Master Thesis: AI‐Driven RF Inverse Design of Multi‐Layer EM structures

Shape the Future of Quantum Computing: Write Your Master's Thesis at SCALINQ Are you a Master's student in physics, engineering, or a related field with a passion for turning theory into reality? SCALINQ offers a unique opportunity to conduct your thesis work at the cutting edge of quantum technology. We believe that the most impactful work comes from a combination of academic curiosity and industrial application. That's why we invite talented students to join our team, not just for a project, but for a chance to contribute to the core of our mission. Why write your thesis with SCALINQ? Solve Real Challenges: Work on problems that are critical to making quantum computers scalable and powerful. Your research will have a direct application Expert Mentorship: Be guided by our team of experienced scientists and engineers in the field Access to State-of-the-Art Technology: Get hands-on experience in a dynamic, high-tech environment A Pathway to a Career: A successful thesis project could be a stepping stone to a full-time role with us. We invest in talent for the long term We are looking for individuals who are: Driven by curiosity and a desire to innovate Collaborative, proactive, and eager to learn Passionate about building the hardware or contributing in closely related areas that will power the quantum revolution Master Thesis Topic: AI‐Driven RF Inverse Design of Multi‐Layer EM structures Suitable period: Jan 2027 - June 2027 Length: 30 credits (ECTS / hp) Number of students (min/max): 1/2 Suitable student background: Artificial Intelligence, Complex Adaptive Systems, Computer Science, Electrical and Electronics Engineering, Software Engineering Description: Summary & Objecti ve Microwave/RF components in cryo-quantum computing systems require compact size, high selectivity, and fast design cycles. Planar 2D pixelated layouts provide design flexibility, but their performance is constrained by single-layer physical limits. 3D multi-layer structures solve this problem, but simulating and tuning 3D layouts manually is computationally heavy and time-consuming. This project, based on previous work in NN-assisted 2D pixelated LPF inverse design, explores the possibility to expand the design space from single-layer to multi-layer EM structure. Combining RF engineering and computer science, we integrate deep generative models with open-source EM solvers (EMerge, OpenEMS). The framework generates DRC-compliant, manufacturable layouts directly from target RF specifications. Key Value: Design Automation: Transitions filter design from manual trial-and-error to an automated "specs-in, layout-out" workflow. Reusable Toolchain: Improve the open-source Python pipeline for batch simulation and AI inverse design that the team can reuse for future projects. Fabrication-Ready Output: Integrates manufacturing rules (DRC) directly into the generative model, producing optimized 3D design (PCB filter with absorber or optimized transistor core) Team Structure & Division of Responsibilities The workload is divided between two students based on their technical backgrounds to cover both electromagnetic modeling and AI model implementation: Student 1: RF Engineering Focus Substrate & DRC Definition: Define substrate properties (e.g., Rogers 4350B), layer stack-up, via geometry, and manufacturing DRC limits (trace width and clearance). Baseline Design & Benchmarking: Design traditional multi-layer components manually (e.g., filter with absorber) to serve as a performance benchmark for the AI model. Computational EM & Data-Efficiency: Find a way to lower the EM computational cost in dataset generation. Performance Analysis: Perform yield and sensitivity analysis on final layout candidates to evaluate performance variations against manufacturing tolerances. Analyze E/H field distributions and resonant coupling mechanisms of the generated topologies to explain the physical working principles behind the designs. Student 2: Computer Science Focus Simulation Automation Pipeline: Write Python scripts to parameterize multi-layer layout grids, run batch EM simulations in parallel, and parse S-parameter data. Neural Network Implementation: Build and train both the forward surrogate model (e.g., CNN/ResNet) and the inverse generative network (e.g., Tandem/cVAE). DRC Constraints & Layout Export: Implement connectivity checks and DRC penalty terms in PyTorch, and create the tool to export final layouts to Gerber/DXF format. (Optional) Potential integration of LLM: Deploy LLM to guide or even build the pipeline for the design workflow. Technical Approach & Workflow The framework operates via a systematic two-phase pipeline designed for low computational latency and high reliability: Phase 1: Training Phase Dataset Generation: The RF student sets up the multi-layer PCB templates (e.g., 2–4 layers on Rogers 4350B). The CS student writes Python scripts to run batch simulations and extract the S-parameter data (S11,S21). Model Training: The CS student trains the forward models and DRC-constrained inverse networks. The RF student reviews the results to verify physical accuracy. Phase 2: Inverse Design Phase Layout Synthesis: Target S-parameters are fed into the trained model, which generates DRC-compliant 3D PCB layouts. Validation & Export: The CS student automates the loop to verify these generated layouts in EMerge and export them. The RF student benchmarks the final designs, comparing the footprint and selectivity of the 3D filters against traditional 2D filters. Project Scope & Boundaries To keep the project feasible within the 4–5 month timeline, the scope is defined as follows: Target Component: Focuses specifically on one or two demo components (e.g., transistor core for internal evaluation, passive filter for thesis presentation). Manufacturing Process: Limited to standard multi-layer PCB technology (2–4 layers). About us : SCALINQ is a prominent player in the growing field of quantum compu