Motivation

The PtM (Power-to-Methanol) pilot plant at KIT is based on a Novel Hybrid Process that enables simultaneous CO₂ capture and methanol synthesis in a single integrated system. The plant has been fully constructed, thoroughly tested, and operated for several hundred hours under diverse conditions, generating a rich and well-validated operational dataset.

Currently, the plant is operated by human operators who monitor key process parameters, make real-time adjustments, and ensure safe and stable operation. However, the PtM plant is designed as a modular system, meaning that multiple identical units can be deployed together at scale. If each unit requires a dedicated human operator, the operational costs will become expensive as the number of deployed plants increases.

To address this challenge, the long-term goal is to achieve fully unmanned (autonomous) plant operation, where the plant can:

1. 1. Monitor its own state using real-time sensor data,
2. 2. Predict deviations from optimal operating conditions,
3. 3. Automatically adjust operating parameters to maximize methanol yield while maintaining safety constraints.

Autonomous operation of modular PtM plants represents a significant step toward economically viable large-scale deployment of PtX technology. Thereby, directly contributing to climate change mitigation efforts.

Currently, such unmanned PtM plant concepts correspond to a Technology Readiness Level (TRL) of 2–3 (basic concept validated in a laboratory/pilot environment). The broader goal of this thesis is to advance the PtM plant to higher TRL by implementing data-driven surrogate modeling as a core component of the autonomous control framework.

Objectives

1. 1. Raw data processing such as handle missing values, outliers, and sensor noise from several hundred hours of operational logs.
2. 2. Develop, train and testing machine learning surrogate models on the cleaned operational dataset.
3. 3. Incorporate physical constraints (mass balance, energy balance, reaction kinetics of CO₂-to-methanol synthesis) into the neural network training to ensure physically plausible predictions.
4. 4. Develop a constraint-aware regression framework that ensures surrogate predictions respect physical and operational boundaries.
5. 5. Integrate the trained model into a predictor that outputs optimal operating parameters for maximum methanol yield under varying feed conditions.

Required Expertise

• - Proficiency in Python (NumPy, Pandas, scikit-learn, PyTorch/TensorFlow) or MATLAB.
• - Basic knowledge of data cleaning, preprocessing, and exploratory analysis.
• - Familiarity with regression models, neural networks, and surrogate modelling concepts.
• - Strong interest in novel research at the intersection of process engineering and machine learning.
• - Ability to write a scientific thesis in English.

Nice to Have:

Basic understanding of Physics-Informed Neural Networks (PINNs).

Familiarity with chemical process fundamentals (reaction kinetics, mass/energy balances).