Circuits that think in light.

I am Bartu Yaman — physicist, MSc, PhD candidate and research associate at the Fraunhofer Heinrich Hertz Institute in Berlin. I build integrated photonic circuits: heterogeneous integration to allow increased functionality on photonic circuit platforms, high-frequency electro-optics, and optical neural networks — with an emphasis on modelling, design and simulations.


The EO workhorse: Mach–Zehnder modulatorFig. I

Conscious parameter-space exploration & surrogatesFig. II

On-chip diffractive neural networksFig. III

How I work01

What I am working on02

Where I have been03


Fig. I — The EO workhorse: Mach–Zehnder modulator

This is the device that puts electronics onto light. Every high-speed link — datacom inside the datacentre, telecom between cities — needs an electrical signal turned into an optical one, and the Mach–Zehnder modulator is the workhorse that does it: broadband, linear enough to shape, and manufacturable. Its bandwidth, drive voltage and loss set the ceiling for everything downstream.

Live simulation · PAM4 symbol stream · push–pull ±Vπ/2 · the same signal drawn electrically, then optically
01 — How I work

A physicist's habits, a process engineer's constraints

I came to photonics from physics — a BSc at METU, an MSc at the Freie Universität Berlin — and I still design like a physicist: start from the field, then find out what the process line will actually let you build.

Working close to fabrication changes how you design. A layout is a hypothesis and the wafer answers in weeks, so the interesting question is rarely what the simulation says — it is which part of the simulation the process will honour, and how few simulations you need to find that out. The three figures on this page are the three parts of that interest: one modulates, one computes, one searches.

Bartu Yaman · Berlin · printed in four inks
02 — What I am working on

High-speed systems on chip

Above a few tens of gigahertz the electrical and the optical design stop being separable: the microwave line and the optical mode have to travel at the same speed, over the same length, without the metal spoiling the light it is steering. So they are designed as one problem — transmission line, waveguide and drive together. That electro-optic co-design is what sets the real RF and optical performance of any electrical-to-optical or optical-to-electrical conversion.

Computing in propagation

A diffractive layer is a weight matrix you cannot see. Stack a few in a slab and the interference performs the multiply–accumulate in the time the light needs to cross the chip — no clock, no memory traffic, and no energy beyond the laser that lit it.

Surrogate modelling & advanced design

A full-wave sweep of a modulator is hours per point, and the design space has more dimensions than patience. So the sweep is replaced by a model of itself: sample a few designs, fit a surrogate to them, and let the acquisition decide where the next simulation is worth running. The search happens on the model; the solver only confirms.

Heterogeneous integration

A passive platform routes light well but cannot modulate it; an active crystal modulates beautifully but makes a poor circuit. Bond one onto the other and each does the job it is good at. The work is in the seam — mode transitions between layers, tolerance to bonding and film thickness, and how much of the crystal's figure of merit survives into the circuit. The method outlives any one material.

Fig. II — Conscious parameter-space exploration & surrogates

Six design knobs and three numbers that fight over them: half-wave voltage, bandwidth, insertion loss. Only two knobs fit on a sheet of paper — the surface is the slice the other four currently cut, and it re-forms whenever they move. It is not the truth either, but a surrogate fitted to the designs sampled so far. The walker goes where the model says the cost is low and where the model is least sure.

Live radial-basis surrogate over sampled designs · expected-improvement-style acquisition, explore then exploit · the objective weighting changes each round

Fig. III — On-chip diffractive neural networks

The input is written as light in a slab, then crosses three phase masks — metalines of sub-wavelength scatterers. Propagation between them mixes every input into every output: the masks are the weights, the interference is the arithmetic, and the brightest detector is the answer. Move your pointer across the figure to steer the input.

Live scalar diffraction · angular-spectrum propagation · λ 1550 nm · 256 samples across a 218 µm aperture · three phase masks solved by wavefront matching onto four detector bins
03 — Where I have been

Research Associate · Fraunhofer HHI, Berlin2025 —

Research Assistant · Fraunhofer HHI, Berlin2023–25

MSc Physics · Freie Universität Berlin2022–25

Undergraduate Researcher · METU, Ankara2020–22

BSc Physics · METU, Ankara2018–22

Summer of HPC · PRACE, Ljubljana2021

Say hello

Best reached by email. I am always glad to talk about modulators, diffractive optics, or anything that turns a wafer into a system.

bartuyamann@gmail.com linkedin.com/in/brtymn