Research & Writings

TempTPI: Informer-Based trajectory prediction for maritime vessels
K. Ferneding, V. Lietavcova, A. M. Blachowiak, P. Heiselberg
Proceedings of IEEE International Geoscience and Remote Sensing Symposium (IGARSS 2026), Washington, D.C., USA · 2026
Status: In Press (Oral Presentation on August 10, 2026)

Abstract: Predicts long-term maritime vessel trajectories from AIS data by combining an Informer architecture with a multi-channel temporal encoding mechanism. By leveraging ProbSparse self-attention and Fourier-based frequency expansions to capture cyclic vessel behavior, TempTPI reduces computational complexity while outperforming state-of-the-art models (TPTrans), achieving a 55% reduction in MSE at a 5-hour forecasting horizon.

[Postprint (PDF)]

© 2026 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future media.

A Hybrid Approach to Automated Test Case Generation using Formal Methods and LLMs
K. Ferneding
M.Sc. Thesis Project · Technical University of Denmark (DTU) · Airbus Defence & Space GmbH · 2026

Abstract: Evaluates a novel neuro-symbolic framework that combines Large Language Models with formal abstract syntax tree (AST) transformations to automate safety-critical test case generation under strict regulatory standards (e.g., DO-178C). By restricting the LLM to semantic translation and offloading test derivation to deterministic Boolean logic, the system mitigates hallucination risks while significantly improving test correctness, consistency, and traceability compared to purely generative baselines.

Note: Full thesis report and code are non-public due to corporate confidentiality.

On the fragility of pluralistic ignorance
K. Ferneding, S. Stange, F. T. Dollas, J. B. Weikop
Course Project: Logical Theories for Uncertainty and Learning · Technical University of Denmark · 2025

Abstract: Analyzes the computational dynamics and fragility of pluralistic ignorance through the lens of The Emperor’s New Clothes. Replaces the standard assumption of belief independence with a Bayesian model for non-independent belief revision under collective misperception. Demonstrates through a multi-agent simulation framework that the sudden collapse and dissolution of pluralistic ignorance can be effectively modeled computationally.

Minchia: A fast, JPS-based Algorithm for Multi-Agent Path Finding
C. Zhou, K. Ferneding, S. G. Murgia Kristensen, S. Wang, T. K. Luyen
Course Project: Artificial Intelligence for Multi-Agent Systems · Technical University of Denmark · 2025

Abstract: Solves complex Multi-Agent Path Finding (MAPF) and box-pushing problems by combining Jump Point Search (JPS) with Conflict-Based Search (JPCBS) to accelerate grid pathfinding and resolve agent interferences. Minchia utilizes map submapping alongside a dynamic Goal Dependency Graph to detect deadlocks and automatically assign intermediate storage subgoals for blocking obstacles. Evaluated across benchmark maps like SAsoko3 and Bispebjerg Hospital, it reduces runtime by over 10–100 compared to standard A* search.

Awards: Winner (1st place out of 39 in 4/5 categories) at the final course competition

Artificial pollination systems using machine learning and remote sensing
A. Maftei, K. Ferneding, M. M. Volkert
Course Project: Intelligent Systems · Technical University of Denmark · 2024

Abstract: Proposes an autonomous drone-based system to address global pollinator decline through targeted artificial pollination. Combines real-time flower detection via YOLO object detection with a soap-bubble dispensing mechanism, reducing pollen consumption by up to 99% compared to traditional methods. Deploys a LiDAR sensor and PID flight control on a modified DJI Mavic 3M platform with edge computing for precise altitude calibration and automated flower-approaching maneuvers.