00 / AI-CEM RESEARCH PROGRAMME

From Maxwell's equations
to working systems — a 30-year roadmap.

AI-CEM (AI-Driven Computational Electromagnetics) is a single research programme built around a four-layer stack. This page brings together its four phases, the P1–P8 publication chain, the keyword rings, the funding ladder, and the 6-box project doctrine. Publications that talk to each other; layers that feed each other; goals that follow each other.

THE 30-YEAR RESEARCH QUESTION

How can AI and computational physics be integrated to solve electromagnetic forward and inverse problems faster, more accurately and more autonomously?

01 / PHASES

Four phases, one unified field.

2026 → 2056+

2026 — 2031 PHASE A · PhD

Differentiable CEM + first AI integration

Build the foundational math and solver capacity. P1–P3 deliver the differentiable FDTD backbone, graph+gradient hybrid optimisation, and Hessian/curvature-aware inverse design. P4–P5 add FNO/GNO and a hybrid AI–physics multi-fidelity solver. P7 closes with a real antenna/RF/6G component. P8 ships an open-source benchmark + dataset.

JAX-FDTDGPU autodiffFNO/GNOMemory-efficient gradientAntenna / RISOpen source

2031 — 2037 PHASE B · EARLY CAREER

AI-native EM platform

FNO/GNO/PINO towards an EM foundation model. The AI-CEM platform: dataset + surrogate + solver + benchmark. MSCA PF postdoc, TÜBİTAK 1001/3501, ERC Starting Grant. Begin the Turkish associate-professor file under ÜAK 90516.

EM foundation modelMSCA PF1001/3501ERC StartingÜAK 90516

2037 — 2045 PHASE C · GROUP LEADER / ASSOCIATE PROFESSOR

Autonomous electromagnetic engineering

Given a specification (frequency, gain, bandwidth, volume), the AI-CEM system generates topology, simulates, optimises, quantifies uncertainty, and checks manufacturability. QuevaTech bridge into RF fingerprinting and EM verification. Associate professorship (ÜAK 90516).

Autonomous designEM verificationQuevaTech bridgeHPC partnerships

2045 — 2056+ PHASE D · FIELD LEADERSHIP

Quantum-classical CEM + industrial impact

Quantum numerics (Maxwell Hamiltonian simulation), AI-native solvers, open standards, quantum-HPC hybrid infrastructure and a spin-off company. The 30-year research question stays; the answers change.

Quantum numericsAI-native solversQuantum-HPC hybridOpen standardsSpin-off

02 / PUBLICATION CHAIN

P1—P8: publications that feed each other.

A research programme, not a collection

P12026-27

Differentiable FDTD backbone

Period Fall → SpringHypothesis Autodiff opens the door to EM inverse problemsOutput Open-source solver + validation

JAX-based differentiable FDTD. Gradient accuracy vs analytical. Conference: ICCEM / ACES. The first academic anchor for differentiable simulation.

P22027

Graph + gradient hybrid optimisation

Period FallHypothesis Topology search + local sensitivityOutput Algorithm paper

Structural/topological information combined with gradient refinement. Convergence / speedup gains. Conference: EuCAP / IEEE APS. Anchors numerical optimization.

P32027-28

Curvature-aware (Hessian) inverse design

Period SpringHypothesis 2nd-order information → faster / more stableOutput H-vector, Newton-CG, learned preconditioner

Hessian-vector product, Newton-CG, Gauss-Newton, quasi-Newton. Journal: TAP or JCP. Combines differentiable simulation + numerical optimization.

P42028

Neural operators for EM forward

Period FallHypothesis Accelerate the full-wave solveOutput FNO-multiscale; error vs full-wave + speedup

Learn the operator from FDTD data. Opportunity for a field review. Anchors neural operators + surrogate modelling.

P52028-29

Hybrid AI–physics multi-fidelity solver

Period SpringHypothesis Accuracy–budget balanceOutput The most valuable paper

FNO gives a fast prediction; high uncertainty triggers the full-wave solve; results retrain the AI. Multi-fidelity concepts applied to EM.

P62029-30

AI-native inverse design (integration)

Period Thesis climaxHypothesis Combine every pieceOutput Thesis backbone

Graph + operator + differentiable solver + 2nd-order optimisation, all in one framework. Builds the full electromagnetic inverse design fingerprint.

P72030-31

Physical application: antenna / metasurface / 6G

Period Final yearHypothesis Real-world evidence of the methodOutput Design + fabrication + measurement

Design a real component, fabricate it, measure it. Simulation–measurement agreement. Anchors antenna design / microwave / 6G.

P8Y3

Benchmark suite + dataset paper

Period Year-3 reviewHypothesis Community standardOutput JOSS / Scientific Data / NeurIPS dataset

AI-CEM benchmark package + open dataset. Critical for citation growth and the community fingerprint.

03 / KEYWORD FINGERPRINT

Three rings. 3–5 core terms in every paper.

The mechanism behind "that Yiğit in this field"

Core 6

computational electromagnetics FDTD electromagnetic inverse design numerical optimization differentiable simulation antenna design

Ring 2

scientific machine learning neural operators surrogate modelling physics-informed GPU/HPC microwave engineering

Ring 3

6G metasurfaces RIS mmWave arrays computational imaging

04 / FUNDING LADDER

Growing support, year by year.

Continuous + periodic

Y1BAP (university)small but continuous

Y1TÜBİTAK 2211PhD scholarship

Y2-Y3TÜBİTAK 2214-Aabroad research 6-12 mo

Y2-Y52224-Bconference support

Y3-Y5TÜBİTAK 1001 / 3501project + early-career

Y2+COST Action networksnetworking

2031+MSCA PFpostdoc

2034+ERC Starting Grant5-8 year horizon

05 / 6-BOX PROJECT DOCTRINE

Every project idea passes through these six boxes.

The systematised version of the mentor's "aim–method–output" advice

A method step with no measurable output is a problem. An output that connects to no objective is also a problem. These six boxes are the honesty check for every paper, thesis chapter and project proposal.

01ProblemWhy does it matter?

02Research gapWhy does the existing method fall short?

03ObjectiveWhat exactly are we changing?

04MethodWhich steps deliver that?

05EvidenceWhich metric shows success?

06OutputPaper / solver / dataset / prototype?

Return to the home page to see the whole stack.

Computational Electromagnetic Systems — 4 layers, one unified field. From Maxwell's equations to working systems.

See the stack ↓ Open portfolio ↗

From Maxwell's equationsto working systems — a 30-year roadmap.