2=1530\mathrm{nm}$) whose energy is much lower (0.2 mW) than the switching energy to achieve data reading. The operation principle of the all-optical fiber memory is shown in Fig. 1. We perform eight-level storage (3-bit) with high repeatability and excellent reversibility using the pulse amplitude modulation (PAM) method, and the optical contrast is about 38%. Moreover, each memory level has an associated PAM parameter that will always result in written memory level, whatever the starting state. We further explore its application to the provision of a nonvolatile logical fiber calculator, achieving “AND” and “OR” computations. The presented all-fiber memory opens new routes toward the intelligent fiber system [38].

GST is a kind of chalcogenide phase-change material. It has a wide range of applications in near-infrared absorbers and modulators [39], data storage [40], and data calculation [41,42]. GST has two phases corresponding to the amorphous structure (a-GST) and cubic structure (c-GST). The two phases have very different material properties, thus providing the contrast requirements of distinguished logical states [43]. The GST can transform phases with electrical pulses or light pulses. People use significantly different electrical/optical properties in different phases to achieve data storage.

The scanning electron microscope images of our all-fiber memory cell are shown in Figs. 2(a)–2(d). The thicknesses of the GST and ITO are 120 nm and 30 nm, respectively. ITO prevents GST from being oxidized. Our all-fiber storage unit realizes optical storage by using the difference in reflectance between a-GST and c-GST. The scheme is the same as the principle used in conventional optical storage (where reflectivity is modulated). The crystalline state (which we term “level 1”) exhibits higher reflectivity than the amorphous state (“level 0”) [44,45]. Therefore, we define the conversion from a-GST to c-GST (crystallization) as the “write” process and the conversion from c-GST to a-GST (amorphization) as the “erase” process; the principle of GST transformation is shown in Fig. 1. We simulate a single 10 ns, 47 mW optical pulse train (the number is 100, and the repetition rate is 1 kHz) to heat GST to 580 K (see Fig. 3) to achieve the crystallization temperature of GST (500–600 K), and the 48.5 mW optical pulse can heat GST to an amorphous state, reaching the amorphous temperature of GST (890 K) [46].

Multi-level storage requires the cell reflectivity to be programmed with higher accuracy as compared to bi-level storage. We must precisely control the crystal fraction of the active GST material to achieve the required cell reflectivity value with adequate accuracy [47]. We select a $15\mathrm{\mu m}\times 15\mathrm{\mu m}$ square area of GST, and the area where the temperature surpasses the melting temperature of GST (890 K) is shown in dark red and denotes the final amorphous region of the memory cell after a single optical pulse train. With the increase of the pulse power, the amorphous area of GST gradually becomes more considerable. In this paper, we assume that the two states with the largest reflectivity differences are completely crystalline and completely amorphous. We define the amorphous ratio of the fully crystallized state as “0” and the amorphous ratio of “1” when the dark red region is the largest. We can obtain the amorphous area of GST by taking boundary points through simulation software. The permittivity of partially amorphous GST can be calculated from effective permittivity approximated by an effective-medium theory [48,49]: $$\frac{\epsilon (p)-1}{\epsilon (p)+2}=p\times \frac{{\epsilon }_a-1}{{\epsilon }_a+2}+(1-p)\times \frac{{\epsilon }_c-1}{{\epsilon }_c+2},$$(1)where ${\epsilon }_c$ and ${\epsilon }_a$ are the complex permittivity of c-GST and a-GST, respectively, and $p$ is the degree of amorphization (ratio of partially amorphous area to completely amorphous area). Moreover, the permittivity can be calculated from the refractive indices of GST by ${\epsilon }^{0.5}=\mathit{n}+\mathit{i}\kappa$, where n is the real part of the refractive index of GST, and $\kappa$ is the imaginary part of the refractive index of GST. Thus, the different levels of amorphization of GST lead to various values of $\epsilon (p)$, leading to different levels of reflectivity. Figure 3 shows the reflectance of GST with different amorphous proportions, which is normalized. The simulated results indicate that the reflectance is inversely proportional to the amorphous area of GST.

Our all-fiber memory device consists of a GST thin film passivated with a layer to prevent GST from oxidizing (ITO) and a single-mode fiber. GST is deposited on the end face of the fiber by radio frequency (RF) sputtering (50 W RF power, 30 mTorr Ar atmosphere, and $9\times {10}^{-4}\mathrm{Torr}$ base pressure), and ITO is deposited by direct current (RF) sputtering (10 W power, 15 mTorr Ar atmosphere, and $9\times {10}^{-4}\mathrm{Torr}$ base pressure). All depositions are performed at room temperature.

We use a pulsed laser (wavelength: ${\lambda }_1=1550\mathrm{nm}$) to switch the state of the fiber memory, and we use a continuous-wave laser (wavelength: ${\lambda }_2=1530\mathrm{nm}$) to monitor the optical reflection (optical memory level). The optical setup in the experimental measurements is sketched in Fig. 4. The probe signal is attenuated by an attenuator (0.2 mW) to reduce potential drifts caused by the thermal-optical effect of GST. Then the two optical beams are injected into the all-fiber memory cell through the optical coupler and the circulator (CIR). The fiber reflects the pump pulses and the probe signal. The optical pulses are filtered by the bandpass filter. A data acquisition card connects a photodetector to a computer to monitor the all-fiber memory’s reflection in real time.

The optical reflection of the a-GST is defined as the baseline; the pulse train (the number is 100 and the repetition rate is 1 kHz) with a fixed width of 10 ns and different peak power ($P_i$, $\mathit{i}=2$, 3, 4, 5, 6, 7, 8) can switch the state of fiber memory. The reflectivity change is defined as $\mathrm{\Delta }{R}_i=R_i-R_1$, where $R_i$ ($i=2$, 3, 4, 5, 6, 7, 8) is the final reflectivity after laser irradiation, and $R_1$ is the initial reflectivity before laser irradiation (lower reflection). The relative reflectivity change is expressed as $(\mathrm{\Delta }{R}_i/R_1)\times 100\%$, and the different relative reflectivity presents the different levels. We define the $P_8$ level as entirely crystalline and the $P_1$ level as entirely amorphous; other levels are partly crystalline or partly amorphous.

Here the powers of different optical pump pulses are $P_1=54.5\mathrm{mW}$, $P_i=53.5$, 52.5, 51.5, 50.5, 49.5, 48.5, 47 mW ($i=2$, 3, 4, 5, 6, 7, 8) in the experiment, and the optical power increase from $P_8$ to $P_7$ is more than the others, which can make the result of the experiment have a better linear relationship. The corresponding energies are 54.5, 53.5, 52.5, 51.5, 50.5, 49.5, 48.5, and 47.0 nJ.

Levels 1 to 8 are sequentially achieved by decreasing the power from $P_1$ to $P_8$. The PAM switching can also work vice versa for the amorphization. For example, levels 8 down to 1 are sequentially achieved by increasing the power from $P_8$ to $P_1$ [Fig. 5(a)]. Importantly, we can use a pulse sequence with a power of $P_i$ to switch to any energy level i, and this switch does not consider the state of GST. Moreover, the reflectivity contrast is about 38% between levels 1 and 8. The result demonstrates that our all-fiber memory cell has good reversibility and high contrast in optical reflection. Figure 5(c) also demonstrates that the same level has the same corresponding $P_1$ and $P_i$, and each level is not affected by the previous state. It implies that every level can be accessed from all others, with accurate reflection levels and remarkable repeatability.

Level 6, for example, can be achieved from the level 1, 5, or 8 with the power of pump pulse of $P_6$. Due to our unique fiber reflection structure (different from the waveguide structure), the impact of a pulse sequence on GST may be divided into two parts. The first part of the energy heats the GST into a crystalline state (initialization process). During this conversion process, the absorption coefficient $k$ of GST gradually increases, so the energy acting on GST gradually will increase. The power is large enough to make the GST amorphized, so the subsequent pulses gradually accumulate to convert the GST to the amorphous state. The final switching level depends on the power of the light pulse [50]; the higher the optical pulse power, the greater the degree of amorphization of GST, resulting in lower reflectivity. Hence, the level and power have a “one-to-one” relationship between values, so the all-fiber memory can realize arbitrary switching by using the power of the optical pump pulse (P_i) corresponding to this level. When the initial complex refractive index of the GST is relatively large, the energy of the initial pulses is easier to act on, so the number of pulses will be small. Conversely, when the initial complex refractive index is small, more pulses are required to act on the GST. Due to the state’s difference after initialization, the same state may still be achieved even if the number of accumulated pulses is different. From Figs. 5(a) and 5(c), we can see that our multi-level switching has a small drifting.

We repeatedly switch GST states between levels 1 and 8 to further describe the storage performance. Figure 5(b) shows the result of repeated switching over 40 cycles, demonstrating that the all-optical data storage on the fiber end face has high repeatability. PCMs have operation cycles up to ${10}^{11}$ and stay in the state for many years until sufficiently enormous energy is applied to PCMs [51]. We repeat the same series of pulses 10 times and plot a histogram of the error between the target reflectivity and actual reflectivity [see Fig. 5(d)]. It can be seen that the energy drift error of our device is within $\pm 1.5\%$. We also have supplemented the experiment for investigating the thermo-optical effect of GST, as shown in Fig. 5(e). It can be seen for a single pulse that the response time of the amorphization process is 158 ns, and the response time of the crystallization process is 145 ns.

We can use memory cells to manufacture all-fiber logic gates (“AND,” “OR”), which can meet the requirement of next-generation optical networks [52]. We name the present optical memory as “memlogic,” which is short for memory logic with both nonvolatile configurations of logic functions and nonvolatile logic output results [53].

Optical pulses of the power ($P_8$ and $P_1$) corresponding to the entirely crystalline state (level 8) and entirely amorphous state (level 1) are used as the logic input “1” and the logic input “0” (the two logic inputs have the same pulse width $W=10\mathrm{ns}$); the enormous optical contrast makes the two logic inputs more easily distinguished. As a result, the logic output has a higher contrast ratio when carrying out logic operations. Here, we define the normalized reflection (NR) of the two input pulses passing through two memlogic units as the logic output.

Figure 6(a) shows the logic operation principle of typical memlogic cells. First, as shown in Fig. 6(a), we connect two memlogic cells in series to perform the “AND” logic operation. Next, we connect two memlogic cells in parallel and combine two outputs with a coupler to achieve the “OR” logic operation. Each memlogic unit must be connected to a CIR because our all-fiber memory measures the reflection of the fiber. The reflectivities of two series memlogic units are defined as R_in1, R_in2, and the reflectivities of “OR” gate memlogic units are defined as R_in3, R_in4, respectively. According to the series and parallel characteristics of two memlogic units, the value of the reflectivity output after series (R_s) and parallel (R_p) connections can be expressed as ${\mathit{R}}_s=R_{\text{in}1}\times R_{\text{in}2}$; $R_p=R_{\text{in}3}+R_{\text{in}4}$. The input information is stored in the memlogic cell and is read by a continuous wavelength light. We conduct an optical erasing process with both memlogic cells to ensure they are in the amorphous state before performing the logic calculation. We take the “AND” gate in series as an example. We use two independent optical pulses as optical inputs (Pin1, Pin2), and the two pulses of fixed width (10 ns) are used to program digital inputs “0,” “1,” corresponding to pulse amplitudes of 54.5 mW, 47 mW, respectively, and are applied to memlogic cells. As shown in Fig. 6(b), we can obtain distinct NR values with different digit combinations of Pin1 and Pin2 (Pin3 and Pin4). We compared the value of NR with the reference value of 0.5 (red line); the logic output (Y1 and Y2) of the optical memory can be either “0” or “1” (below and above the reference line). When the inputs are “00,” “01,” “10,” and “11”, respectively, the NRs are 0.00, 0.29, 0.28, and 0.99 in the logic device of the “AND” gate, and the outputs are 0.01, 0.76, 0.78, and 0.96 in the logic device of the “OR” gate. We set the reference line to 0.5. Thus, the logic “AND” and logic “OR” outputs are “0,” “0,” “0,” “1,” and “0,” “1,” “1,” “1”, respectively. The truth tables of the memlogic cells are shown as Figs. 6(c) and 6(d), corresponding to the results of the traditional electronic “AND” and “OR” gates, respectively.

Importantly, when the logic inputs of $\mathrm{P}\mathrm{in}1$ and $\mathrm{P}\mathrm{in}2$ are “1,” “0” or “0,” “1”, respectively, the values of NR are different because the reflectivity contrast is different of two memlogic cells. Moreover, the differences between the two outputs will increase because of the increasing difference in the memlogic units’ reflectivity contrast. However, the final logic results are not affected by this slight difference. Finally, it is worth mentioning that we can extend our logic devices to achieve more complex optical logic functions (normalizing reflectivity, a standardized device is added after each logic gate). Therefore, our fiber memlogic cells have excellent potential in all-optical networks.

In summary, we have shown the demonstration of 3-bit data storage with an all-fiber memory cell. We use the PAM switching technique to realize partial crystallization of the PCM memory cell, and we can control the reflectivity in the GST layer by pulse power ($\sim 50\mathrm{nJ}$). Furthermore, we show the readily and directly switching ability between the multiple memory levels, with high accuracy of the readout signal and excellent repeatability. Moreover, we use the memlogic units in series and parallel to realize the essential logic operation (“AND” and “OR”), and we can recognize more complex logic operations by using more fiber optic memlogic cells.

The results are a broad and fundamental step toward a fast, low-power fiber memory for applications. The all-fiber nonvolatile multi-level memory can have applications such as neuromorphic computing and in-memory computing. The primary logic gate is expected to realize more complex logic operations. The application of the memlogic unit in optical fiber communication networks is expected to lead to a new revolution in optical fiber communication technology.

Acknowledgment. The authors thank the College of Physics and Optoelectronic Engineering of Harbin Engineering University for providing technical support under the Key Laboratory of In-fiber Integrated Optics, Ministry of Education, China.