Light-Based Computing Takes Step Towards Efficiency with Stable Lithium Niobate Tuning

Researchers are actively pursuing nonvolatile optical methods for performing energy-efficient matrix-vector multiplications, crucial operations in modern computing. Abhiram Devata from Northwestern University, Axel Magaña Ponce from Elmhurst University, and David Barton from Northwestern University et al. demonstrate a novel approach utilising electrochemical lithiation of thin film lithium niobate to achieve programmable and nonvolatile control of refractive index. This work is significant because it overcomes a key limitation of lithium niobate photonics, the lack of nonvolatility, and paves the way for scalable, low-energy photonic computing systems capable of performing complex image processing and iterative weight updates with high accuracy, validated at an average relative error of only 1.5%.

Electrochemical lithiation tuning refractive index for nonvolatile photonic matrix operations

Scientists have developed a pathway to nonvolatile photonic computing using thin film lithium niobate, addressing a critical need for energy-efficient matrix operations at scale. Matrix-vector multiplications, essential for artificial intelligence and high-throughput computations, currently consume significant energy in electronic processors.

This work introduces electrochemical lithiation as a method to non-volatility tune the refractive index of thin film lithium niobate, enabling programmable and low-loss matrix-vector multiplications. Researchers computationally demonstrate the feasibility of controlling matrix weights by altering the composition of lithium niobate, a material already favoured for its strong electro-optic effect and low optical loss.

The study establishes that the lithium niobate phase remains stable at room temperature across a 2% lithium composition window, exhibiting a corresponding change in refractive index. This composition-dependent refractive index is then harnessed to design Mach-Zehnder interferometers for image processing and microring resonators for iterative weight updates.

Simulations incorporate realistic material loss constraints, validating the accuracy of matrix-vector multiplication with an average relative error of 1.5%. These components are designed to be reprogrammable, mimicking the learning process of neural networks through gradient descent training. This innovative approach leverages the material as an electrolyte, allowing lithium ions to be added and removed, effectively tuning the refractive index without continuous power consumption.

The achieved index change of 0.008 is substantial enough to enable the design of compact photonic circuits for computing applications. By cascading and reprogramming these interferometers and ring resonators, the research demonstrates the potential for performing complex matrix operations, such as those required for image processing and neural network training. The findings represent a facile route towards nonvolatile photonic computing in thin film lithium niobate, offering a promising solution for reducing the energy footprint of future AI hardware.

Fabrication and Characterisation of Lithium Niobate Waveguides for Electrochemical Matrix Multiplication

Electrochemical lithiation serves as a route to nonvolatile matrix-vector multiplications within thin film lithium niobate (TFLN). The LiNbO3 phase remains stable at room temperature across a 2% lithium composition window, exhibiting a composition-dependent refractive index. Computational demonstrations confirm this as a programmable and low-loss approach to matrix-vector multiplication, utilising composition to modulate matrix weights.

Mach-Zehnder interferometers and microring resonators were designed to perform image processing tasks and iterative weight updates respectively, accounting for realistic material loss. Monolithically etched rib waveguides were fabricated on 600nm TFLN, incorporating a metal back contact beneath a silicon dioxide buried oxide and a polymer electrolyte cladding with a lithium source above the TFLN.

This configuration prevents absorption-induced propagation loss of guided modes. Waveguides were designed with 60-degree sidewall angles and a refractive index of 1.486 was used for the electrolyte to model mode confinement. The impact of organic polymer absorption in the near infrared regime was considered, with potential redesign for the O-band at 1310nm if necessary.

Applying a positive bias drives lithium ions into the TFLN, altering both the ordinary and extraordinary refractive indices. Simulations were restricted to the 2% composition window starting from congruent lithium niobate to avoid phase transitions. Experimentally measured index changes at 1550nm demonstrate a nearly linear relationship with added lithium across the composition window.

The anisotropy of lithium addition is highlighted, with the change in extraordinary index (-0.016) being eightfold greater than the ordinary index change (+0.002). This contrasts with typical TFLN electro-optic modulators which demonstrate index changes of up to 0.003, or around 0.0001 in most cases. Mach-Zehnder interferometers (MZIs) exploit the refractive index change for interferometry.

The critical design parameter, the π-phase shift length Lπ, depends on the difference in mode index between the MZI arms, calculated as L = λ/2Δn, where λ is the free space wavelength (1550nm) and Δn is the difference between lithiated and unlithiated effective TE00 mode indices. Δn ranges from 8.6×10-4 to 5.5×10-3 depending on waveguide dimensions. MZIs were designed in X-cut TFLN using TE modes to maximise index modulation and minimise device footprint. Nonvolatile optical trimming was also considered to correct for potential fabrication errors.

Lithium concentration modulates refractive indices enabling enhanced electro-optic performance

A refractive index change of 0.008 was observed, facilitating the design and simulation of interferometers and ring resonators for computing applications. The research demonstrates a composition-dependent waveguide mode index change sufficient for performing matrix-vector multiplications. Waveguides were designed with a 60-degree sidewall angle on 600nm thin film lithium niobate, utilising a refractive index of 1.486 for the electrolyte to model mode confinement.

Simulated TE00 mode profiles for unlithiated material at a wavelength of 1550nm were included as insets for reference. Lithiation induces a nearly linear relation between lithium concentration and refractive index across a 2% composition window, starting from congruent lithium niobate. The absolute change in the extraordinary index reached -0.016 over the composition window, which is eight times greater than the ordinary index change of +0.002.

These index changes surpass those demonstrated in typical TFLN electro-optic modulators, which generally produce index changes around 0.0001, and are comparable to the up to 0.003 index changes seen in some TFLN electro-optic modulators. Mach-Zehnder Interferometers were designed with a π-phase shift length, Lπ, dependent on the difference in mode index between the two arms.

Depending on waveguide width and etch depth, the effective mode index difference, Δneff, ranged from 8.6×10-4 to 5.5×10-3. A waveguide width of 800nm and etch depth of 315nm were selected, yielding a Δneff of 0.0078 and an Lπ of 100μm. This Lπ length is comparable to thermo-optic and carrier accumulation MZIs on silicon on insulator, and is three to fifteen times shorter than carrier depletion MZIs.

Furthermore, optical trimming is possible, with up to 40μm of additional length required to correct for up to 10% decreases in waveguide geometry parameters, ensuring a full π-phase shift. Larger rib geometries do not require additional length, as they would experience a Δneff greater than the target from lithiation. Matrix-vector multiplication accuracy was validated at an average relative error of 1.5%, demonstrating the potential for nonvolatile photonic computing in TFLN.

Electrochemically programmed lithium niobate for passive and scalable photonic matrix operations

Researchers have demonstrated a nonvolatile approach to matrix-vector multiplication using electrochemical lithiation in thin film lithium niobate. This technique exploits the composition-dependent refractive index of lithium niobate to program and store matrix weights, enabling energy-efficient photonic computing.

Computational modelling and simulations confirm the feasibility of this method for image processing and neural network training tasks. The designs incorporate both Mach-Zehnder interferometers and microring resonators to perform matrix operations, achieving an average relative error of 1.5% in validating matrix-vector multiplication accuracy.

These compact and scalable structures allow for the monolithic integration of electro-optic and nonvolatile index control within a single platform, a capability previously unavailable. The system operates passively once trained, representing the target matrix without continuous energy input. Limitations acknowledged include material loss of 1.5 dB cm-1 and 10% noise, which were accounted for during the training process.

Future work may focus on optimising these parameters and exploring more complex neural network architectures to further enhance performance and scalability. This advancement addresses a critical need for energy-efficient photonic matrix operations, potentially enabling significant improvements in high-throughput computing applications.