Nathan Belmore Email & Phone Number

Currently, I am employed as a senior machine learning engineer at TicketSwap, where I have primarily focused on enhancing the MLops pipeline. The existing pipeline was lacking in crucial components such as experiment tracking, model registry, data versioning, metrics and hyperparameter logging, orchestration of compute resources (including pipelines, pools, agents, and autoscalers), as well as monitoring and automation functionalities. Despite being part of a small team, I also undertake various data science projects and serve as a reliable resource for statistical modeling, develop ML models, and advise other MLEs & data analysts. During my tenure as a graduate student at the Max Planck Institute for Plasma Physics in Greifswald, I had the privilege of contributing to the development of the Wendelstein 7-X (W7X) experimental fusion reactor. I collaborated with the NEISS (Neural Extraction of Information, Structures, and Symmetries from Images) research group, which focuses on designing an automated control system for the W7X. This ambitious endeavor will span years and encompasses projects involving Generative Adversarial Networks (GANs), reinforcement learning, and Deep Neural Networks (DNNs).My contributions to the W7X project can be categorized into two integrated ML projects. The first project involves employing Convolutional Neural Networks (CNNs) to analyze thermal load images captured from the experiment's internal thermal cameras. The goal is to reconstruct crucial plasma parameters, such as the toroidal current and rotational transform, which are challenging to directly measure or determine accurately. To produce the training and validation data, simulations were conducted using real-world experimental results to serve as proxies for the true rotational transform. By incorporating these simulated results into the CNN framework, we can infer the rotational transform with greater precision.The second area of my work centers around a diffusion model similar to DALL-E, which is utilized to generate heat load images based on given plasma states. To ensure the fidelity of the generated heat load images and preserve the inherent physics, the model is fine-tuned using the aforementioned CNN in the loss function, akin to a Physics-Informed Neural Network (PINN) approach. This integration allows us to validate the generated heat load images adhered to the underlying physics principles.