Dr. Ali Mahdavi

Set up a simulation and design department at Sanan Optoelectronics, starting from zero, for advanced emitters, displays and integrated optical interconnects.

Established optical and electrical design workflows for LEDs, lasers and optical interconnects, built the technical interfaces between design, epitaxy, process and characterization, and aligned these teams around emitter and display roadmaps.

Technical leadership for micro-LED device and optical design in AR display programmes, covering device architecture, optics, process integration and wafer-level test, for red AlInGaP and green and blue InGaN emitters.

Contributed to a world-record AR micro-LED display through device architecture, inverse design and optimization. Designed wafer-level characterization platforms and test vehicles that sped up decisions and closed the loop between simulation and measurement, and coordinated suppliers and cross-functional teams across the US, Europe and Asia.

Light extraction from a micro-LED is the average over thousands of mutually incoherent dipole emitters inside a high-index semiconductor. Simulating each of them with FDTD makes gradient-based design impractical. This work uses the Fourier modal method instead, with a vector formulation for fast convergence in structures that contain metal and Brillouin-zone integration to represent isolated dipoles in a periodic cell.

The method matches FDTD while running about 3 × 10³ times faster per dipole, a projected 2 × 10⁷ for the full source set. Because the solver (FMMAX, written in JAX) is differentiable, a metasurface on the mesa can be optimized directly: extraction at 630 nm rises from 28.4 % for the unoptimized device to 40.1 % with optimized layer thicknesses and 56.2 % with the metasurface.

Inverse design of a thin-film coating is ill-posed: very different stacks can produce almost the same reflectance spectrum, so a network trained to return a single answer averages between them. A conditional invertible neural network instead learns the full distribution of layer thicknesses given a target spectrum. Sampling its Gaussian latent space returns several distinct candidate designs at once.

The network was trained on 10⁶ nine-layer Nb₂O₅/SiO₂ stacks with layer thicknesses of 10–200 nm and normal-incidence reflectance from 400 to 700 nm, generated with TMM-Fast. Its proposals serve as starting points for local optimization (downhill simplex or Levenberg–Marquardt). Across 2000 test targets this beats random initialization, and it also returns useful designs for hand-drawn target spectra with no known exact solution.

The transfer-matrix method gives exact reflection and transmission of planar multilayers, but common implementations loop over wavelengths, angles and stacks one at a time. TMM-Fast evaluates the characteristic matrices of all layers as a single tensor operation across these dimensions, which makes large training sets for machine learning and population-based optimization inexpensive.

Against the widely used reference implementation it is about 100 times faster on one core, with a further gain from parallel execution across cores: a million nine-layer stacks at 100 wavelengths and 10 angles take roughly 40 minutes on an 8-core machine. A PyTorch version supplies automatic gradients with respect to layer thicknesses and refractive indices, and an OpenAI Gym environment poses layer-by-layer design as a reinforcement-learning task.

In a dense micro-LED array, light that travels sideways in the layers above the active zone leaks into neighbouring pixels and lowers contrast. Here each emitter's semiconductor body, two semiconductor layers with the active zone between them, is contacted through a cover region (a current-spreading layer, a transparent conductive oxide, or both) that is laterally wider than the active zone.

The part of the cover region outside the footprint of the active zone carries vertical recesses, arranged as closed or interrupted frames or rings around the emitter. The recesses and the cover material between them form a lateral Bragg mirror, n₁d₁ = n₂d₂ = (2k − 1)λ/4, that reflects sideways-travelling light back and also narrows the emission, while the contact above the active zone is unchanged. Arrays sharing one continuous cover region, for example for AR and VR displays, are also claimed.

Nanofins are narrow semiconductor walls grown up from a base plane by selective-area growth, with the active quantum wells on the fin. When the InGaN fins of a red micro-LED all run parallel, its far field depends on the azimuthal angle.

Here each light-emitting fin runs along at least two directions in plan view, as concentric hexagons, triangles or circles, with the fin directions 60° apart and following the GaN crystal planes. The emission becomes independent of azimuth. Passive fins of similar height, grown without the active zone, surround or fill the pattern so that growth, and with it the emission, stays uniform across the pixel.

Designing a multilayer coating mixes discrete choices, which material comes next and when to stop, with continuous ones, how thick each layer is. Here the stack is built layer by layer by a reinforcement-learning agent in a parameterized action space: every discrete material choice carries its own continuous thickness, and a terminate action lets the agent decide the number of layers. A multi-pass deep Q-network learns from transfer-matrix simulations, with an exponentially shaped reward and an optional penalty on total thickness.

On target reflectance spectra in the visible, the agent reached rewards at least 20 % higher than a discretized baseline with a fifth of the simulation calls. For a reflector at 445–455 nm over 0–60° it found a 19-layer, 1.31 µm design that outperforms a 34-layer expert reference.

In a planar LED most of the light generated in the active layer meets the chip surface beyond the critical angle and stays trapped. In a micro-LED, in addition, a large share of the active area lies close to etched mesa edges, where carriers recombine without emitting light.

Here the first doped region is structured into a stepped, tapering body (a trapezoid, a step pyramid or nearly a hemisphere) by repeated etching or overgrowth. The active region covers its side faces as well as its top, and the second doped region follows. Light is then generated almost perpendicular to the curved outer surface, which reduces total internal reflection, and less of the active region ends at an etched edge. The concept applies to InGaAlP, InGaAlAs and InGaAlN emitters.

In a micro-LED array the transparent current-spreading layer on top is shared by all pixels, so it also guides light sideways into the neighbouring pixels. Here a two-dimensional photonic crystal, a hexagonal lattice of holes, is etched into that common layer, between the separate active regions and the light-emitting surface.

The lateral photonic band gap blocks in-plane propagation and suppresses crosstalk, while the lattice couples trapped and guided modes into the escape cone, which raises extraction and narrows the beam. The layer still spreads current: a 300 nm ITO layer with holes has a lower resistance than a plain 150 nm layer for r/a below 0.33. In the granted claims each active region sits in a funnel-shaped mesa.

The first work at ams-OSRAM was VCSEL design. A transfer-matrix program, written from scratch, computes for a complete 850 nm epitaxial stack the reflection spectrum of the Bragg mirrors, the lasing wavelength from the round-trip phase condition, and the standing-wave field inside the resonator. Graded aluminium profiles are resolved into sub-layers, and a refractive-index library for Al_xGa_1−xAs, built on the Adachi and Afromowitz models, was checked against ellipsometry.

The code set thickness tolerances for the passivation and contact layers, showed the effect of the oxide aperture on the mirror, and turned measured reflectivity into a diagnosis of the growth: the shape of the spectrum and the position of the resonance dip in the stop band reveal deviations from the target layers.

Alongside the optics, carrier transport was modelled in APSYS (Crosslight) together with a thermal model, and ray-tracing models of 850 nm and 940 nm infrared chips were built in Zemax and LightTools. With precisely measured inputs, the LightTools models predicted the out-coupled light without fitted parameters.

Residual fabrication disorder in a photonic-crystal W1 waveguide turns the slow-light region near the band edge into a chain of randomly formed cavities: Anderson localization. Instead of coupling through an access waveguide, these modes were excited directly from free space with a tightly focused beam of tailored polarization, linear or radial, and the out-of-plane scattering was imaged along the guide.

Integrating the scattering spectra along the waveguide at the modal cut-off reveals the individual localized cavities and their coupling, with quality factors up to 2 × 10⁵. The dependence on input polarization and position measures how well the focal field overlaps each mode, a route to beam-steered addressing of cavities and to scanning-microscopy sensing.

Silicon is an indirect-gap semiconductor. Its band-edge photoluminescence is weak, and most of it is trapped inside the high-index material. In a photonic-crystal slab, emission can couple to flat, low-group-velocity bands at the Γ point that lie above the light line and radiate vertically. Counting these modes inside the emission band, from plane-wave and FDTD band-structure calculations over lattice constant and hole radius, gives a map of where extraction should peak.

Thirty-six triangular air-hole lattices were written by electron-beam lithography in 220 nm silicon-on-insulator around the predicted optimum and measured by micro-photoluminescence. The best slab (a = 480 nm, r/a = 0.29) emits more than 500 times more light than unpatterned silicon at the 975 nm peak, and over 200 times integrated over the spectrum. Randomly placed holes with the same fill factor reach at most 40 times, which separates the band-structure effect from plain scattering.

In an ambipolar light-emitting field-effect transistor, electrons and holes enter from opposite electrodes and recombine in a narrow zone whose position along the channel follows the gate voltage. Gold nanorods, acting as plasmonic nanoantennas, were placed in one half of the channel of a top-gated transistor with the near-infrared polymer DPPT-BT as the emitter.

Sweeping the gate at constant drain current moves the recombination zone onto or away from the rods and switches the near-field coupling between excitons and plasmons on and off. After correcting for the mode area, electroluminescence is enhanced locally by up to 12 times at 1100 nm for short rods and 47 times at 1350 nm for long rods. Three-dimensional FDTD simulations of the radiative-rate enhancement match the measured spectra, and charge transport is almost unaffected by the rods.

Silicon photonic-crystal slabs with a W1 line defect support guided modes, deliberately designed defect cavities, and random high-Q cavities created by disorder. Reaching them from free space usually requires a tapered fibre brought into the near field of the slab.

This work instead shapes the polarization distribution inside the diffraction-limited spot of a normally incident, tightly focused beam to match the mode being addressed. The approach is more flexible and more robust when the sample is scanned, and it led to the 2017 study of Anderson-localized modes.

Plasmonic resonances of periodic, infinitely long chains of metallic nanopins, and the red shift of the collective resonance that near-field coupling between neighbouring pins produces relative to an isolated pin.

Metallic nanoparticles with a 10 × 10 nm cross-section, resonant at hundreds of terahertz, were simulated full-wave and their electric-field distributions compared. A pair of coupled rods reproduces the resonant behaviour of split rings and split squares, so a geometrically complex particle can be understood, and replaced, by simpler elements.

The fall in resonance frequency as the gap narrows, usually attributed to the self-capacitance of the gap, is reinterpreted as surface plasmons coupled across the gap.

A liquid film less than a micrometre thick is suspended in a frame between two electrodes and placed between two capacitor plates. When a current flows through the film and an external field is applied perpendicular to it, the film starts to rotate steadily in its own plane. Its sense of rotation is fixed by E_ext × E_in. This is the liquid film motor, first reported by the same group in 2008.

This study used films of the nematic liquid crystal MBBA. As the external field is raised, the electrohydrodynamic vortices driven by the current turn into a single rotation, and the current required to drive the vortices drops sharply. Rotation sets in once the product of external field and film voltage passes a threshold, which was mapped against both.

The mechanism was explained later (Phys. Rev. E 92, 033002, 2015). The electrodes and the capacitively coupled plates induce free charges over the whole surface of the film, not only at its edges, and the in-plane field acting on this charge distribution exerts the torque.