Skip to Content

Modelling of Manufacturing Processes via Machine Learning Algorithms

This dissertation explores the use of machine learning techniques in manufacturing to enhance efficiency and quality. It investigates the real potential of AI through addressing three challenges: surrogate modelling, AI driven by experimental data, and reinforcement learning. Six AI models developed within this study tackle these issues. DeepForge and CrystalMind, as surrogate models, offer rapid prediction capabilities, notably reducing the time for optimizing forging processes from years to minutes. MeltPoolGAN and AIBead, empirical AI models, leverage real-world data to refine predictions in complex manufacturing tasks, with MeltPoolGAN notably improving melt pool image classification. RLPlanner and RLTube demonstrate the application of reinforcement learning in path planning for Wire Arc Additive Manufacturing, optimizing process parameters efficiently.


Type of Work: PhD Thesis

Main Author: Jan Petrik

Affiliation: ETH Zurich

Supervisors: Markus Bambach

Date: 20th October 2023

Journal: Six Publications in Journals

Online: ResearchCollection

The research took a well-rounded approach, integrating machine learning with manufacturing technologies such as wire arc additive manufacturing, laser powder bed fusion and forging. Using both experimental data and simulations, surrogate models were developed to replace traditional, computationally intensive simulations, facilitating real-time process control and optimisation. Emphasis was placed on experimental validation and the development of robust machine learning algorithms that could accurately predict and improve manufacturing outcomes. Challenges included the complexity of modelling multi-physical interactions and the need for large, high-quality data sets to effectively train the algorithms.
Key findings demonstrate the potential of AI to significantly improve manufacturing efficiency, quality and adaptability. Surrogate models such as DeepForge and CrystalMind proved effective in speeding up optimisation workflows, while empirical AI models provided insights into complex manufacturing processes that were previously difficult to simulate accurately. Reinforcement learning applications in path planning further highlighted AI's ability to improve operational decision making. These contributions not only address the pressing need for more efficient and adaptive manufacturing solutions, but also open up new avenues for future research, particularly in hybrid model development and real-time control.


  • Image 1
    AI-based algorithms have been developed for three manufacturing technologies: Wire Arc Additive Manufacturing, Laser Powder Bed Fusion, and both Closed and Open Die Forging.
  • Image 2
    Closed and open loop systems tailored for manufacturing processes, incorporating Artificial Intelligence (AI) models. These AI models are developed based on simulation or experimental data from the manufacturing systems. Furthermore, these AI models can be used within an optimization algorithm to improve manufacturing processes.
  • Image 3
    Summary of six AI-based models deployed within the dissertation. Their more detailed description can be found in my Portfolio.