Machine Learning for Signals and Scientific Data

1Research

My work is machine learning for signals, time series and other scientific data. Figure 1 gives an overview: the methods I work with, and the listed pieces of work in which each has been used. The current application domain is seismic recordings. At the Kandilli Observatory and Earthquake Research Institute (KOERI) I am a research assistant on TRACE,¹1TRACE project website: trace.bogazici.edu.tr. Its funding is noted in the Acknowledgements. where I develop and evaluate AI models for earthquake detection and location and build the scientific backend that connects data preparation, model training, inference and catalog analysis.

1.1Generative modelling

I work with diffusion models, flow matching, variational autoencoders and GANs. In research, I work on conditional generative models for seismic waveform synthesis, including latent diffusion. The question is how the parameters of an event and the physical properties of a signal can inform both how waveforms are generated and how the generated waveforms are evaluated. High-resolution waveform generation for the Marmara region with conditional latent diffusion is the subject of an oral presentation at ESC 2026 in Istanbul [1]. In teaching, the MATH482 term projects explore the StyleGAN2 latent space and conditional latent diffusion for image editing, and a course notebook covers variational autoencoders (Appendix A).

1.2Learning on structured data

I work with convolutional and residual networks, Transformers, and graph neural networks. A seismic network is a set of sensors whose recordings have to be read together, which makes it a natural setting for graph neural networks. I adapt and evaluate GENIE, a graph neural network model, for regional seismic networks,²2Code: a fork of GENIE at github.com/Pancarsuyu/GENIE. working on model training, inference efficiency, and comparison against observatory catalogs. AI-driven event detection and association for monitoring at KOERI is the subject of a contribution to the EGU General Assembly 2025 [2]. Residual learning is also one of the MATH482 topics (Table A1).

1.3Probabilistic and physics-informed learning

I work with probabilistic neural networks; with mixture density networks, which predict a full conditional distribution instead of a single value; and with physics-informed neural networks (PINNs). The first two, together with the derivation of loss functions, are part of the MATH482 practical material (Table A1). Physics-informed networks are part of my current work.

Across all of this the common requirement is evaluation: generated signals judged with physically informed criteria, detections compared against observatory catalogs, and coursework that ties each method to an experiment.

2Selected work

The two projects on seismic data are described in Sections 1.1 and 1.2. Three further pieces of work, on economic time series and language models, are listed here; all five appear in Figure 1.

2.1Machine learning for inflation forecasting — time series

A replication of an inflation forecasting study, extended with feature engineering and regime-aware models on macroeconomic time series.

Code and reports: github.com/Pancarsuyu/EC48E-Project

2.2PatentAI — LLM applications

An LLM application for patent research and startup analysis. It combines tool calling, information retrieval and report generation behind a conversational interface.

2.3Turkish LLM fine-tuning dataset — language models

A Turkish question–answer dataset, co-created for the T3 AI Hackathon, with a workflow for generating synthetic data and assessing its quality.

Dataset: huggingface.co/datasets/barathanasln/turkish_llm_finetune_dataset_4_topics

3Teaching

I am a course assistant for MATH482, Machine Learning Methods and Neural Networks, at Boğaziçi University.³3Fall 2025 and Fall 2026. I design and lead the practical coursework in deep learning, probabilistic modelling and generative AI, with the aim of connecting the mathematical foundations to implementation and to careful experimental evaluation. The 2025 lab and assignment notebooks, published on Kaggle, and the term projects are collected in Appendix A.

4Experience

5Education and recognition

5.1Education

B.S. in Mathematics, Boğaziçi University, 2021 – present (fourth year).

5.2Recognition

AlgoRun Hackathon (Invent.ai & Compec), second place, 2025.
Kaggle Competitions Expert: kaggle.com/batuhankalem.

Acknowledgements

TRACE is supported by Google.org through a $1 million grant (university announcement).

References

Conference contributions, numbered in order of citation.

1. [1] Erdem Ata, Batuhan Kalem, Yusuf Sezer Car, Çağrı Diner, Gülüm Tanırcan, Nurcan Meral Özel, Yasin Arkan. High-Resolution Seismic Waveform Generation in the Marmara Region Using Conditional Latent Diffusion. 40th General Assembly of the European Seismological Commission (ESC 2026), Istanbul, 2026. Oral presentation, abstract #9688. esc2026.org/programme.
2. [2] Nurcan Meral Özel et al. (including Batuhan Kalem). Artificial Intelligence-Driven Seismic Event Detection and Association for Enhanced Monitoring at KOERI. EGU General Assembly 2025, Vienna, 2025. doi: 10.5194/egusphere-egu25-11172.