Felix
Stocker

CrashDefNet is a graph neural network pipeline I developed entirely on my own, predicting the deformation of finite element meshes directly on the graph structure. Every mesh node becomes a graph node, element connections become edges; through iterative message passing the network propagates the mechanical interactions between neighboring nodes and reconstructs the post-crash displacement field from them.

Architectures

Three custom models compared head-to-head — an EmbeddedGNN with learned node embeddings, a Graph Transformer (GTR) with attention over the neighborhood, and an MGN-Transformer following the MeshGraphNet principle (Encoder–Processor–Decoder).

Data pipeline

Custom mesh-to-graph preprocessing including face extraction, construction of node and edge features, and an adaptive test infrastructure for fast iteration cycles.

Infrastructure

Two-tier setup — local development under WSL, GPU training on a remote H100 server; reproducible conda/Miniforge environment on Linux.

Objective

Data-driven approximation of expensive crash solvers — predicting component deformation in seconds instead of hours, as an accelerator for design loops in passive safety.

The CBR-augmented R&D-Agent extends an autonomous LLM research agent (Microsoft's R&D-Agent) with Case-Based Reasoning (CBR). Instead of starting every task from scratch, the system stores experiences from previous experiment iterations as structured "cases" and retrieves the most relevant ones for new tasks — the classical CBR cycle of Retrieve, Reuse, Revise, and Retain, applied to data-driven ML research.

Knowledge base

Custom seed knowledge base builder; similarity-based retrieval via fingerprints and embeddings with an upstream relevance filter.

LLM backend

Gemma 4 31B Dense via Google AI Studio.

Evaluation

Reproducible comparison pipeline (eval_cbr, plot_cbr) on the Kaggle competitions NOMAD 2018 and Spaceship Titanic — with per-iteration metrics, directional IQR outlier detection, Mann-Whitney U test, Cohen's d, and bootstrap confidence intervals.

Publication

Own preprint (arXiv, cs.SE) in IEEE format, with a fully verified bibliography (43 references) and a conference poster.

A production-ready system for automated evaluation of crash test sensor signals. It combines two data modalities: a ResNet-18 processes the signals as an image, an MLP processes the extracted tabular features; both representations are fused for the final OK/NOK classification.

Feature engineering

15+ features across three domains — time domain (peak amplitude, RMS, crest factor, kurtosis, skewness, zero-crossing rate, derivatives), frequency domain (spectral centroid, spectral bandwidth, energy concentration via Welch's method), and statistics (mean absolute value, variance, coefficient of variation).

Preprocessing

MAD-based outlier detection, signal trimming, and min-max normalization.

Evaluation

5-fold stratified cross-validation with bootstrap confidence intervals (1000 iterations), balanced accuracy, Matthews correlation coefficient, and a custom W-score metric.

Processing

Concurrent channel processing via ThreadPoolExecutor, TensorBoard integration.

From just four scenario parameters — impact position, impact angle, ego and opponent velocity — this system predicts complete acceleration time series of 400 to 1000 time steps. It's a sequence-to-sequence problem for which I implemented and systematically compared six architectures.

Models

1D-CNN (temporal convolution), classical LSTM (RepeatVector + two LSTM layers), GRU variant, hybrid CNN-LSTM with BatchNormalization and Dropout, an MLP baseline (64→128→256→512), and a PyTorch CrashNet with Conv1d encoder and ConvTranspose2d decoder.

Data preparation

LS-DYNA raw data with resampling, unit conversion (mm/s² → G), StandardScaler normalization, and JSON export.

Statistical grounding

Evidence of nonlinearity in the data via Mutual Information, the Ramsey RESET test, and polynomial regression (degree 2–3) as justification for deep learning approaches; partial dependence plots via Random Forest.

From scratch

Custom gradient descent implementation with a parametric pulse function (exponential + sinusoidal); target quantities include peak, mean, and AUC acceleration.

Training

Callback system (EarlyStopping, ReduceLROnPlateau, ModelCheckpoint), GPU acceleration, metrics MSE/RMSE/MAE/R²/MAPE.

arTIco is a standardized development environment for multivariate time series classification that I co-developed and released on GitHub. It enables systematic experimentation with several deep learning architectures under uniform data and evaluation logic.

Architectures

Configurable 1D-CNN, LSTM-FCN (LSTM branch + FCN branch + attention), and InceptionTime (LSTM + Inception modules + Squeeze-Excite).

Optimization

Optuna-based search with a custom search space (batch size, conv layers, filters, adaptive kernel size, pooling, BatchNorm, dense layers, dropout, learning rate on log scale); an experiment class encapsulates the pipeline (prepare → run → get_balanced_accuracy_score → optimize).

Evaluation

5-fold stratified CV with bootstrap CIs, balanced accuracy, per-class recall, weighted precision, F1, and Matthews correlation.

Collaboration

Joint release with F. Plaschkies (CARISSMA/THI) and Z. Yuan (IDIADA).

A complete end-to-end computer vision system that detects and classifies EV charging stations from images — from automated data acquisition through training to production serverless deployment on AWS. The MLOps pipeline consists of six clearly separated stages.

1 · Data acquisition

CSV-based URL filtering by model name, parallel image downloading from PlugShare (ThreadPoolExecutor, up to 10 concurrent workers).

2 · Auto-labeling

YOLO predictions with JSON export, crop extraction, Base64 encoding, and configurable confidence threshold.

3 · Quality filter

ResNet50 feature extraction with cosine similarity filtering (threshold 0.8), multiprocessing over 8 workers.

4 · Dataset

Conversion to YOLO format, URL-based train/val split (80/20), class mapping, and YAML generation.

5 · Training

YOLO training locally or via AWS SageMaker (PyTorch estimator with GPU).

6 · Deployment

Serverless inference via AWS Lambda in a Docker container (ECR), accessed through a Lambda caller API from a Tkinter desktop GUI and a Flask web app.

The AIDA Toolbox is an internal automation system that orchestrates LLM-based analysis of source code and presentations. It provides two pipelines and is separated into clean layers — presentation (Streamlit), business logic, Selenium integration, and infrastructure (shell).

PPT→HTML

Headless conversion of PowerPoint to PDF via LibreOffice, multi-stage upload with verification, detection of LLM streaming completion, and HTML extraction of the result.

Code→HTML

Recursive file discovery (40+ extensions), validation against limits (100 files, 20,000 lines), intelligent conversion of non-native files to .txt and path flattening; typed data classes (CodeFileInfo, ValidationResult).

Automation

AIDAAutomation class based on Selenium WebDriver (Firefox/geckodriver or Edge) for full browser control.

Security

Path traversal protection (os.path.commonpath), hardened file permissions (0o700/0o600), filename sanitization, whitelist authorization, and PID-based lock file management.

Meta

The architecture reports that feed this portfolio were generated with exactly this tool.

A methodology that systematically transfers expert knowledge into a machine learning model to automate the correlation between virtual and physical models in passive vehicle safety. Structured expert interviews yield evaluation criteria and rating catalogs; a model trained on simulation data and expert judgments then takes over the validation of virtual models, complementing classical objective correlation methods.

Application

Validation of WorldSID-50M crash test dummy models.

Presentation

Presented at the crash.tech conference at Technische Hochschule Ingolstadt.

A calibrated and validated sequential thermo-mechanical finite element analysis of an additively manufactured stent lattice structure (Laser Powder Bed Fusion). The work investigates how geometric parameters — strut angle, strut thickness, and transition radius — govern distortion, and shows that FEA-based inversion of the predicted deformation brings the part geometry back close to the CAD target.

Significance

Peer-reviewed publication in Materialia (Elsevier, 2023) — the engineering foundation of simulation, optimization, and scientific validation on which my data science work builds.