About Milan Rother

I build simulation tools and infrastructure and solve numerical modeling problems for engineering teams.

Long ago during my studies I was already building small numerical side projects. A symbolic circuit analysis tool, an eigenmode solver for the Schrödinger equation, a CUDA-accelerated 2D FDTD solver, a MIMO vector fitting implementation, and so on. I even wrote a 120-page LaTeX manuscript on linear system theory during the summer of 2021 to help out a friend.

During my master's I dipped my toes into the RFIC EDA world and worked on compact modeling and layout generation of RFIC passives (inductors, transformers, baluns...). This has now become RapidPassives, an open-source library.

At the beginning of my PhD (early 2023) I diverged from the circuit path to the more general system level, where you have signal flow diagrams and ODEs instead of Maxwell and KVL/KCL. During that time PathSim started as a side project. Initially just an analog computer emulator, it evolved into a fully fledged continuous and discrete (hybrid, we have events!) system modeling and simulation framework. I announced it in March 2025 as an open alternative to Simulink, and a modern take on system modeling in general. Clean API, mutability from the get-go, robust solvers (stiffness, algebraic loops, no problem). Today (mid 2026) we are at roughly 400 stars on GitHub and have many users and collaborators worldwide, some of them prominent figures such as the MIT Plasma Science and Fusion Center, scikit-rf, and JSBSim developers.

Now I'm making the full roundtrip back through circuits and fields, bringing it all together into one vertically integrated stack: EM solvers (RapidFEM, RapidMoM), a symbolic circuit engine (SANE), and system-level simulation (PathSim, FastSim). One architecture: SSA-style compute graphs at the heart of the engines, Rust cores, Python APIs, browser interfaces. The commercial EDA world splits these levels across vendors and decades of legacy. I think one coherent, modern stack can do better.

Those who have talked to me know that I can't stop talking about numerics, modeling, and simulation principles and ideas.

I work as a freelancer and license my tools commercially, free for academia. If you have something you want my take on, or need some ideas and even implementation, don't hesitate to reach out.

Timeline

• 2026: one stack. The full roundtrip back through circuits and fields, built as one stack rather than six tools: RapidFEM and RapidMoM for electromagnetic fields, RapidMesh generating what RapidMoM solves on, RSLAB underneath as the sparse direct solver, SANE for symbolic and numeric circuit analysis, FastSim compiling system models to native code. One architecture throughout: SSA-style compute graphs at the heart of the engines, Rust cores, Python APIs, browser interfaces. The commercial EDA world splits these levels across vendors and decades of legacy; the bet is that one coherent stack does better. Most of it is one year: RapidFEM and FastSim from April, then SANE, RSLAB, RapidMoM and RapidMesh from June, the fields, circuits and foundations levels growing together rather than in sequence. RSLAB took PARDISO's place and is benchmarked against it. RapidMesh is the least finished: its 2D path meshes for RapidMoM and does exact element budgets, which gmsh does not, while the 3D path is slower than gmsh and still leaves slivers on some geometries, so RapidFEM meshes with gmsh a while longer.
• 2025-2026: fusion. The MIT Plasma Science and Fusion Center and the UK Atomic Energy Authority build tritium fuel-cycle models on PathSim: a zero-dimensional residence-time model of an ARC-class power plant, a one-dimensional liquid-metal bubble column, and full FESTIM finite-element transport, all coupled inside one timestepping loop. The same system model runs against lab experiments like BABY and LIBRA and against whole power-plant concepts, and a thirty-component bubbler transient solves in seconds, which is what makes the Monte Carlo and uncertainty quantification runs affordable. PathView came out of the same collaboration: it started at the Plasma Science and Fusion Center as a way to build and inspect PathSim models visually. From November 2025 I rewrote it on my own stack and it moved to the PathSim organization, where it is hosted free for everyone. I presented the framework at the IAEA workshop on digital engineering for fusion energy in December 2025, and I am a co-author on the multi-fidelity fuel-cycle paper that came out of the collaboration.
• 2025: open source. The repository had been public since the summer of 2024, but March 2025 is when I actually announced PathSim: an open alternative to Simulink, with a clean API, mutability from the get-go, and solvers that handle stiffness and algebraic loops without ceremony. Published in the Journal of Open Source Software, released on PyPI and conda-forge, and picked up faster than I expected. It is past 400 stars and nine contributors. The domain toolboxes cover chemical engineering, batteries, vehicles, flight dynamics, RF and FMI co-simulation, several of them contributed by people I have never met. JSBSim picked it up for flight dynamics. PathView followed: the same models as drag-and-drop block diagrams in the browser, running entirely client-side through Pyodide, nothing to install. PySimHub came out of the same year, a catalog for Python's scattered simulation communities.
• 2024-2025: Analog Insydes, revisited. From December 2024 I worked with Ralf Sommer, the inventor of Analog Insydes, on reviving that tool: symbolic circuit analysis in Mathematica, from the Fraunhofer ITWM, largely out of use today. Picking those ideas back up on my own stack in June 2026 is what became SANE.
• 2023-2024: PathSim. Started in early 2023 at the beginning of the PhD as an analog computer emulator side project, PathSim grew into a full hybrid system modeling and simulation framework: custom solvers, an event system for discrete behaviour alongside continuous dynamics, and a block-diagram API built from first principles rather than copied from the tool everyone already had. It stayed on my own machine for a year and a half, versioned as 31 zip snapshots, before the repository went public in the summer of 2024. The earliest snapshot, December 2023, already contains Simulation, Connection and the block names the API uses today. The same stretch brought three more: a minimalistic harmonic balance framework for nonlinear periodic steady-state response, a symbolic circuit analysis tool inspired by Analog Insydes, and a finite volume solver for quasi-electrostatic frequency-domain analysis that came out of the sensor modeling at the institute. The solver needed a mesh, so it got one: a 2D QuadTree that refines toward edges and balances itself. Writing my own mesher for the first time is where the idea started that a solver and its mesher should be designed against each other, which is what RapidMesh is now.
• 2023-2025: research associate, CMOS design. Research associate at the Institut fuer CMOS Design, TU Braunschweig. Three strands ran in parallel: numerical modeling of electrochemical sensors for analog circuit design, system design and hardware for bio-impedance spectroscopy, and an EDA pipeline for RFIC passives that was validated in silicon for cryogenic quantum applications. That last one started as a Tkinter window that drew inductors and exported GDS. It became RapidPassives: browser-based layout generation with real-time preview and a WebGL viewer that renders million-polygon GDS at 60 fps.
• 2023-now: PhD candidate, TU Ilmenau. External PhD program in electrical engineering, run alongside the work in Braunschweig and everything that grew out of it.
• 2022-2023: first repos. The first experiments that went public: an interactive 2D electromagnetic FDTD environment with a live field view, a fast relaxed vector fitting implementation that turns measured MIMO frequency responses into compact circuit models, and an interactive Bode plot tool for teaching. Vector fitting had been sitting in a folder since 2020 as rational interpolation experiments, complex fitting by iteratively reweighted least squares, and a stack of papers on passivity enforcement. December 2022 is when it went into a repository.
• 2021-2022: things you could click. A run of interactive desktop toys, all pygame, all packaged with PyInstaller into a single executable so I could hand someone a file instead of an install guide. A Verlet physics sandbox with numba-compiled constraint solving, a few thousand particles on collision courses, and a random polygon generator built out of packed circles. The one I kept using was the interactive Laplace domain: drag poles and zeros around the complex plane and watch magnitude, phase and step response follow in real time. It went into the tutorials.
• 2021-2023: M.Sc. electrical engineering, TU Braunschweig. Master of science with honors, thesis graded excellent. Written at the Institut fuer CMOS Design between October 2022 and March 2023: area optimized neural network based modelling of passive magnetic structures for chip-level integration. Three things had to be built for it. Polygonal windings parametrised down to the geometry, so a spiral inductor is a handful of numbers rather than a drawing. Infineon's B12HFC layerstack reduced to the layers an EM solver actually needs, because the full stack makes the simulation unaffordable. And a completely analytical derivation of the first and second partial derivatives of RLCk transfer functions with respect to the circuit parameters, read straight off the block structure of the MNA matrices. Those derivative stencils are what SANE now runs symbolically over a whole compute graph, and the windings and layerstacks are what RapidPassives generates today. At the time the point was to stop the optimizer finite-differencing.
• 2021-2023: research assistant, CMOS design. High-frequency measurement calibration methods, design and implementation of verification hardware, teaching the circuit theory seminar.
• 2020-2021: writing it down. The 120-page LaTeX manuscript on linear system theory, written over the summer of 2021 to help out a friend: Laplace analysis, poles and zeros in the complex plane, frequency response, state space. It has since been translated to English. A separate writeup on state-space representations of linear time-invariant network models came first, in 2020. And for the tutorials I animated the parts that are hard to draw once on a blackboard: the rotating phasor, resonance sweeping through a Bode plot, the step response of a damped oscillator.
• 2019-2021: before GitHub. The first real project never made it to a repository. MiCir was a symbolic network analysis library written in Python from nothing: my own matrix class, my own complex arithmetic, my own polynomial and symbolic element types in s. Not because there was no numpy, but because I wanted to know what numpy was doing before I let it do it for me. It read netlists, solved networks symbolically, synthesised Cauer ladders from a driving-point impedance, and drew its own Bode and locus plots. In February 2020 I typeset the whole thing in LaTeX. It is the same problem SANE solves today: keep the circuit's symbolic structure instead of collapsing it into numbers, and read the answers off it. The same folder holds the detours: Mandelbrot sets in the complex plane at resolutions my laptop resented, Monte Carlo estimates of pi, a DFT written out by hand.
• 2019-2020: internship, Volkswagen AG. Low-voltage and redundant power supply systems in Wolfsburg: concept, schematic, hardware, ASIL considerations.
• 2018-2021: teaching assistant, TU Braunschweig. Weekly independent tutorials on linear dynamical systems and circuit theory for 50+ students. The Fachschaft tutorial for Netzwerke / Wechselstroeme und Netzwerke came with its own problem set, typeset and drawn from scratch and never quite finished: networks with switching instants, initial values, asymptotic stability, harmonic steady state.
• 2015-2021: B.Sc. electrical engineering, TU Braunschweig. Bachelor of science, thesis graded excellent. Written over the winter of 2020/21: order reduction of interconnect network models for efficient calculation with SPICE simulators. A reduced RLC network model is only usable if it stays passive and stable, and both properties are readable off the structure of the MNA system matrices before any reduction happens. That is where I first met rational fitting, passivity enforcement and the papers that later became my vector fitting implementation.
• 2014-2021: climbing coach, DAV Braunschweig. Led the competition climbing team from entry level to international level, over seven years.

Publications and writing

• PathSim: a system simulation framework. Journal of Open Source Software 10(109), 8158, 2025. [ doi ]
• Physics-informed tritium fuel cycle modelling workflow for fusion reactors, with R. Delaporte-Mathurin, R. MacDonald, J. Dark, T. Zulfiqar and K. B. Woller. arXiv:2603.25751, 2026. [ arxiv ]
• I built an open-source Simulink alternative that runs in your browser. dev.to. [ article ]
• Linear time-invariant systems: a 120-page manuscript on system theory, in German and English. [ de ] [ en ]

Talks

• PathSim: block diagram simulation in pure Python. PyCon DE & PyData, April 2026. [ talk ]
• An open-source Python framework for dynamic system simulation in fusion energy applications. OSSFE 2026, March 2026.
• PathSim, at the IAEA Workshop on Digital Engineering for Fusion Energy Research, MIT Plasma Science and Fusion Center, December 2025. [ workshop ]
• PathSim, at the GAMM RSE/RDM workshop, December 2025.
• The decentralized perspective on dynamical system simulation. Guest lecture at the MUSEN colloquium, TU Braunschweig, May 2025.
• PathSim, at Gauss Fusion, May 2025, and at IFAS, May 2025.
• Die Anatomie eines Systemsimulators. IMMS, March 2025.

Curriculum vitae (PDF)