M.S. Candidate: Baran Özden
Program: Information Systems
Date: 25.06.2026 / 14:30
Place: A-212

Abstract: Foundation models have revolutionized artificial intelligence in language and vision, yet no comparable architecture exists for the continuous production environments of heavy industry. Today, process industries deploy task-specific models that treat a unified physical plant as disconnected parts. This thesis argues that the path forward lies neither in these fragmented models nor in universal time-series forecasters, but in domain specialized architectures with necessary inductive biases. To this end, we propose Prometheus, a compact 1.6M-parameter bidirectional Transformer encoder that processes multivariate sensor data as a language-like sequence, with fully-fused spatio-temporal attention. The proposed model is pretrained by self-supervised reconstruction on five years of crude distillation unit data through Four-Teacher geometric masking strategy, then tested on reconstruction and downstream quality tasks. Results show that, on short-horizon forecasting, Prometheus leads on all 38 channels across every metric (mean intra-patch R² 0.76). Its margin is widest on the coupled flow and level channels, where only Prometheus stays positive in R² while every baseline falls below zero, despite being up to ~300× smaller than zero-shot foundation models. It also achieves a virtual sensing capability the baselines cannot natively perform: reconstructing entirely missing channels from cross-channel inference alone (mean intra-patch R² 0.84), indicating it has captured the system's underlying thermodynamics. Fine-tuned as a soft sensor, it predicts heavy-diesel T90-T95 qualities more accurately than every competitor, including the deployed T95-Specialist. Prometheus offers a blueprint for an Industrial Foundation Model: a first step toward a unified intelligence that can interpret, and ultimately control, the physical reality of heavy industry.