Skip to content
Frontiers
Search

Review

Frontiers in Photonics ·

Volume 7 - 2026 | https://doi.org/10.3389/fphot.2026.1918762

Photonic devices based on phase-change materials for neuromorphic computing

  • Ranjeet Dwivedi1
  • Charles Pourny2
  • Anthony Albanese3
  • Agraj Yadav4
  • Régis Orobtchouk1
  • Gaya Benane2
  • Rajath Sawant2
  • Martina Tomelleri3
  • Yoann Brulé3
  • Arnaud Rogemont2
  • Aurélien Coillet2
  • Benoit Charbonnier3
  • Pierre Noé3
  • Benoit Cluzel2
  • Fabio Pavanello4*corresponding author
  1. 1INSA Lyon, Ecole Centrale de Lyon, CNRS, Université Claude Bernard Lyon 1, CPE Lyon, INL, Écully, France
  2. 2Laboratoire Interdisciplinaire Carnot de Bourgogne UMR 6303, CNRS-Université Bourgogne Europe, Dijon, France
  3. 3University Grenoble Alpes, CEA, LETI, Grenoble, France
  4. 4University Grenoble Alpes, University Savoie Mont Blanc, CNRS, Grenoble, France
Article metrics0 viewsView details
PDFCiteMetrics

Abstract

This paper reviews recent advances in the field of photonic devices integrating phase-change materials (PCMs). In particular, certain PCMs showcase interesting properties such as non-volatile behaviour and ultra-low optical losses that can be leveraged in computing architectures. For example, they can operate as complex-valued weights enabling ultra-low power consumption matrix-vector multiplications. Here, we first review the optical properties of multiple PCM compounds as interesting candidates for photonic computing. Then, we discuss about GeSe-based electrically activated photonic devices that exploit such properties in novel device concepts. Specifically, we present a hybrid-mode interferometer and an asymmetric heater geometry along the propagation direction for electrical PCM actuation. These approaches lead to a higher number of encoded states as well as ultra-low optical insertion losses compared to previous implementations, while retaining a compact footprint and practical driving conditions. Finally, we present experimental results for optically actuated Ge 2 Sb 2 Te 5 (GST225) based waveguides in terms of encoding capabilities and reliability for all-optical reconfigurability and processing.

Keywords

1 Introduction

Modern artificial intelligence (AI) workloads are increasingly limited not only by the energy required to perform arithmetic operations, but also by the energy and latency associated with moving data between memory and processing units, also called the communication bottleneck. As complementary metal-oxide-semiconductor (CMOS) scaling slows, the classical von Neumann separation between storage and compute has become a dominant bottleneck: data movement can cost orders of magnitude more energy than local arithmetic, motivating new architectures based on in-memory, near-memory, and domain-specialized computing (Horowitz, 2014; Leiserson et al., 2020).

Photonic computing offers an attractive route because optical signals can propagate with very high bandwidth, low latency, and inherent parallelism through wavelength, spatial, and temporal multiplexing. These properties have motivated optical neural networks, photonic matrix-vector multiplication (MVM) engines, and photonic tensor cores as potential accelerators for the linear-algebra operations that dominate inference and signal-processing workloads (Shen et al., 2017; Harris et al., 2018; Zhou et al., 2022; Peserico et al., 2023; Fu et al., 2024). However, a central challenge remains: practical photonic processors require dense, programmable, reliable, and energy-efficient memory elements to store weights, biases, routing states, and calibration parameters directly inside the photonic circuit.

Phase-change materials (PCMs) address this missing memory problem in a remarkable way. In fact, chalcogenide PCMs such as GeSbTe (GST), GeSbSeTe (GSST), Sb 2 S 3, Sb 2 Se 3, and related compounds exhibit large, reversible changes in optical refractive index and absorption between amorphous and crystalline states, enabling compact photonic weight cells, switches, memories, phase shifters, and synapses (Ríos et al., 2015; Zhang et al., 2019; Delaney et al., 2020; Ríos et al., 2022; Prabhathan et al., 2023; Li et al., 2023). Crucially, these states are nonvolatile: once programmed, the optical state can be retained without continuous electrical power, reducing the static energy overhead to hold a given configuration that affects thermo-optic or carrier-based photonic circuits (Fang et al., 2022a; Chen et al., 2023; Sun et al., 2025; Chen et al., 2024). This makes PCM-based devices especially relevant for photonic computing, where many weights must remain fixed during inference while optical signals perform MVM operations at high speed. Demonstrations have already shown PCM-based photonic in-memory computing, all-optical neurosynaptic systems, integrated photonic memories, and photonic tensor cores combining PCM weight banks with wavelength multiplexing (Ríos et al., 2019; Feldmann, 2017; Feldmann et al., 2020; Feldmann et al., 2021). More recent work further emphasizes electrically programmable, CMOS-compatible, low-loss, and endurance-improved PCM platforms, indicating a path toward scalable nonvolatile programmable photonic processors (Fang et al., 2022a; Chen et al., 2023; Wei and Saxena, 2024a; Gong et al., 2024; Fang et al., 2024; Sawant et al., 2025; Dwivedi et al., 2025; Pavanello et al., 2025; Wu et al., 2026). Therefore, PCMs are not merely an optional tuning layer; they are a key enabling material for bridging photonics high-throughput compute capability with the dense, persistent, and low-energy programmability required by practical AI hardware.

Besides synaptic weights for photonic neural networks (PNNs), PCMs can also be used to emulate functions such as neuron activation thanks to their plasticity (“Calculating with light using a). Once combined with photonic devices, e.g., microring resonators, this property enables to build neurons implementing an integrate and fire as well as non-linear behaviors (Chakraborty et al., 2018; Biasi et al., 2024). This concept has been used to build neuromorphic systems based on arrays of microring resonators leveraging wavelength multiplexing strategies (Chakraborty et al., 2019; Feldmann et al., 2019). Such systems have shown learning capabilities using rules like spiking time-dependent plasticity (STDP). PCM plasticity allows also to build systems whose learning rules are based on self-adaptation principles, leading to more biologically-plausible rules (Lugnan et al., 2025). Therefore, PCMs can be a key enabler for neuromorphic photonics by providing the possibility of building all the necessary building blocks to develop complex neuromorphic systems.

Although the vast majority of PCM-based demonstrations of neuromorphic photonic systems are based on chalcogenide PCMs, non-chalcogenide materials such as VO 2, V 2 O 3, and NbO 2 have also attracted considerable attention for their unique phase-change properties (Yang et al., 2011; O’Hara et al., 2014). These materials are characterized by a strong change in their optical and electrical properties similarly to chalcogenide PCMs, resulting from a metal to insulator transition (e.g., VO 2, V 2 O 3, and NbO 2) acquiring different crystal configurations. They have been often studied for their ultra-fast optical properties which do not require a reset step contrarily to GST225 and other chalcogenide PCMs because of their volatile behavior (O’Hara et al., 2014; Briggs et al., 2010; Navarro et al., 2021; Lee and Durand, 1984). Depending on the system update and operation frequency, this property can be of interest alongside their low switching temperatures (340 K) for building energy-efficient, rapidly reconfigurable systems (O’Hara et al., 2014). For example, VO 2 has been investigated for building spiking neurons and activation functions (Morcillo et al., 2024; Parra et al., 2026). However, these materials are highly lossy (especially in their metallic state), thus limiting their use in scalable neuromorphic systems compared to chalcogenide PCMs which allow to achieve a broader range of functions, notably low-loss synaptic weights and non-volatile signal summation. Besides, from an integration point of view, VO 2 large annealing temperature hinders the thermal budget available for CMOS back-end of line (BEOL) integration differently from other chalcogenides such as GST225, already available for BEOL integration (Haddad et al., 2022; Wei et al., 2024). Besides, its fabrication is much more challenging than for other chalcogenides such as GST225 because of the capability of Vanadium to form several stable oxides, e.g., V 2 O 3, VO 2, V 2 O 5, thus demanding a very tight window in terms of fabrication process conditions (deposition temperature, oxygen concentration…) to achieve high-quality VO 2 thin films with predictable properties (Ibrahim et al., 2026).

In what follows, we will therefore review the main properties of interest for PCMs, focusing on non-volatile, chalcogenide compounds due to their large-scale integration potential as well as their wider range of optical properties achievable, in photonics. We will present ellipsometric results of various PCM compounds showing good performance trade-offs between refractive index contrast and low optical absorption in both phases. Then, we will present novel photonic designs taking advantage of these properties by focusing on electrical activation through heaters as well as a comparison with other approaches leveraging similar PCM compounds. Afterwards, we will discuss experimental results of PCM patches integrated above silicon waveguides and actuated by optical pulses. We will also The combination of devices in photonic architectures based on electrical and optical actuation using different PCM compounds can provide a path towards achieving both static synaptic weighting (electrically) and dynamic synaptic reconfigurability as well as signal summation and activation functions (optically).

2 PCM properties

2.1 PCM compounds of interest for photonic applications

Metavalent bonding (MVB) has been recently identified as a distinctive bonding mechanism that characterizes the crystalline phase of PCMs. Unlike conventional solids, PCM crystals exhibit a unique combination of properties that place them between covalently bonded and metallically bonded materials (Wuttig et al., 2018; Raty et al., 2019; Raty and Wuttig, 2020; Raty et al., 2021; Wuttig et al., 2023). As a result, materials exhibiting metavalent bonding possess a characteristic set of features, including a high optical dielectric constant ( ϵ ∞ ), enhanced chemical bond polarizability reflected by large Born effective charges ( Z ∗ ), strong lattice anharmonicity associated with high Grüneisen parameters, and moderate electrical conductivity and optical reflectivity (Raty et al., 2019). Thus, PCMs represent a class of alloys distinguished by their unique and remarkable properties. From an optical perspective, they exhibit significant differences between their amorphous (covalent) and crystalline (metavalent) phases, making them ideal candidates for optical data storage technologies such as CD-RW, DVD-RAM, and Blu-ray disks (Noé et al., 2018a). The operation of PCM-based optical storage relies on the contrast in reflectivity and refractive index between the two phases. As illustrated in Figure 1, the amorphous phase is obtained by heating the material above its melting temperature ( T m ) and then rapidly cooling it (quenching). Conversely, the crystalline phase is achieved by annealing the amorphous material above its crystallization temperature ( T x ), followed by slow cooling. In recent years, PCMs have also gained considerable interest for non-volatile electronic memory applications. Similar to optical storage, where optical contrast is exploited, electronic memories take advantage of the large difference in electrical resistivity between the amorphous and crystalline states (Noé et al., 2018b). Among all PCMs, Ge 2 Sb 2 Te 5 (GST225) is the most extensively studied material. Located on the GeTe–Sb 2 Te 3 pseudo-binary line of the Ge–Sb–Te ternary phase diagram, GST225 is particularly attractive because of its large refractive index contrast across the visible, near-infrared (NIR), and mid-infrared (MIR) spectral ranges. This contrast arises when the material switches from the amorphous phase to the face-centered cubic (fcc) crystalline phase, typically induced by laser pulses.

2D ball-and-stick schematic representations of the amorphous and crystalline phases of a phase-change material. The amorphous phase is produced by heating the alloy above its melting temperature ( T m ) and subsequently rapidly cooling it (quenching). In contrast, the crystalline phase is obtained by annealing the amorphous material above its vitreous transition temperature T g (commonly at the so-called crystallization temperature ( T x )), followed by slow cooling to promote crystal formation.
Figure 1. 2D ball-and-stick schematic representations of the amorphous and crystalline phases of a phase-change material. The amorphous phase is produced by heating the alloy above its melting temperature ( T m ) and subsequently rapidly cooling it (quenching). In contrast, the crystalline phase is obtained by annealing the amorphous material above its vitreous transition temperature T g (commonly at the so-called crystallization temperature ( T x )), followed by slow cooling to promote crystal formation.

Figure 2 presents the reflectivity (a) and resistivity (b) of GST225 and GeTe thin films (100 nm) as a function of temperature. For GST225, crystallization occurs at approximately T x ≈ 150 °C–180 °C and is marked by an abrupt change in both reflectivity and resistivity. Prior to crystallization, the as-deposited amorphous GST225 exhibits lower reflectivity and higher resistivity than the crystalline fcc phase. For practical applications, the amorphous-to-crystalline transition must not only provide a strong optical and electrical contrast but also remain reversible over a large number of switching cycles (Noé et al., 2018a).

Temperature-dependent reflectivity at 670 nm **(a)** and four-point probe resistivity **(b)** of 100 nm GST225 (blue) and GeTe (red) thin films measured during heating at 10 °C/min. Crystallization is marked by a sharp change in reflectivity and resistivity, with T x ≈ 170 °C for GST225 and T x ≈ 230 °C for capped, non-oxidized GeTe films. Adapted with permission from Ref. (Noé et al., 2018a). Ⓒ IOP Publishing. All rights reserved license: CC BY 4.0.
Figure 2. Temperature-dependent reflectivity at 670 nm (a) and four-point probe resistivity (b) of 100 nm GST225 (blue) and GeTe (red) thin films measured during heating at 10 °C/min. Crystallization is marked by a sharp change in reflectivity and resistivity, with T x ≈ 170 °C for GST225 and T x ≈ 230 °C for capped, non-oxidized GeTe films. Adapted with permission from Ref. (Noé et al., 2018a). Ⓒ IOP Publishing. All rights reserved license: CC BY 4.0.

GST225 undergoes significant changes in its optical properties upon crystallization. At 1550 nm, it exhibits a large refractive index contrast ( Δ n ≈ 2.97 ), which is beneficial for optical modulation (Figure 3). However, crystallization also increases the extinction coefficient by about an order of magnitude (k ≈ 0.1–1.5), leading to high optical losses. Therefore, integrated photonic applications favor PCMs that offer a large refractive index change while maintaining a small increase in optical absorption, enabling efficient and low-loss optical tuning. To evaluate and compare different PCMs for photonic applications, a material figure of merit (FOM) is commonly used. This parameter provides a quantitative measure of the trade-off between optical modulation efficiency and insertion losses in optical devices (Soref et al., 1988). The FOM is defined as: FOM = Δ n / Δ k where Δ n = n c r − n a m and Δ k = k c r − k a m ≈ k c r, with n and k representing the refractive index and extinction coefficient of the crystalline ( c r ) and amorphous ( a m ) phases, respectively.

Wavelength-dependent refractive index n **(a)** and extinction coefficient k **(b)** of GST225 in the fcc crystalline phase (green) and the as-deposited amorphous phase (red). The significant contrasts in refractive index ( Δ n ) and extinction coefficient ( Δ k ) between the two phases are evident (see text for details).
Figure 3. Wavelength-dependent refractive index n (a) and extinction coefficient k (b) of GST225 in the fcc crystalline phase (green) and the as-deposited amorphous phase (red). The significant contrasts in refractive index ( Δ n ) and extinction coefficient ( Δ k ) between the two phases are evident (see text for details).

As shown in Figure 4, conventional GST225 thin films exhibit a FOM of approximately 2 at the telecommunication wavelength of 1550 nm. Higher FOM values have been reported for other chalcogenide alloys, including GeTe (FOM ≈ 6.2) and Ge 2 Sb 2 Se 4 Te 1 (FOM ≈ 4.2), indicating a more favorable balance between optical contrast and losses (Wuttig et al., 2017; Abdollahramezani et al., 2020; Zhang et al., 2018; Zhang et al., 2019).

Figure of merit (FOM) as a function of wavelength for some of the most widely studied PCMs. The comparison illustrates the spectral dependence of their optical performance and enables assessment of their suitability for photonic applications. Adapted from Ref. (Abdollahramezani et al., 2020). under the CC BY 4.0 licence. Modifications were made.
Figure 4. Figure of merit (FOM) as a function of wavelength for some of the most widely studied PCMs. The comparison illustrates the spectral dependence of their optical performance and enables assessment of their suitability for photonic applications. Adapted from Ref. (Abdollahramezani et al., 2020). under the CC BY 4.0 licence. Modifications were made.

A distinct class of ultra-low-loss PCMs, such as Sb 2 S 3 and Sb 2 Se 3 is of particular relevance (Yo et al., 2023). For these materials, the extinction coefficient remains close to zero in both the amorphous and crystalline states. Consequently, the FOM defined above becomes less meaningful because Δ k approaches zero, causing the denominator to vanish.

Interestingly, these materials not only exhibit negligible absorption changes (Δ k ≈ 0) but also a substantially smaller refractive index contrast ( Δ n ) compared with GST225 (Wuttig et al., 2018). This reduced optical contrast is attributed to the predominantly covalent bonding character of Sb 2 S 3 and Sb 2 Se 3 (Delaney et al., 2020). In contrast, GST225 benefits from the emergence of metavalent bonding in its crystalline phase, which is responsible for its significantly larger refractive index change and superior optical switching performance (Yo et al., 2023; Wuttig et al., 2018; Raty et al., 2019).

Metavalent bonding (MVB) plays a key role in determining the physical properties of PCMs. In particular, it endows the crystalline phase with an electronic conductivity that is several orders of magnitude higher than that of the amorphous, insulating phase (Raty et al., 2019). Furthermore, the high electronic polarizability associated with MVB results in elevated optical dielectric constants, which underpin the substantial differences in optical properties between the amorphous and crystalline states of these materials. At the boarder between covalent and metavalent materials lies the GeSe–GeTe pseudo-binary line (GeSe 1 − x Te x alloys), which represents an intriguing material system for investigation. This GeSe 1 − x Te x alloy series bridges GeSe, characterized by a covalently bonded orthorhombic crystalline structure and a relatively small variation in electronic properties upon crystallization at approximately 385 °C, and GeTe, a prototypical phase-change material whose rhombohedral crystalline phase is governed by metavalent bonding. Consequently, GeTe exhibits a remarkable contrast in physical properties between its amorphous and crystalline states, a hallmark of metavalent-bonded materials (Wuttig et al., 2018; Raty et al., 2019; Zhu et al., 2018; Kooi and Wuttig, 2020; Tomelleri et al., 2021; Ren et al., 2019). GeSe 1 − x Te x thin films are promising PCMs for low-loss reconfigurable photonics and nonlinear optical devices. Most compositions exhibit a large refractive index contrast ( Δ n > 1.7 ) with limited absorption changes, resulting in FOM values up to 550 percent higher than GST225 and up to three times greater than GSST2241 (Albanese et al., 2024). The optical properties can be tuned by adjusting the Se/Te ratio. GeSe offers ultra-low optical losses but a smaller refractive index contrast ( Δ n ≈ 0.2 ), making it a potential candidate for phase modulation applications despite challenges related to void formation after crystallization. Additionally, all GeSe 1 − x Te x alloys exhibit exceptionally high nonlinear Kerr coefficients and negligible two-photon absorption, outperforming many conventional nonlinear photonic materials. Combined with their thermal stability, these properties make them attractive for future nonlinear photonic technologies such as supercontinuum and frequency-comb generation. Ongoing device fabrication studies aim to evaluate their practical integration into photonic platforms (Albanese et al., 2024).

Nitrogen-doped GeSe 1 − x Te x thin films are also promising PCMs for low-loss reconfigurable photonics and nonlinear optical applications. Nitrogen incorporation improves thermal stability, increases crystallization temperatures, and refines the grain structure compared with undoped GeSe 1 − x Te x alloys. Among the investigated compositions, N-doped GeSe 0.4 Te 0.6 exhibits the best performance, combining a high figure of merit (FOM ≈ 11.6 at 1550 nm), a large refractive index contrast ( Δ n > 1.9 ), low optical losses, and a high crystallization temperature (T x ≈ 345 °C). Additionally, the material properties can be tailored by adjusting the Se/Te ratio and nitrogen content during deposition, making N-doped GeSe 0.4 Te 0.6 a strong candidate for future low-loss, non-volatile photonic devices operating at telecommunication wavelengths (Albanese et al., 2025).

2.2 Experimental characterization of refractive indices

To get insight into the optical properties of PCM thin films, Spectroscopic Ellipsometry (SE) is a powerful tool. SE is a non-destructive optical characterization technique used to determine the thickness and wavelength-dependent optical constants of thin films, namely, the refractive index ( n ) and extinction coefficient ( k ). It works by measuring changes in the polarization state of light reflected from a sample surface, which are described by the ellipsometric parameters ψ and Δ. These experimental data are fitted using optical models to accurately extract film thickness and optical properties (Fujiwara, 2007; Woollam et al., 1999). Variable Angle Spectroscopic Ellipsometry (VASE) extends this technique by performing measurements over a broad spectral range and at multiple angles of incidence. The additional information improves sensitivity and reduces parameter correlations, leading to more reliable and accurate determination of n, k, and thickness, particularly for complex or absorbing thin-film structures. SE and VASE are widely used for the characterization of PCMs and photonic thin films because they provide precise optical constants over a broad wavelength range. This information is essential for evaluating refractive index contrast, optical losses, and figures of merit, which are key parameters for designing reconfigurable photonic and optoelectronic devices. Their high sensitivity and non-contact nature make them particularly suitable for studying the optical differences between the amorphous and crystalline phases of PCMs. The optical properties of the amorphous phase are usually modeled using the Cody–Lorentz dispersion relation. This model is particularly suitable for amorphous semiconductors because it accurately describes the gradual absorption edge arising from structural disorder while maintaining Kramers–Kronig consistency. Owing to its ability to reproduce both sub-bandgap and above-bandgap optical transitions, the Cody–Lorentz model has become a standard approach in spectroscopic ellipsometry studies of amorphous chalcogenides and PCMs, enabling reliable extraction of refractive index, extinction coefficient, optical bandgap, and film thickness from experimental data (Orava et al., 2008; Němec et al., 2009; Němec et al., 2011; Dory et al., 2020; Dory et al., 2023; Albanese et al., 2024). For the crystalline phase of PCMs, a Drude term accounting for free-carrier absorption may be incorporated into the Cody–Lorentz (CL) model, in order to significantly enhancing the quality of the fit (Albanese et al., 2024). The fitting procedure are usually carried out using regression analysis based on minimization of the mean square error (MSE).

3 Energy-efficient multi-level devices

3.1 High-resolution, low-loss non-volatile phase shifters

High-resolution, multilevel, low-loss phase shifters are fundamental building blocks in programmable photonic integrated circuits, particularly for applications such as photonic neural networks (PNNs), optical signal processing, and coherent computing (Ahmed et al., 2025; Bogaerts et al., 2020; Sun et al., 2015). In these systems, precise control of the optical phase enables the accurate implementation of weighted operations, such as matrix–vector multiplications in Mach-Zehnder interferometer (MZI) meshes (Shen et al., 2017).

PCMs have emerged as a promising platform for realizing non-volatile, low-power phase shifters. Unlike thermo-optic or free-carrier based devices, PCM-based phase shifters retain their programmed state without continuous power consumption, significantly improving energy efficiency. They provide a large refractive-index contrast between the crystalline and amorphous phases, which is orders of magnitude higher than that of conventional refractive-index tuning mechanisms, thereby enhancing device compactness (Ríos et al., 2022; Brunner et al., 2025). Among PCMs, low-loss materials such as GeSe, Sb 2 Se 3, and Sb 2 S 3 have attracted attention because of their minimal absorption in the telecom band (Soref et al., 2015; Delaney et al., 2020; Teo et al., 2022; Fang et al., 2024), overcoming the limitations of high-loss materials such as GST (Sawant et al., 2025).

3.1.1 Multilevel phase encoding

A key requirement for advanced photonic computing is the ability to encode multiple discrete phase levels. This is typically achieved through partial amorphization of the PCM layer, where intermediate refractive indices correspond to intermediate phase shifts. Electrical programming using microheaters enables control of the amorphization fraction through pulse-width modulation (PWM) and pulse-amplitude modulation (PAM).

Spatially inhomogeneous temperature distributions within the PCM layer cross-section allow multiple intermediate amorphization levels. However, conventional designs based on uniform heaters suffer from abrupt phase transitions because of the accumulation of highly non-uniform phase shifts originating from homogeneous temperature distributions along the light propagation direction within the PCM layer (Dwivedi et al., 2026). As shown in Figure 5, for a 1550 nm laser wavelength, small variations in pulse parameters can produce disproportionately large phase shifts, thereby limiting resolution.

Variation of **(a)** phase shift and **(b)** difference between adjacent phase shift levels ( d ϕ ) with pulse duration. Adapted with permission from Ref. (Dwivedi et al., 2026). ⒸOptica Publishing Group license: OAPA, permissions obtained.
Figure 5. Variation of (a) phase shift and (b) difference between adjacent phase shift levels ( d ϕ ) with pulse duration. Adapted with permission from Ref. (Dwivedi et al., 2026). ⒸOptica Publishing Group license: OAPA, permissions obtained.

3.1.2 Segmented heater design for high resolution

To overcome these limitations, a segmented heater architecture has been proposed in which the heater width varies along the propagation direction, as shown in Figure 6. This design introduces a controlled thermal gradient, enabling different segments of the PCM to experience distinct levels of heating and, consequently, different degrees of amorphization. A highly non-uniform phase change within each segment, when averaged across multiple segments with varying degrees of amorphization, results in an overall uniform amorphization process (Dwivedi et al., 2026).

**(a)** Cross-section schematic of the phase-shifter structure, including copper/tungsten thermal vias for enhanced cooling of GeSe. **(b)** Schematic illustrating the increasing heater width along the propagation direction and the via placement for electrical connections. Adapted with permission from Ref. (Dwivedi et al., 2026). ⒸOptica Publishing Group license: OAPA, permissions obtained.
Figure 6. (a) Cross-section schematic of the phase-shifter structure, including copper/tungsten thermal vias for enhanced cooling of GeSe. (b) Schematic illustrating the increasing heater width along the propagation direction and the via placement for electrical connections. Adapted with permission from Ref. (Dwivedi et al., 2026). ⒸOptica Publishing Group license: OAPA, permissions obtained.

As illustrated in Figure 7a, increasing the heater width reduces local power dissipation and temperature, resulting in a spatially varying phase-change profile along the propagation direction. The cumulative effect is a smooth overall amorphization (Figure 7b) and a nearly continuous phase-shift response (Figure 7c), despite the inherently discrete and highly non-uniform switching behavior of individual segments.

Variation of **(a)** heater power and maximum temperature across TiN with heater width. Variation of **(b)** amorphous fraction and **(c)** phase shift ( ϕ ) with pulse duration for the combined GeSe segments. Adapted with permission from Ref. (Dwivedi et al., 2026). ⒸOptica Publishing Group license: OAPA, permissions obtained.
Figure 7. Variation of (a) heater power and maximum temperature across TiN with heater width. Variation of (b) amorphous fraction and (c) phase shift ( ϕ ) with pulse duration for the combined GeSe segments. Adapted with permission from Ref. (Dwivedi et al., 2026). ⒸOptica Publishing Group license: OAPA, permissions obtained.

This approach enables up to 164 distinguishable phase levels (greater than 7-bit resolution) and smooth phase variation across the range from 0 to π, with low insertion loss (approximately 0.6 dB in the worst case) (Dwivedi et al., 2026). Compared with conventional constant-width heaters, the segmented design significantly reduces the variation between adjacent phase levels ( d ϕ ), improving linearity and programmability.

The combination of low loss, non-volatility, and high resolution positions PCM-based phase shifters with heater geometry that varies along the light propagation direction as strong candidates for next-generation programmable photonic systems. The proposed design complies with standard foundry constraints, employing a minimum heater width of 1 µm with incremental increases of 50 nm. The average power dissipation across the segmented heater is approximately 0.94 W, while the programming pulse duration ranges from 142 to 1015 ns? This corresponds to an energy consumption ranging from 133 nJ to 955 nJ, enabling access to multiple phase-shift levels between the fully crystalline and fully amorphous states of GeSe.

Although reducing the heater width could further lower power consumption, the current configuration is particularly well suited for applications characterized by low reconfiguration rates, where the energy cost is dominated by weight programming. Importantly, the segmented-heater concept is not restricted to this specific platform and can be extended to other waveguide geometries and PCMs, potentially allowing reduced-power operation under alternative or relaxed design constraints.

Overall, the design represents a trade-off between optical performance, thermal efficiency, and fabrication limitations. Although additional optimization is feasible, it inherently involves compromises. For instance, shortening the device length by increasing the PCM thickness or enhancing modal overlap would lead to higher insertion loss, as indicated by the simulations. Similarly, modifying the heater geometry could improve heating efficiency and cooling dynamics, but it would require careful co-optimization to maintain both optical and thermal performance.

3.2 Hybrid mode interferometer for multilevel transmission

While standalone phase shifters provide phase control, achieving compact multilevel transmission typically requires interferometric architectures such as MZIs or resonators (Ríos et al., 2022; Chen et al., 2023; Sun et al., 2025; Pang et al., 2025; Chen et al., 2024; Yang et al., 2023; Gong et al., 2024). The hybrid mode interferometer (HMI) represents an alternative approach that directly integrates phase-change modulation with modal interference in a single compact structure (Dwivedi et al., 2025).

3.2.1 Operating principle: hybridized mode interference

The HMI operates by exploiting interference between hybridized transverse modes in a silicon waveguide integrated with a PCM layer at a 1550 nm laser wavelength. By utilizing the vertical asymmetry in the structure, the waveguide is specifically engineered to support hybridization between the fundamental TM mode (TM 0) and the first higher-order TE mode (TE 1). At optimized waveguide dimensions (e.g., width ≈ 0.71 µm), these two modes exhibit comparable TE polarization fractions ( ∼ 42 % ) (Dwivedi et al., 2025).

The interferometric device structure shown in Figure 8 consists of a PCM-integrated waveguide interconnected between two tapered SOI waveguides. This configuration facilitates coupling to the hybridized modes from the input taper and enables observation of the interference-induced intensity variation at the output taper. The transmission can be approximated as:where the output depends on the phase difference accumulated between the two hybrid modes with propagation constants β 1 / 2 and coupling coefficients (between fundamental input/output mode and hybrid modes) C 1 / 2 over a device length L (Dwivedi et al., 2025).

T ≈ C 1 e i β 1 L + C 2 e i β 2 L 2

Schematic of the hybrid mode interferometer based on GeSe integrated onto a silicon waveguide. Adapted from Ref. (Dwivedi et al., 2025). under the CC BY 4.0 licence.
Figure 8. Schematic of the hybrid mode interferometer based on GeSe integrated onto a silicon waveguide. Adapted from Ref. (Dwivedi et al., 2025). under the CC BY 4.0 licence.

The PCM layer (e.g., GeSe) introduces a state-dependent perturbation to the modal effective indices. In the crystalline state, while unequal modal coupling arising from imperfect hybridization is achieved, the length of the PCM integrated waveguide is selected such that it allows constructive interference to provide high transmittance with ultra-low insertion loss (< 0.1 dB). In the amorphous state, nearly equal modal amplitudes arise from near-perfect hybridization, resulting in very low transmittance because the phase difference approaches π (destructive interference). This mechanism enables a high extinction ratio (∼ 14 dB) within a compact footprint (∼ 24 µm length), making it suitable for analog encoding.

3.2.2 Multilevel transmission via partial amorphization of GeSe

A distinctive advantage of the HMI is its ability to achieve multilevel transmission without requiring external interferometers. By gradually increasing the amorphous fraction of the PCM layer (e.g., via segmented heating or incremental switching (Dwivedi et al., 2026)), the device exhibits continuous variation in both amplitude and phase, as shown in Figure 9. It provides multi-bit storage capability with nearly linear transmission scaling over a significant range and sub-μs switching speeds. In the demonstrated design, the transmittance varies from ∼0.98 to ∼0.04 with a contrast of ≈ 14 dB, making it suitable for analog encoding.

Variation of **(a)** transmittance and **(b)** phase with the length of the amorphous region. Reproduced from Ref. (Dwivedi et al., 2025). under the CC BY 4.0 licence.
Figure 9. Variation of (a) transmittance and (b) phase with the length of the amorphous region. Reproduced from Ref. (Dwivedi et al., 2025). under the CC BY 4.0 licence.

The HMI concept eliminates the need for separate interferometric structures, reducing system complexity and footprint once compared with more common ring resonator and MZI employing PCM phase shifter configurations (Dwivedi et al., 2025). The device offers several advantages, including a compact footprint, low insertion loss, and the ability to achieve intrinsic non-volatile multilevel transmission. However, it also presents certain limitations. Its performance relies on precise modal engineering and is sensitive to variations in waveguide dimensions. Additionally, there is a trade-off between contrast and insertion loss, which depends on the device length.

The combination of high-resolution PCM phase shifters and hybrid mode interferometers represents a powerful paradigm for analog photonic signal processing. While segmented heater designs address the challenge of precise phase encoding, HMIs provide a compact and efficient route to directly translating phase changes into multilevel transmission states. Together, these approaches enable scalable, low-power, and high-density implementations of photonic memory and computing systems, with strong potential for applications in neuromorphic and in-memory photonic architectures.

3.3 Alternative approaches for phase and amplitude modulation

This section describes alternative solutions that have been adopted to implement PCM-based devices that can serve as synaptic weights in PNNs. In particular, the main designs investigated for amplitude and phase modulation consists of straight waveguides, which can be then integrated as interferometric arms in MZIs, and of ring resonators. For both categories, the PCM introduces a phase shift either in an MZI arm or inside the ring resonator through evanescent mode interaction with the PCM layer in proximity of the waveguide core. Approaches based on MZIs have shown good multilevel capability, e.g., . However, even though phase shifters as compact as ≈ 11-μ m-long based on Sb 2 Se 3 have been reported (Ríos et al., 2022), their integration in an MZI configuration for amplitude modulation increases their footprint considerably (e.g., hundreds to thousands of μ m 2) because of the presence of input/output couplers and physical design constraints such as cross-talk between arms. Ring resonator approaches solve these issues because, due to the field enhancement effect, they can achieve more compact footprints while not requiring input/output couplers and interferometric arms. However, only a limited number of levels can be readily encoded due to the strong phase sensitivity of the output response which does not favor a tailored, well distributed encoding. For example, Fang et al. were able to achieve up to only 14 distinct optical levels using a graphene-actuated Sb 2 Se 3-based ring resonator even though the microheater design resulted in a highly energy-efficient actuation (Fang et al., 2022b).

Table 1 compares the performance of the HMI with conventional architectures based on MZIs and ring resonators. It can be noted that the HMI solution described in the previous section provides a good trade-off between compactness, possibility to encode many levels, and low insertion losses which are critical for system scalability.

TABLE 1. Comparison of the HMI-based structure with conventional architectures.

ArchitectureFootprintMultilevel capabilityLoss
MZI + PCM phase shifter (Ríos et al., 2015; Chen et al., 2023; Pang et al., 2025; Chen et al., 2024; Yang et al., 2023; Gong et al., 2024)LargeYesLow–moderate
Ring resonator (Ríos et al., 2015; Fang et al., 2022b; Sun et al., 2025; Chen et al., 2024)CompactLimitedWavelength sensitive
HMI (Dwivedi et al., 2026)Very compactIntrinsic multilevel< 0.1 dB
( ∼ 24 μ m 2 )​​

4 Optical device characterization

In the following, we describe standard techniques used for characterization of PCM-based devices. Then, we present some recently published experimental results on GST225 using the interferometric approach described below to extract both amplitude and phase as a function of PCM crystallization. All the experimental data presented in this section were conducted on SOI waveguides integrated with a layer of GST225, a reference for microelectronic foundries. Other families of PCMs have been explored in Ref (Albanese et al., 2024). and Ref (Albanese et al., 2025). but are still not yet integrated onto photonic circuits. To couple light into the chip through the waveguide, grating couplers are patterned on either sides. Light is injected into and collected from the grating couplers using optical fibers as shown in Figure 10. The sources of losses are grating coupling losses ( η ), interface losses at the edges of the PCM patch ( κ ) owing to the refractive index change along transmission and absorption loss through the PCM patch ( α ). Therefore, the linear transmission is given by the relation depicted in Figure 10. The transmission for different lengths L of PCM patches are recorded. By plotting log(Pout/Pin) vs. patch length L, propagation loss through PCM ( α ) can be determined. To do so, continuous wave laser at 1550 nm wavelength is transmitted through the waveguide. As the grating couplers are designed to couple the Transverse Electric (TE) polarization into the waveguide, a polarization controller is utilized at the insertion fiber to maximize the input through the waveguide. For different lengths of PCM patches ranging from 1 μ m to 3 μ m, output power at the collection fiber is recorded using a power meter. The log of transmission is plotted as a function of patch length as shown in Figure 11. SOI waveguides integrated with GST225 are found to have ≈ 5dB/μ m loss.

Schematic view of the coupling scheme into the silicon waveguide decorated with phase change materials patches. Adapted with permission from Ref. (Sawant et al., 2025). Copyright Ⓒ 2025 Wiley-VCH GmbH license: CC BY 4.0.
Figure 10. Schematic view of the coupling scheme into the silicon waveguide decorated with phase change materials patches. Adapted with permission from Ref. (Sawant et al., 2025). Copyright Ⓒ 2025 Wiley-VCH GmbH license: CC BY 4.0.
SOI waveguide transmission for different GST225 patch lengths. The dashed line shows the linear fit providing an estimation of linear losses. Adapted with permission from Ref. (Sawant et al., 2025). Copyright Ⓒ 2025 Wiley-VCH GmbH license: CC BY 4.0.
Figure 11. SOI waveguide transmission for different GST225 patch lengths. The dashed line shows the linear fit providing an estimation of linear losses. Adapted with permission from Ref. (Sawant et al., 2025). Copyright Ⓒ 2025 Wiley-VCH GmbH license: CC BY 4.0.

4.1 Characterization techniques for PCM-based devices

In general, both real and imaginary parts of the effective refractive index of an optical mode are strongly modified by the addition of PCMs to optical waveguides. Therefore, techniques capable to extract both of these contributions are of the utmost importance to allow careful design and operation of PCM-based devices as well as accurate data analysis. A well-known technique consists in placing the PCM-based device in one arm of an interferometer as shown in Figure 12. To switch the state of the PCM layer and probe it, programming optical pulses and probe signal (read) are then sent through the waveguide in the counter-propagating direction (see Figure 12). One cycle of programming consists of sending a short (400 ps) and intense amorphization pulse (write), followed 25 μs later by a longer (∼1 μs) but less intense pulse to recrystallize and reset the PCM state (erase). Owing to non-zero extinction coefficient of GST225 in both amorphous and crystalline phases, the optical pulses are absorbed providing the thermal energy for the phase transition. The programming pulses are tailored from a CW laser emitting in our experiments at 1552 nm using standard telecommunication fiber optic components, modulated by an electro-optic modulator (EOM) driven by an Arbitrary Waveforme Generator (AWG) and finally amplified with an erbium doped fiber amplifier (EDFA). The phase-state change of the patches is then detected in real time thanks to an optical heterodyne detection based on an MZI and excited by a low power probe (P = 0.6 mW in our experiments) CW laser emitting at 1550 nm. As shown in Figure 12, the sample is sandwiched between circulators that allow the counter-propagative combination of the write and read lasers inside the sample. A spectral filter is added to the read laser path after the sample to suppress the residual programming pulses reflected from the sample and passing through the circulator. The other arm is the reference arm, which contains an acousto-optic modulator (AOM) driven by a local oscillator (LO) that shifts the laser wavelength frequency, and a delay line that synchronizes the optical paths of the read signal and reference arms to minimize phase noise. The interference between the two arms is split with a 50:50 coupler and detected using a balanced amplified photodiode that maximizes the contrast of the interferometric signal x(t). This signal is further fed to a lock-in amplifier, where it is mixed with the LO signal to produce two separate outputs: relative changes in transmittance ( Δ T ) and phase ( Δ ϕ ). While this technique is highly accurate thanks to its tunable delay line and heterodyne detection, it requires also a bulky setup. Therefore, other techniques where the PCM-based device is placed into integrated on-chip MZIs or ring resonators have also been investigated (Ríos et al., 2022; Fang et al., 2022b; Yang et al., 2023; Brunner et al., 2025). These approaches allow to extract the amplitude and phase contributions introduced by the PCM presence through the use of wavelength spectroscopy, i.e., acquiring the transmittance spectrum of an MZI or a ring resonator as a function of wavelength. This enables the extraction, from fitting the transmittance spectra (leveraging the acquisition over multiple transmittance peaks/dips), of the insertion loss and phase-shift (which are linked to both real and imaginary parts of the refractive index) introduced by the device for different PCM crystallization states. Complementary measurements can also be used to extract the real and imaginary parts of the effective refractive index by using nominally equivalent devices, e.g., a simple transmittance measurement of a phase-shifter allows to extract the insertion loss which is mostly related to the imaginary part of the effective index (for example, scattering losses can be affected by the real part as well), while a measurement of the resonance shift in a ring resonator can provide the information on the real part of the effective index. This allows to reduce the fitting errors, however it requires that multiple PCM-based devices can be fabricated in a highly close manner in terms of optical, thermal, and electrical (if actuated by heaters) behavior. Besides, the devices need to be driven and reply to stimuli in the exact same manner. Ultimately, accurate device design simulations and cross-section measurements are needed to extract the standalone PCM properties.

**(a)** Programming pulse generation setup, AWG -Arbitrary Waveform Generator, EOM -Electro-Optic Modulator, EDFA -Erbium Doped Fiber Amplifier. **(b)** A sequence of amorphization (write) and crystallization (erase) pulses from the pulse generation setup. **(c)** Heterodyne interferometer setup to detect transmission and optical phase change: AOM - Acousto-Optic Modulator, LO - Local Oscillator, Filt.- Spectral filter, D0, D1, D2 - Photodiodes. Reproduced with permission from Ref. (Sawant et al., 2025). Copyright Ⓒ 2025 Wiley-VCH GmbH license: CC BY 4.0.
Figure 12. (a) Programming pulse generation setup, AWG -Arbitrary Waveform Generator, EOM -Electro-Optic Modulator, EDFA -Erbium Doped Fiber Amplifier. (b) A sequence of amorphization (write) and crystallization (erase) pulses from the pulse generation setup. (c) Heterodyne interferometer setup to detect transmission and optical phase change: AOM - Acousto-Optic Modulator, LO - Local Oscillator, Filt.- Spectral filter, D0, D1, D2 - Photodiodes. Reproduced with permission from Ref. (Sawant et al., 2025). Copyright Ⓒ 2025 Wiley-VCH GmbH license: CC BY 4.0.

4.2 Write and read operations

Typical measurements obtained with the interferometric setup (see Figure 12) are presented in Figure 13 for a GST patch with a length of 1.5 μ m. The two output signals are reported together with the sequence of programming pulses measured on photodetector D 0. These programming pulses induce large, reversible, and nonvolatile modifications of both the transmitted intensity and the optical phase. In particular, amorphization pulses with a duration of ≈ 400 p s and an energy of approximately ≈ 100 p J, increase the transmittance by about ( Δ T = 60 % ), while reducing the optical phase delay by ( Δ ϕ = 55 ° ). These changes are consistent with the decrease of both the refractive index and the extinction coefficient when GST switches from its crystalline to its amorphous phase. The transmittance variation is further amplified by a slight modification of the mode mismatch between the waveguide sections covered and not covered by the PCM patch.

Transmission and optical phase change recorded upon exposure to programming pulses, i.e., amorphization (write) and crystallization (erase) pulses on GST of patch length 1.5 µm. Reproduced with permission from Ref. (Sawant et al., 2025). Copyright © 2025 Wiley-VCH GmbH license: CC BY 4.0.
Figure 13. Transmission and optical phase change recorded upon exposure to programming pulses, i.e., amorphization (write) and crystallization (erase) pulses on GST of patch length 1.5 µm. Reproduced with permission from Ref. (Sawant et al., 2025). Copyright © 2025 Wiley-VCH GmbH license: CC BY 4.0.

4.3 Reliability study

A key parameter for the practical implementation of PCM-based photonic devices is their endurance, defined as the number of write–erase cycles that can be performed without significant degradation of the optical response. Since excess energy delivered to the PCM may adversely affect the cycling performance, the endurance of the GST patch was evaluated after optimizing the pulse energies required for amorphization and crystallization. The write-erase sequence was repeated 1 0 9 times using the optimized pulse energies of 100 p J and 7.4 n J, selected to maximize the transmission contrast. In practice, this corresponds to a 14-h endurance test, during which the PCM patch was repeatedly melted, quenched, and recrystallized every 50 μ s using optical programming pulses. To avoid saturation of the acquisition-card memory, the transmitted amplitude and phase through the PCM patch were not recorded continuously, but only over selected programming sequences, sampled every 1 0 N cycles, as shown in Figure 14.

Transmitted amplitude and phase changes through the silicon-decorated PCM patches recorded each 1 0 N sequences, with N ranging from 1 to 9, up to 1 billion. Reproduced with permission from Ref. (Sawant et al., 2025). Copyright Ⓒ 2025 Wiley-VCH GmbH license: CC BY 4.0.
Figure 14. Transmitted amplitude and phase changes through the silicon-decorated PCM patches recorded each 1 0 N sequences, with N ranging from 1 to 9, up to 1 billion. Reproduced with permission from Ref. (Sawant et al., 2025). Copyright Ⓒ 2025 Wiley-VCH GmbH license: CC BY 4.0.

5 Conclusion

PCMs have emerged as key enabling materials for reconfigurable and non-volatile photonic devices owing to their large and reversible optical property changes upon phase transition. While GST225 remains the benchmark PCM, alternative compositions such as GeTe, Ge2Sb2Se4Te1, Sb2S3, Sb2Se3, and GeSe 1 − x Te x alloys have demonstrated improved trade-offs between refractive-index contrast and optical losses. In particular, GeSe 1 − x Te x and nitrogen-doped GeSe 1 − x Te x alloys offer a promising combination of high figures of merit, low absorption, and enhanced thermal stability. Continued investigation of the relationship between bonding mechanisms, especially metavalent bonding, and optical performance will be essential for the rational design of next-generation low-loss photonic PCMs and devices.

Spectroscopic ellipsometry (SE) and variable angle spectroscopic ellipsometry (VASE) are key techniques for the characterization of phase-change material thin films. By providing accurate measurements of film thickness and optical constants, they enable the evaluation of refractive-index contrast, optical losses, and figures of merit. Combined with appropriate optical models, such as Cody–Lorentz and Drude–Cody–Lorentz, these methods provide reliable insight into the optical properties of both amorphous and crystalline phases, supporting the development of advanced photonic devices.

Integrating low-loss PCMs onto silicon waveguides enables non-volatile control of optical mode propagation by tuning the degree of amorphization or crystallization. However, precise encoding of multiple phase-shift levels requires controlled and reproducible amorphization of the PCM. This, in turn, demands carefully engineered heater geometries and appropriate cooling dynamics to achieve well-defined intermediate material states. A segmented heater design provides an effective route to smoother phase control by spatially tailoring the PCM amorphization profile along the light propagation direction. In such a configuration, different PCM sections experience different thermal conditions, resulting in a more gradual and controllable phase-shift response. While PCM-based phase shifters are directly suitable for coherent photonic neural networks, certain applications, such as photonic memories and dot-product engines in tensor-core architectures, require PCM programming to be directly translated into transmission modulation. Conventional approaches usually convert programmed phase shifts into transmission levels using interferometric or resonant structures, such as Mach - Zehnder interferometers or ring resonators. However, these additional components increase the overall footprint and can introduce extra optical losses. In contrast, a hybrid mode interferometer directly maps the PCM-induced refractive-index change into transmission modulation by exploiting high-contrast interference between hybridized TE - TM modes. This HMI approach enables compact, low-loss, intrinsic multilevel transmission without requiring an external interferometric architecture.

Finally, GST225 patches integrated onto SOI waveguides were characterized under all-optical programming and readout conditions. The propagation losses associated with the PCM-covered waveguide were estimated to be approximately 5 dB/μm at 1550 nm. A heterodyne interferometric setup was used to monitor the transmission and phase variations induced by successive amorphization and crystallization pulses. For a 1.5 µm-long GST225 patch, 400 ps amorphization pulses with an energy of approximately 100 pJ produced a transmission increase of about 60% and a phase-delay reduction of 55°. The reversibility and endurance of the device were evaluated using optimized write and erase pulse energies of 100 pJ and 7.4 nJ, respectively. The write–erase sequence was repeated up to 1 0 9 times over a 14-h cycling test, providing a quantitative assessment of the device response under repeated optical programming.

Energy-efficient, high-resolution, and reliable devices leveraging platforms integrating different PCM compounds can provide the missing ingredient to enable the next-generation of computing systems where all the key functions can be addressed by using either optical or electrical stimuli depending on the exact required function.

Generative AI statement

The author(s) declared that generative AI was not used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Footnotes

Edited by: Ripalta Stabile, Eindhoven University of Technology, Netherlands

Reviewed by: Lin Chen, University of Shanghai for Science and Technology, China

Dezhi Tan, Zhejiang University, China

Statements

Conflict of interest

The author(s) declared that this work was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Funding

The author(s) declared that financial support was received for this work and/or its publication. Horizon Europe research and innovation (101070238). All optical tests were conducted at the facilities of the SMARTLIGHT platform, funded by France 2030 (EQUIPEX + contract ANR-21-ESRE-0040) and Région Bourgogne Franche-Comté, and at the facilites of CEA-LETI (ellipsometry).

Author contributions

RD: Conceptualization, Investigation, Writing – original draft, Writing – review and editing. CP: Conceptualization, Investigation, Writing – original draft, Writing – review and editing. AA: Conceptualization, Investigation, Writing – original draft, Writing – review and editing. AY: Writing – original draft, Writing – review and editing. RO: Supervision, Writing – original draft, Writing – review and editing. GB: Writing – original draft, Writing – review and editing. RS: Writing – original draft, Writing – review and editing. MT: Writing – original draft, Writing – review and editing. YB: Writing – original draft, Writing – review and editing. AR: Writing – original draft, Writing – review and editing. AC: Writing – original draft, Writing – review and editing. BCh: Funding acquisition, Project administration, Writing – original draft, Writing – review and editing. PN: Funding acquisition, Supervision, Writing – original draft, Writing – review and editing. BCl: Funding acquisition, Supervision, Writing – original draft, Writing – review and editing. FP: Funding acquisition, Supervision, Writing – original draft, Writing – review and editing.

Acknowledgements

Views and opinions expressed are, however, those of the author(s) only and do not necessarily reflect those of the European Union. Neither the European Union nor the granting authority can be held responsible for them. French authors also thank the CNRS research group “GdR CHALCO”, for fruitful discussions on PCM materials in the context of their annual workhops (https://gdrchalco.cnrs.fr/).

References

  1. Abdollahramezani S., Hemmatyar O., Taghinejad H., Krasnok A., Kiarashinejad Y., Zandehshahvar M. (2020). Tunable nanophotonics enabled by chalcogenide phase-change materials. Nanophotonics 9 (5), 1189 – 1241. doi:10.1515/nanoph-2020-0039
  2. Ahmed S. R., Baghdadi R., Bernadskiy M., Bowman N., Braid R., Carr J. (2025). Universal photonic artificial intelligence acceleration. Nature 640 (8058), 368 – 374. doi:10.1038/s41586-025-08854-x
  3. Albanese A., Tomelleri M., Karam L., Dory J.-B., Licitra C., Charbonnier B. (2024). Optical properties of GeSe1Te chalcogenide materials promising for on-Chip low and ultra-low loss reconfigurable photonics and nonlinear devices. Physica Status Solidi (RRL) – Rapid Res. Lett. 18 (10), 2400129. doi:10.1002/pssr.202400129
  4. Albanese A., Tomelleri M., Dory J.-B., Karam L., Licitra C., Coillet A. (2025). Nitrogen-doped GeSe1-xTex chalcogenide thin films: a candidate for On-Chip low-loss reconfigurable photonics and nonlinear device applications. Physica Status Solidi (RRL) – Rapid Res. Lett. 19 (7), 2400413. doi:10.1002/pssr.202400413
  5. Biasi S., Donati G., Lugnan A., Mancinelli M., Staffoli E., Pavesi L. (2024). Photonic neural networks based on integrated silicon microresonators. Intell. Comput. 3, 0067. doi:10.34133/icomputing.0067
  6. Bogaerts W., Pérez D., Capmany J., Miller D. A. B., Poon J., Englund D. (2020). Programmable photonic circuits. Nature 586 (7828), 207 – 216. doi:10.1038/s41586-020-2764-0
  7. Briggs R. M., Pryce I. M., Atwater H. A. (2010). Compact silicon photonic waveguide modulator based on the vanadium dioxide metal-insulator phase transition. Opt. Express 18 (11), 192 – 201. doi:10.1364/OE.18.011192
  8. Brunner D., Shastri B. J., Al-Qadasi M. A., Ballani H., Barbay S., Biasi S. (2025). Roadmap on Neuromorphic Photonics.
  9. Chakraborty I., Saha G., Sengupta A., Roy K. (2018). Toward fast neural computing using all-photonic phase change spiking neurons. Sci. Rep. 8 (1), 12980. doi:10.1038/s41598-018-31365-x
  10. Chakraborty I., Saha G., Roy K. (2019). Photonic In-Memory computing primitive for spiking neural networks using phase-change materials. Phys. Rev. Appl. 11 (1), 014063. doi:10.1103/physrevapplied.11.014063
  11. Chen R., Fang Z., Perez C., Miller F., Kumari K., Saxena A. (2023). Non-volatile electrically programmable integrated photonics with a 5-bit operation. Nat. Commun. 14, 3465. doi:10.1038/s41467-023-39180-3
  12. Chen R., Tara V., Choi M., Dutta J., Sim J., Ye J. (2024). Deterministic quasi-continuous tuning of phase-change material integrated on a high-volume 300-mm silicon photonics platform. Npj Nanophot. 1, 7. doi:10.1038/s44310-024-00009-6
  13. Delaney M., Zeimpekis I., Lawson D., Hewak D. W., Muskens O. L. (2020). A new family of ultralow loss reversible phase-change materials for photonic integrated circuits: Sb2S3 and Sb2Se3. Adv. Funct. Mater. 30 (36), 2002447. doi:10.1002/adfm.202002447
  14. Dory J.-B., Castro-Chavarria C., Verdy A., Jager J.-B., Bernard M., Sabbione C. (2020). Ge–Sb–S–Se–Te amorphous chalcogenide thin films towards on-chip nonlinear photonic devices. Sci. Rep. 10 (1), 11894. doi:10.1038/s41598-020-67377-9
  15. Dory J.-B., Ibnoussina M., Raty J.-Y., Jager J.-B., Verdy A., Coillet A. (2023). Origin of the unusual high optical nonlinearities observed in glassy chalcogenides. Adv. Opt. Mater. 11 (24), 2301154. doi:10.1002/adom.202301154
  16. Dwivedi R., Pavanello F., Orobtchouk R. (2025). An ultra-low-loss compact phase-change material-based hybrid-mode interferometer for photonic memories. Apl. Mater. 13 (4), 041123. doi:10.1063/5.0245618
  17. Dwivedi R., Yadav A., Orobtchouk R., Charbonnier B., Malhouitre S., Noé P. (2026). High-resolution, low-loss multilevel phase shifters based on electrically driven phase-change materials integrated onto silicon waveguides. Opt. Express 34 (11), 896 – 908. doi:10.1364/OE.592015
  18. Fang Z., Chen R., Zheng J., Saxena A., Whitehead J., Majumdar A. (2022a). Ultra-low-energy programmable non-voilatile silicon photonics based on phase-change materials with graphene heaters. Nat. Nanotechnol. 17, 842 – 848. doi:10.1038/s41565-022-01153-w
  19. Fang Z., Chen R., Zheng J., Khan A. I., Neilson K. M., Geiger S. J. (2022b). Ultra-low-energy programmable non-volatile silicon photonics based on phase-change materials with graphene heaters. Nat. Nanotechnol. 17 (8), 842 – 848.
  20. Fang Z., Mills B., Chen R., Zhang J., Xu P., Hu J. (2024). Arbitrary programming of racetrack resonators using low-loss phase-change material Sb2Se3. Nano Lett. 24, 97 – 103. doi:10.1021/acs.nanolett.3c03353
  21. Feldmann J., Stegmaier M., Gruhler N., Ríos C., Bhaskaran H., David Wright C. (2017). Calculating with light using a chip-scale all-optical abacus. Nat. Commun. 8, 1256. doi:10.1038/s41467-017-01506-3
  22. Feldmann J., Youngblood N., Wright C. D., Bhaskaran H., Pernice W. H. P. (2019). All-optical spiking neurosynaptic networks with self-learning capabilities. Nature 569 (7755), 208 – 214. doi:10.1038/s41586-019-1157-8
  23. Feldmann J., Youngblood N., Li X., Wright C. D., Bhaskaran H., Pernice W. H. P. (2020). Integrated 256 cell photonic phase-change memory with 512-bit capacity. IEEE J. Sel. Top. Quantum Electron. 26 (2), 1 – 7. doi:10.1109/jstqe.2019.2956871
  24. Feldmann J., Youngblood N., Karpov M., Gehring H., Li X., Stappers M. (2021). Parallel convolutional processing using an integrated photonic tensor core. Nature 589 (7840), 52 – 58. doi:10.1038/s41586-020-03070-1
  25. Fu T., Zhang J., Sun R., Zhang R., Liang X., Sun Y. (2024). Optical neural networks: progress and challenges. Light Sci. and Appl. 13 (1), 263. doi:10.1038/s41377-024-01590-3
  26. Fujiwara H. (2007). Spectroscopic Ellipsometry: Principles and Applications. John Wiley and Sons.
  27. Gong J., Xia J., Wang T., Li Z., Wang Z., Dong Y. (2024). Six bit optical phase states realized in nonvolatile phase shifter based on N-doped Sb2Se3. ACS Photonics 11, 4334 – 4341.
  28. Haddad E., Kruzelecky R. V., Murzionak P., Jamroz W., Tagziria K., Chaker M. (2022). Review of the VO2 smart material applications with emphasis on its use for spacecraft thermal control. Front. Mater. 9, 1013848. doi:10.3389/fmats.2022.1013848
  29. Harris N. C., Carolan J., Bunandar D., Prabhu M., Hochberg M., Baehr-Jones T. (2018). Linear programmable nanophotonic processors. Optica 5 (12), 1623 – 1631. doi:10.1364/optica.5.001623
  30. Horowitz M. (2014). “ Computing’s energy problem (and what we can do about it),” in 2014 IEEE International Solid-State Circuits Conference Digest of Technical Papers (ISSCC) (IEEE), 10 – 14.
  31. Ibrahim M., Xiao Z., Xu J. (2026). Effect of vanadium layer thickness on the phase transition of fabricated vanadium dioxide thin films. Sci. Rep. 16 (1), 20734. doi:10.1038/s41598-026-50768-9
  32. Kooi B. J., Wuttig M. (2020). Chalcogenides by design: functionality through metavalent bonding and confinement. Adv. Mater. 32 (21), 1908302. doi:10.1002/adma.201908302
  33. Lee J. C., Durand W. W. (1984). Electrically stimulated optical switching of NbO2 thin films. J. Appl. Phys. 56 (11), 3350 – 3352. doi:10.1063/1.333863
  34. Leiserson C. E., Thompson N. C., Emer J. S., Kuszmaul B. C., Lampson B. W., Sanchez D. (2020). There’s plenty of room at the top: what will drive computer performance after moore’s law? Science 368 (6495), eaam9744. doi:10.1126/science.aam9744
  35. Li T., Gao S., Xie Z., Xu X., Zhang X. (2023). Neuromorphic photonics based on phase change materials. Nanomaterials 13 (11), 1755.
  36. Lugnan A., Aggarwal S., Brückerhoff-Plückelmann F., Wright C. D., Pernice W. H. P., Bhaskaran H. (2025). Emergent self-adaptation in an integrated photonic neural network for backpropagation-free learning. Adv. Sci. 12 (2), 2404920. doi:10.1002/advs.202404920
  37. Morcillo J.-F., Sanchis P., Parra J. (2024). On-chip electro-optical spiking VO 2 /Si device with an inhibitory leaky integrate-and-fire response. Opt. Mater. Express 14 (11), 2681 – 2693. doi:10.1364/ome.537717
  38. Navarro H., del Valle J., Kalcheim Y., Vargas N. M., Adda C., Lee M.-H. (2021). A hybrid optoelectronic mott insulator. Appl. Phys. Lett. 118 (14), 141901. doi:10.1063/5.0044066
  39. Němec P., Moreac A., Nazabal V., Pavlišta M., Přikryl J., Frumar M. (2009). Ge–sb–Te thin films deposited by pulsed laser: an ellipsometry and raman scattering spectroscopy study. J. Appl. Phys. 106 (10), 103509. doi:10.1063/1.3259435
  40. Němec P., Přikryl J., Nazabal V., Frumar M. (2011). Optical characteristics of pulsed laser deposited ge–sb–te thin films studied by spectroscopic ellipsometry. J. Appl. Phys. 109 (7), 073520. doi:10.1063/1.3569865
  41. Noé P., Vallée C., Hippert F., Fillot F., Raty J.-Y. (2018a). Phase-change materials for non-volatile memory devices: from technological challenges to materials science issues. Semicond. Sci. Technol. 33 (1), 013002. doi:10.1088/1361-6641/aa7c25
  42. Noé P., Hippert F. (2018b). “ Structure and properties of chalcogenide materials for PCM,” in Phase Change Memory: Device Physics, Reliability and Applications. Editor Redaelli A. (Cham: Springer International Publishing), 125 – 179. doi:10.1007/978-3-319-69053-7_6
  43. Orava J., Wágner T., Šik J., Přikryl J., Frumar M., Beneš L. (2008). Optical properties and phase change transition in Ge2Sb2Te5 flash evaporated thin films studied by temperature dependent spectroscopic ellipsometry. J. Appl. Phys. 104 (4), 043523. doi:10.1063/1.2970069
  44. O’Hara A., Nunley T. N., Posadas A. B., Zollner S., Demkov A. A. (2014). Electronic and optical properties of NbO2. J. Appl. Phys. 116 (21), 213705. doi:10.1063/1.4903067
  45. Pang J., Ren H., Chen R., Ren Y., Wu Y., Yi Q. (2025). Nonvolatile continuously tunable integrated optical router. Laser and Photonics Rev. 19 (23), 2500281. doi:10.1002/lpor.202500281
  46. Parra J., Morcillo J.-F., Schouteden K., Seo J. W., Locquet J.-P., Sanchis P. (2026). Broadband and ultra-compact all-optical nonlinear activation function enabled by VO2-integrated photonics. J. Phys. Photonics 8 (2), 025010. doi:10.1088/2515-7647/ae6004
  47. Pavanello F., Marchand C., O’Connor I., Orobtchouk R., Mandorlo F., Letartre X. (2025). “ NEUROPULS: NEUROmorphic energy-efficient secure accelerators based on phase change materials aUgmented siLicon photonicS,” in 2023 IEEE European Test Symposium (ETS), May 2023
  48. Peserico N., Shastri B. J., Sorger V. J. (2023). Integrated photonic tensor processing unit for a matrix multiply accelerator. J. Light. Technol. 41 (12), 3704 – 3716. doi:10.1109/jlt.2023.3269957
  49. Prabhathan P., Sreekanth K. V., Teng J., Ko J. H., Yoo Y. J., Heong H.-H. (2023). Roadmap for phase change materials in photonics and beyond. iScience 26 (11), 108149.
  50. Raty J.-Y., Wuttig M. (2020). The interplay between peierls distortions and metavalent bonding in IV–VI compounds: comparing GeTe with related monochalcogenides. J. Phys. D Appl. Phys. 53 (23), 234002. doi:10.1088/2F1361-6463/2Fab7e66
  51. Raty J.-Y., Schumacher M., Golub P., Deringer V. L., Gatti C., Wuttig M. (2019). A quantum-mechanical map for bonding and properties in solids. Adv. Mater. 31 (3), 1806280. doi:10.1002/adma.201806280
  52. Raty J.-Y., Gatti C., Schön C.-F., Wuttig M. (2021). How to Identify Lone Pairs, Van der Waals Gaps, and Metavalent Bonding Using Charge and Pair Density Methods: from Elemental Chalcogens to Lead Chalcogenides and Phase-Change Materials. Physica Status Solidi (RRL) – Rapid Res. Lett. 15 (11), 2000534. doi:10.1002/pssr.202000534
  53. Ren K., Zhu M., Song W., Lv S., Xia M., Wang Y. (2019). Electrical switching properties and structural characteristics of GeSe–GeTe films. Nanoscale 11 (4), 1595 – 1603. doi:10.1039/c8nr07832g
  54. Ríos C., Stegmaier M., Hosseini P., Wang D., Scherer T., Wright C. D. (2015). Integrated all-photonic non-volatile multi-level memory. Nat. Photonics 9, 725 – 732. doi:10.1038/nphoton.2015.182
  55. Ríos C., Youngblood N., Cheng Z., Le Gallo M., Pernice W. H. P., Wright C. D. (2019). In-memory computing on a photonic platform. Sci. Adv. 5 (2), eaau5759. doi:10.1126/sciadv.aau5759
  56. Ríos C., Du Q., Zhang Y., Popescu C. C., Shalaginov M. Y., Miller P. (2022). Ultra-compact nonvolatile phase shifter based on electrically reprogrammable transparent phase change materials. PhotoniX 3, 26. doi:10.1186/s43074-022-00070-4
  57. Sawant R., Albanese A., Rogemont A., Gonzalez-Cortes G., Brûlé Y., Karam L. (2025). High-endurance and energy-efficient all-optical programming of scalable silicon waveguides with integrated phase-change material patches. Adv. Opt. Mater. 13, 1 – 8. doi:10.1002/adom.202500775
  58. Shen Y., Harris N. C., Skirlo S., Prabhu M., Baehr-Jones T., Hochberg M. (2017). Deep learning with coherent nanophotonic circuits. Nat. Photonics 11, 441 – 446.
  59. Soref R., McDaniel D., Bennett B. (1988). Guided-wave intensity modulators using amplitude-and-phase perturbations. J. Light. Technol. 6 (3), 437 – 444. doi:10.1109/50.4021
  60. Soref R., Hendrickson J., Liang H., Majumdar A., Mu J., Li X. (2015). Electro-optical switching at 1550 nm using a two-state GeSe phase-change layer. Opt. Express 23 (2), 1536 – 1546. doi:10.1364/OE.23.001536
  61. Sun C., Wade M. T., Lee Y., Orcutt J. S., Alloatti L., Georgas M. S. (2015). Single-chip microprocessor that communicates directly using light. Nature 528 (7583), 534 – 538. doi:10.1038/nature16454
  62. Sun H., Lian C., Vásquez-Aza F., Rahimi Kari S., Huang Y. S., Restelli A. (2025). Microheater hotspot engineering for spatially resolved and repeatable multi-level switching in foundry-processed phase change silicon photonics. Nat. Commun. 16, 4291. doi:10.1038/s41467-025-59399-6
  63. Teo T. Y., Krbal M., Mistrik J., Prikryl J., Lu L., Simpson R. E. (2022). Comparison and analysis of phase change materials-based reconfigurable silicon photonic directional couplers. Opt. Mater. Express 12 (2), 606 – 621. doi:10.1364/ome.447289
  64. Tomelleri M., Hippert F., Farjot T., Licitra C., Vaxelaire N., Dory J.-B. (2021). Overcoming the thermal stability limit of chalcogenide phase-change materials for high-temperature applications in GeSe1-xTex thin films. Physica Status Solidi (RRL) – Rapid Res. Lett. 15 (3), 2000451. doi:10.1002/pssr.202000451
  65. Wei M., Saxena A., (2024a). Monolithic back-end-of-line integration of phase change photonics. Nat. Commun. 15 (1), 3004.
  66. Wei M., Xu K., Tang B., Li J., Yun Y., Zhang P. (2024). Monolithic back-end-of-line integration of phase change materials into foundry-manufactured silicon photonics. Nat. Commun. 15 (1), 2786. doi:10.1038/s41467-024-47206-7
  67. Woollam J. A., Johs B. D., Herzinger C. M., Hilfiker J. N., Synowicki R. A., Bungay C. L. (1999). Overview of variable-angle spectroscopic ellipsometry (VASE): I. Basic theory and typical applications. Opt. Metrol. A Crit. Rev. 10294, 3. doi:10.1117/12.351660
  68. Wu K., Zhang M., Ma S., Chen L. (2026). Electrically tunable terahertz metadevices: materials, functions, and applications. Apl. Eng. Phys. 1 (1), 011501. doi:10.1063/5.0312371
  69. Wuttig M., Bhaskaran H., Taubner T. (2017). Phase-change materials for non-volatile photonic applications. Nat. Photonics 11 (8), 465 – 476. doi:10.1038/nphoton.2017.126
  70. Wuttig M., Deringer V. L., Gonze X., Bichara C., Raty J.-Y. (2018). Incipient metals: functional materials with a unique bonding mechanism. Adv. Mater. 30 (51), 1803777. doi:10.1002/adma.201803777
  71. Wuttig M., Schön C.-F., Lötfering J., Golub P., Gatti C., Raty J.-Y. (2023). Revisiting the nature of chemical bonding in chalcogenides to explain and design their properties. Adv. Mater. 35 (20), e2208485. doi:10.1002/adma.202208485
  72. Yang Z., Ko C., Ramanathan S. (2011). Oxide electronics utilizing ultrafast metal-insulator transitions. Annu. Rev. Mater. Res. 41, 337 – 367. doi:10.1146/annurev-matsci-062910-100347
  73. Yang X., Lu L., Li Y., Wu Y., Li Z., Chen J. (2023). Non-volatile optical switch element enabled by low-loss phase change material. Adv. Funct. Mater. 33, 2304601. doi:10.1002/adfm.202304601
  74. Youngblood N., Ríos Ocampo C. A., Pernice W. H. P., Bhaskaran H. (2023). Integrated optical memristors. Nat. Photonics 17 (7), 561 – 572. doi:10.1038/s41566-023-01217-w
  75. Zhang Q., Zhang Y., Li J., Soref R., Gu T., Hu J. (2018). Broadband nonvolatile photonic switching based on optical phase change materials: beyond the classical figure-of-merit. Opt. Letters.
  76. Zhang Y., Chou J. B., Li J., Li H., Du Q., Yadav A. (2019). Broadband transparent optical phase change materials for high-performance nonvolatile photonics. Nat. Commun. 10 (1), 1 – 9. doi:10.1038/s41467-019-12196-4
  77. Zhou H., Dong J., Cheng J., Dong W., Huang C., Shen Y. (2022). Photonic matrix multiplication lights up photonic accelerator and beyond. Light Sci. and Appl. 11 (1), 30. doi:10.1038/s41377-022-00717-8
  78. Zhu M., Cojocaru-Mirédin O., Mio A. M., Keutgen J., Küpers M., Yu Y. (2018). Unique bond breaking in crystalline phase change materials and the quest for metavalent bonding. Adv. Mater. 30 (18), 1706735. doi:10.1002/adma.201706735

Citation

Dwivedi, R., et al. (2026). Photonic devices based on phase-change materials for neuromorphic computing. Frontiers in Photonics, Volume 7. https://doi.org/
Received
Accepted
Published
Volume
7 - 2026

Copyright

Copyright © 2026 Dwivedi, Pourny, Albanese, Yadav, Orobtchouk, Benane, Sawant, Tomelleri, Brulé, Rogemont, Coillet, Charbonnier, Noé, Cluzel and Pavanello.

License: https://creativecommons.org/licenses/by/4.0/

*Correspondence: Fabio Pavanello

Article metrics

Views
0
Downloads
0
Citations
0
Shares
0

Related articles

Original ResearchInterleaving SLIDE microscopyChristian Stock, Gregory Berg, Matthea Thielking, Jonas Jurkevičius, Sebastian KarpfOriginal ResearchClosed-form Fresnel analysis of double-slit diffraction with multi-wave coherent illuminationJ. Sumaya-Martínez, M. A. Ortiz-Ferreyro, O. Rojas-HernandezEditorialEditorial: Advances in optical methods for biological sample characterizationEden Morales-Narváez, Remy Avila, Pablo Loza-Alvarez