Marvin May, M.Sc.

  • 76131 Karlsruhe
    Kaiserstraße 12

Marvin May, M.Sc.

Area of Research:

  • Machine Learning in Production Planning and Control
  • Control of flexible Production Systems
  • Digital Twin and Matrix Production
  • Big Data Analysis in Production
  • Sim2Real: fitting simulations to real behavior
  • Knowledge Graphs in Production
  • Industry 4.0

General Tasks:

  • Coordination of lecture…
    • Production Operations Management
    • Data Mining in Production (Seminar)
    • Process Mining in Production (Seminar)
  • Learning factory on global production scalable automatization and lean & industry4.0


  • EU Digiman4.0 
  • Innovation Center SAP – Development of an AI-based multi-agent production control for matrix production
  • BaWue Robust

Curriculum Vitae:

since 09/2019

Research Associate at the Institute of Production Science (wbk) at Karlsruhe Institute of Technology (KIT) 


Industrial Engineering and Management student at KIT, graduated with M.Sc. & B.Sc.


Exchange Student at Shanghai Jiaotong University (上海交通大学)


Exchange student & Research/Teaching Associate at the University of Massachusetts, Amherst and Isenberg School of Management


Exchange student at Université de Strasbourg


Exchange student at Beijing Institute of Technology (北京理工大学)




[ 1 ] May, M. C.; Kuhnle, A. & Lanza, G. (2020), "Digitale Produktion und Intelligente Produktionssteuerung", wt Werkstattstechnik online, vol. 4, pp. 255-260. 10.5445/IR/1000119555
Im Rahmen der stufenweisen Umsetzung von Industrie 4.0 erreicht die Vernetzung und Digitalisierung die gesamte Produktion. Den physischen Produktionsprozess nicht nur cyber-physisch zu begleiten, sondern durch eine virtuelle, digitale Kopie zu erfassen und optimieren, stellt ein enormes Potential für die Produktionssystemplanung und -steuerung dar. Zudem ermöglichen digitale Modelle die Anwendung intelligenter Produktionssteuerungsverfahren und stellen damit einen Beitrag zur Verbreitung optimierender adaptiver Systeme dar.

[ 2 ] Kain, M.; Nadimpalli, V.; Miqueo, A.; May, M. C.; Yagüe-Fabra, J.; Häfner, B.; Pedersen, D.; Calaon, M. & Tosello, G. (2020), "Metal Additive Manufacturing of Multi-Material Dental Strut Implants". Proceedings of the 20th International Conference of the European Society for Precision Engineering and Nanotechnology, Elsevier, pp. 175-176.
The demand for high flexibility and mass customization within medical implants has resulted in the early adoption of additive manufacturing technologies in the medical sector. Apart from offering unique customization opportunities, metal additive manufacturing of implants also enables the generation of functionally structured surfaces showing increased biocompatibility. The need for improving properties like biocompatibility is essential since many patients suffer from dental implant failures such as screw loosening, integration failure or implant breakage. The high requirements for implant materials depending on the specific implant environment and function complicates the implant design and results in multi-component implants. This could be overcome by the use of multi-material metal additive manufacturing that enables functionally graded part production. However, current commercial laser powder-bed fusion systems are limited to a single material at a time. By utilizing an open-architecture laser powder-bed fusion system by Aurora Labs it was possible in a controlled way to mix AISI 316L Stainless Steel with AISI 440C Stainless Steel during the fabrication of a dental strut-like geometry thereby functionally grading the properties of the material to have a low hardness near the gums and a higher hardness around the crown. The material distribution was evaluated by cutting and etching and the hardness was measured at various locations. The current work demonstrates the technical viability of manufacturing functionally graded dental strut implants by leveraging multi-material metal additive manufacturing.

[ 3 ] May, M. C.; Overbeck, L.; Wurster, M.; Kuhnle, A. & Lanza, G. (2020), "Foresighted Digital Twin for situational Agent Selection in Production Control". Elsevier, pp. 27-32. 10.1016/j.procir.2021.03.005
As intelligent Data Acquisition and Analysis in Manufacturing nears its apex, a new era of Digital Twins is dawning. Foresighted Digital Twins enable short- to medium-term system behavior predictions to infer optimal production operation strategies. Creating up-to-the-minute Digital Twins requires both the availability of real-time data and its incorporation and serve as a stepping-stone into developing unprecedented forms of production control. Consequently, we regard a new concept of Digital Twins that includes foresight, thereby enabling situational selection of production control agents. One critical element for adequate system predictions is human behavior as it is neither rule-based nor deterministic, which we therefore model applying Reinforcement Learning. Owing to these ever-changing circumstances, rigid operation strategies crucially restrain reactions, as opposed to circumstantial control strategies that hence can outperform traditional approaches. Building on enhanced foresights we show the superiority of this approach and present strategies for improved situational agent selection.

[ 4 ] Kapp, V.; May, M. C.; Lanza, G. & Wuest, T. (2020), "Pattern Recognition in Multivariate Time Series: Towards an Automated Event Detection Method for Smart Manufacturing Systems", Journal of Manufacturing and Materials Processing, vol. 3, pp. 28-00. 10.3390/jmmp4030088
This paper presents a framework to utilize multivariate time series data to automatically identify reoccurring events, e.g., resembling failure patterns in real-world manufacturing data by combining selected data mining techniques. The use case revolves around the auxiliary polymer manufacturing process of drying and feeding plastic granulate to extrusion or injection molding machines. The overall framework presented in this paper includes a comparison of two different approaches towards the identification of unique patterns in the real-world industrial data set. The first approach uses a subsequent heuristic segmentation and clustering approach, the second branch features a collaborative method with a built-in time dependency structure at its core (TICC). Both alternatives have been facilitated by a standard principle component analysis PCA (feature fusion) and a hyperparameter optimization (TPE) approach. The performance of the corresponding approaches was evaluated through established and commonly accepted metrics in the field of (unsupervised) machine learning. The results suggest the existence of several common failure sources (patterns) for the machine. Insights such as these automatically detected events can be harnessed to develop an advanced monitoring method to predict upcoming failures, ultimately reducing unplanned machine downtime in the future.

[ 5 ] May, M. C.; Kiefer, L.; Kuhnle, A.; Stricker, N. & Lanza, G. (2020), "Decentralized Multi-Agent Production Control through Economic Model Bidding for Matrix Production Systems". Elsevier, pp. 3. 10.1016/j.procir.2021.01.043
Due to increasing demand for unique products, large variety in product portfolios and the associated rise in individualization, the efficient use of resources in traditional line production dwindles. One answer to these new challenges is the application of matrix-shaped layouts with multiple production cells, called Matrix Production Systems. The cycle time independence and redundancy of production cell capabilities within a Matrix Production System enable individual production paths per job for Flexible Mass Customisation. However, the increased degrees of freedom strengthen the need for reliable production control systems compared to traditional production systems such as line production. Beyond reliability a need for intelligent production within a smart factory in order to ensure goal-oriented production control under ever-changing manufacturing conditions can be ascertained. Learning-based methods can leverage condition-based reactions for goal-oriented production control. While centralized control performs well in single-objective situations, it is hard to achieve contradictory targets for individual products or resources. Hence, in order to master these challenges, a production control concept based on a decentralized multi-agent bidding system is presented. In this price-based model, individual production agents - jobs, production cells and transport system - interact based on an economic model and attempt to maximize monetary revenues. Evaluating the application of learning and priority-based control policies shows that decentralized multi-agent production control can outperform traditional approaches for certain control objectives. The introduction of decentralized multi-agent reinforcement learning systems is a starting point for further research in this area of intelligent production control within smart manufacturing.

[ 6 ] May, M. C.; Schmidt, S.; Kuhnle, A.; Stricker, N. & Lanza, G. (2020), "Product Generation Module: Automated Production Planning for optimized workload and increased efficiency in Matrix Production Systems". Elsevier, pp. 45-50. 10.1016/j.procir.2021.01.050
Ever increasing demand for individualized and customized products induce the need for high variability in production and manufacturing through Mass Customisation. Mass Customisation requires more flexibility and adaptability capabilities in production systems. Matrix Production is a tact free job-shop like production system enabling variable production routing through a matrix shaped layout of partially redundant machines. Hence, it is one way to increase a production system's flexibility and adaptability. A more powerful production control system comes hand in hand with the evolution towards a tact free Matrix Production System. However, the additional degree of freedom due to the flexibility not only touches production control, but also production planning, thus enabling the production of portfolio external products. Implementation of a Product Generation Module optimizes the workload of Matrix Production Systems to increase their efficiency by assessing the suitability of co-production of portfolio external products. Generation of suitable production orders increase machine utilization without impeding the original multi-dimensional production goals. Thus, reaching new production strategies that include the creation of value through effective manipulation of minor products and byproducts. The flexibility of Matrix Production Systems acts as the Product Generation Modules enabler, insofar as flexibility is the ability of a system to perform within an acceptable production corridor without layout and planning adjustments. This can be enhanced by making use of the Matrix Production Systems adaptability to increase the set of portfolio external products through layout and planning adjustments. Hence, this strategy leads to a continuous automated planning phase and additional revenue due to the additional manufacture of minor products within a Matrix Production System. Doing so allows the Matrix Production System to react towards external demand related and internal capacity related events without sacrificing precious value creation opportunities. Ş 2020 Elsevier B.V.. All rights reserved.

[ 7 ] May, M.; Albers, A.; Fischer, M. D.; Mayerhofer, F.; Schäfer, L. & Lanza, G. (2021), "Queue Length Forecasting in Complex Manufacturing Job Shops", forecasting, no. 3, pp. 322-338. 10.3390/forecast3020021
Currently, manufacturing is characterized by increasing complexity both on the technical and organizational levels. Thus, more complex and intelligent production control methods are developed in order to remain competitive and achieve operational excellence. Operations management described early on the influence among target metrics, such as queuing times, queue length, and production speed. However, accurate predictions of queue lengths have long been overlooked as a means to better understanding manufacturing systems. In order to provide queue length forecasts, this paper introduced a methodology to identify queue lengths in retrospect based on transitional data, as well as a comparison of easy-to-deploy machine learning-based queue forecasting models. Forecasting, based on static data sets, as well as time series models can be shown to be successfully applied in an exemplary semiconductor case study. The main findings concluded that accurate queue length prediction, even with minimal available data, is feasible by applying a variety of techniques, which can enable further research and predictions.

[ 8 ] Kuhnle, A.; May, M.; Schäfer, L. & Lanza, G. (2021), "Explainable reinforcement learning in production control of job shop manufacturing system", International Journal of Production Research, 10.1080/00207543.2021.1972179
Manufacturing in the age of Industry 4.0 can be characterised by a high product variety and complex material flows. The increasing individualisation of products requires adaptive production planning and control systems. Research in the area of Machine Learning demonstrates the applicability and potential of Reinforcement Learning (RL) systems for the control of complex manufacturing. However, a major disadvantage of RL-methods is that they are usually considered as "black box" models. For this reason, this paper investigates methods of explainable reinforcement learning in production control. Based on a comprehensive literature review an approach to increase the plausibility of RL-based control strategies is presented. The approach combines the advantages of high prediction accuracy (e.g. neural networks) and high explainability (e.g. decision trees). In doing so, understandable control strategies such as heuristics can be generated, and an advanced RL-system can be designed including specific domain expertise. The results are demonstrated based on a real-world system, taken from semiconductor manufacturing, which is investigated in a simulated approach.

[ 9 ] Overbeck, L.; Hugues, A.; May, M. C.; Kuhnle, A. & Lanza, G. (2021), "Reinforcement Learning Based Production Control of Semi-automated Manufacturing Systems". Procedia CIRP, Elsevier, pp. 170-175.
In an environment which is marked by an increasing speed of changes, industrial companies have to be able to quickly adapt to new market demands and innovative technologies. This leads to a need for continuous adaption of existing production systems and the optimization of their production control. To tackle this problem digitalization of production systems has become essential for new and existing systems. Digital twins based on simulations of real production systems allow the simplification of analysis processes and, thus, a better understanding of the systems, which leads to broad optimization possibilities. In parallel, machine learning methods can be integrated to process the numerical data and discover new production control strategies. In this work, these two methods are combined to derive a production control logic in a semi-automated production system based on the chaku-chaku principle. A reinforcement learning method is integrated into the digital twin to autonomously learn a superior production control logic for the distribution of tasks between the different workers on a production line. By analyzing the influence of different reward shaping and hyper-parameter optimization on the quality and stability of the results obtained, the use of a well-configured policy-based algorithm enables an efficient management of the workers and the deduction of an optimal production control logic for the production system. The algorithm manages to define a control logic that leads to an increase in productivity while having a stable task assignment so that a transfer to daily business is possible. The approach is validated in the digital twin of a real assembly line of an automotive supplier. The results obtained suggest a new approach to optimizing production control in production lines. Production control shall be centered directly on the workers' routines and controlled by artificial intelligence infused with a global overview of the entire production system.