Strained KNbO3 thin-film electro-optic modulator for datacenter co-packaged optics

Strained orthorhombic KNbO3 thin-film electro-optic modulators represent a credible materials-science path to driving voltages that thin-film lithium niobate cannot reach. The key figure of merit is the half-wave voltage-length product (Vpi*L): the claimed device stack achieves 0.36-0.59 V·cm at a bandwidth-qualified 70 GHz or higher operating point, compared to the 2.2 V·cm TFLN baseline reported by Wang et al. in Nature (2018). That is a 3.7-6.1x reduction in drive voltage at the bandwidth required for high-speed datacenter links, which directly shrinks the modulator footprint and cuts SerDes power in co-packaged optical transceivers. The inventive step is not the KNbO3 composition — that is expired background art — but the specific strain-engineered thin-film stack: c-axis-tilted orthorhombic KNbO3 on an MgO buffer with a SiN ridge waveguide. Epitaxial strain and c-axis tilt together stabilize the polar Amm2 phase and rotate the electro-optic tensor into a favorable orientation. Without that stack architecture, bulk KNbO3 is dynamically unstable (detailed below), so the claimed device is what converts a notoriously soft ferroelectric into a manufacturable modulator. The timing is practical rather than urgent: a race condition that had been flagged in Lattice Graph's internal tracking was subsequently withdrawn, so the asset now proceeds at normal prosecution priority. A buyer has runway to fund the two remaining validation gates — a filing-grade Pockels-tensor DFPT calculation and a physical modulator coupon — before committing to full commercialization, while still entering a photonic-integration market where TFLN adoption is not yet locked in at the module-maker level.

The engines did not fully agree here — the asset carries that uncertainty openly rather than overstating confidence.

The material is potassium niobate (KNbO3) in its orthorhombic ferroelectric phase (space group Amm2), a perovskite niobate known for a large Pockels coefficient that, in bulk, is suppressed by a soft transverse optical phonon mode that drives the structure toward dynamic instability at room temperature. Computationally, two independent machine-learning interatomic potentials — MACE and CHGNet — disagree sharply on the phonon spectrum of the bulk phase: MACE predicts the relevant mode at +0.761 THz (stable), while CHGNet predicts it at -1.149 THz (imaginary, indicating instability). This inter-potential disagreement is the quantitative signal that the bulk phase sits near a lattice instability and cannot be relied upon without strain engineering. The bandgap of 2.293 eV positions the material in the near-UV absorption edge, well above telecom wavelengths, which is a necessary condition for low-loss electro-optic operation. The central materials-science argument is that epitaxial strain on an MgO buffer layer, combined with a c-axis tilt of the film, moves the ferroelectric soft mode to positive frequency, stabilizing the polar Amm2 phase. A 15-configuration strain-tilt phonon sweep was run to quantify this: at 1.0% biaxial strain and a 25-degree c-axis tilt, the previously imaginary soft mode shifts to +0.289 THz — confirming dynamic stability in the thin-film geometry. This is not multi-engine consensus on the bulk phase (the bulk engines disagree); instead, it is a targeted simulation demonstrating that the specific strained-and-tilted configuration is stable. The claim is appropriately limited to that configuration, not to bulk KNbO3. Waveguide design was optimized using an effective-index-method Pareto sweep over 3,750 configurations based on the Marcatili approximation, mapping the trade-off between confinement, overlap integral, and Vpi*L across SiN ridge geometries. The best configurations in the 70 GHz-qualified regime yield Vpi*L of 0.36-0.59 V·cm. A 0.173 V·cm Pareto point also appears in the sweep but is qualified only at 40 GHz and is not relied upon for the commercial claim. The underlying database draws on four DFT source calculations, and a recipe base of 75 exact KNbO3, LiNbO3, and LiTaO3 depositions via pulsed laser deposition and sputtering provides process-level credibility for the MgO-buffered stack.

The total addressable market is estimated at $5 billion or more, driven by hyperscaler demand for high-bandwidth, low-power optical links in AI datacenter fabrics. Co-packaged optics — where the optical transceiver is mounted directly on or near the switch ASIC package — is the primary use case and is creating a pull for modulators that fit within tight power and footprint budgets. Every reduction in Vpi*L reduces the drive voltage the SerDes must deliver, which at 70+ GHz operating bandwidth is a hard constraint: silicon photonics Mach-Zehnder modulators are limited in bandwidth, and TFLN is currently the leading high-performance alternative, but its 2.2 V·cm Vpi*L still demands relatively high RF swing from the driver electronics. The customers are co-packaged-optics module makers who integrate EO modulators into transceiver assemblies, and hyperscaler photonics teams that design captive optical interconnect solutions. Both groups price component performance against power consumption at the rack or switch level. A royalty or field-of-use license priced against the performance premium — the 3.7-6.1x reduction in drive voltage versus TFLN — is supportable because it directly reduces system power in high-margin, power-constrained AI infrastructure. Relevant use cases named include Mach-Zehnder modulators and WDM co-packaged-optics modules. Market figures here are estimates, and no commercial commitments exist at this stage. The asset is part of Lattice Graph's broader dielectric, ferroelectric, and wide-bandgap oxides portfolio, which spans related electro-optic stack compositions including x-cut TFLN, z-cut lithium tantalate, PMN-PT at the morphotropic phase boundary, KTaO3, KTa(1-x)Nb(x)O3, and ScBiO3. This portfolio breadth means a strategic licensee can optionally acquire coverage across multiple EO-stack architectures, not just the KNbO3 lead.

~3.7-6.1x lower Vpi*L than TFLN at >=70 GHz bandwidth-qualified operating point

strained thin-film device stack + MgO buffer + process window; KNbO3 composition not claimed

The natural acquirers and licensees are co-packaged-optics module makers and hyperscaler photonics groups, both of which are under pressure to reduce modulator drive voltage at 70+ GHz bandwidth. Module makers — integrating Mach-Zehnder and WDM components into transceiver assemblies at scale — are well-positioned for a field-of-use license covering those specific applications; the performance premium supports a per-device royalty against the cost reduction in driver electronics. Hyperscaler photonics teams building captive optical interconnects have a stronger incentive to pursue exclusivity: a 3.7-6.1x lower Vpi*L versus TFLN is a differentiated interconnect specification that a hyperscaler can protect. For that buyer, an exclusive license covering datacenter optical interconnect applications could be structured around the two remaining proof gates, with milestone payments tied to DFPT and coupon validation. Strategic photonics companies — particularly those already licensing or developing TFLN platforms — are also plausible buyers who would use the KNbO3 lead to either hedge against TFLN or position a higher-performance product tier. Given that the broader dielectric, ferroelectric, and wide-bandgap oxides portfolio includes TFLN, lithium tantalate, and PMN-PT EO stacks alongside the KNbO3 lead, a photonics strategic acquiring multiple positions from the portfolio could cover both the mature TFLN baseline and the higher-performance KNbO3 architecture in a single licensing arrangement.

The central risk is that the entire value proposition depends on the strained thin-film stack remaining stable and manufacturable — and dynamic stability has been demonstrated computationally in a 15-configuration phonon sweep, not in a fabricated film. The bulk phase is dynamically unstable under two independent ML potentials, so if the epitaxial strain window in real deposition conditions is narrower than the sweep predicts, or if the MgO buffer introduces defects that relax the strain, the Pockels response may not match the modeled Vpi*L. The headline 0.36-0.59 V·cm is a modeled, bandwidth-qualified figure; it has not yet been measured on a physical coupon. A buyer must treat this as a projected specification until the coupon gate is closed. A secondary risk is the §103 obviousness surface over BaTiO3-on-Si art. While the specific strain-tilt-MgO architecture provides differentiation, an examiner could argue that strain engineering of perovskite ferroelectrics on oxide buffers was a known technique, making the KNbO3 variant potentially obvious. The de-risking path is straightforward: complete the Pockels-tensor DFPT to establish the computed electro-optic coefficients quantitatively, fabricate the modulator coupon to generate measured Vpi*L data, and use both to build a specification that ties the non-obviousness argument to specific, measured performance at the claimed strain-tilt operating point. With the race condition withdrawn, a buyer has the time to execute both steps before filing rather than under deadline pressure.