PhD Thesis: Model Hybridization in the Context of Digital Twins

Context

Complex software-intensive systems are evolving at an accelerating pace, operating in increasingly dynamic environments and contending with ever-increasing uncertainty. This requires a high level of adaptability, through a continuous engineering of complex cyber-physical, socio-technical, ecosystems. Digital twins are key enablers, and leverage on both model simulation and data science. Modeling & Simulation is a time-honored activity consisting in building complex analytical models to be simulated to evaluate natural or engineered phenomena. Conversely, data science relies on the availability of data to build complex predictive AI-based learning models. While both could be confused or even opposed, we argue they better complement each other to enhance the ability to best engineer complex systems continuously.

The sound hybridization of model simulation and data science enables a coordinated use of both techniques in complex scenarios (e.g., analytical models for explanation, and data model for recurrent pattern retrieval). Moreover, the hybridization also opens the door to adaptive modeling, where one model is inferred or refined by the others, and vice-versa (e.g., inferring or refining an analytical model from a learning model, and better tuning and explaining a learning model thanks to an analytical model).

Challenges are related to the identification of well-defined interfaces for each model, and the required protocols and operators to support the proposed scenarios. We aim to establish the first unifying theory for both model simulation and learning models, and demonstrate its applicability in practice through concrete tool support.

Objectives

Unifying theory for inductive and deductive reasoning

• Hybrid modeling: coordinated use of heterogeneous predictive models. This objective focuses on the definition of well-defined concepts to specify complex hybrid modeling scenarios through the coordinated use of different techniques involved in digital twins, e.g., Modeling & Simulation, Machine Learning, Data Mining, etc. These concepts will provide the semantic foundations to enact hybrid models in digital twin services such as recommenders, linters and decision-making tools.

• Adaptive modeling: model adaptation (inference/refinement/configuration). This objective focuses on the definition of well-defined concepts to specify complex adaptive modeling scenarios through a retro-action in between the different models involved in the different techniques (e.g., Modeling & Simulation, Machine Learning, Data Mining, etc.) These concepts will provide the semantic foundations to enact adaptive modeling scenarios in digital twin services such as modeling environment, and decision-making tools.

• Model interfaces and protocols. This objective aims at formalizing the required model interfaces and protocols to leverage on the two aforementioned objectives. The outcome if a unifying predictive platform, supporting both the orchestration of service requests on the different available predictive models, but also possibly the adaptation of them from others.

Application domain

Application to sustainable manufacturing

• Frugal digital twins (i.e., data and computations). The aim here is to create more frugal digital twins by reducing their footprint, both in terms of the models used and the potential computing costs involved. Properly dimensioning the data collected in terms of volume, frequency and type is a first way. Hybridization would be a way of reducing the cost of using models in the twin through various levers: model accuracy, type of model to be used (analytical, learning or other), reduction in the cost of training ML models by cross-fertilization with analytical models, etc.
• Digital twins for sustainable systems (i.e., decision making). Digital twins make it possible to assess the impact of the system itself, but also to design and evaluate adaptation (through feedback) to external phenomena or disturbances. The impact of the system may, for example, concern energy consumption or other ecological indicators linked to the system and processes with material or immaterial output. The creation of specific semantics enabling the creation of adaptive models is a response to these challenges relating to ecological sustainability and digital twins. In addition, model hybridization enables the capture or refinement of energy consumption or other ecological models through data. This enables us to produce decision-support tools for better design, adaptation and evolution.

Environment

The candidate will be involved in the DiverSE team, joint to the CNRS (IRISA) and Inria. The DiverSE team is located in Rennes, France. DiverSE’s research is in the area of software engineering. The team is actively involved in European, French and industrial projects and is composed of 13 faculty members, 20 PhD students, 4 post-docs and 3 engineers. The main advisors of the PhD thesis will be Prof. Benoit Combemale (University of Rennes 1, DiverSE team), Prof. Mathieu Acher (INSA Rennes, DiverSE team), and Dr. Quentin Perez (INSA Rennes, DiverSE team). The candidate will register to the doctoral school in computer science of the University of Rennes.

The PhD research project is funded by the Inria Foundation and will be conducted in collaboration with La Poste.

References

• R. Eramo, F. Bordeleau, B. Combemale, M. v. d. Brand, M. Wimmer and A. Wortmann, “Conceptualizing Digital Twins,” in IEEE Software, vol. 39, no. 2, pp. 39-46, March-April 2022,
• R. Verdecchia, L. Cruz, J. Sallou, M. Lin, J. Wickenden and E. Hotellier, “Data-Centric Green AI An Exploratory Empirical Study,” 2022 International Conference on ICT for Sustainability (ICT4S), Plovdiv, Bulgaria, 2022, pp. 35-45
• Narciso, Diogo AC, and F. G. Martins. “Application of machine learning tools for energy efficiency in industry: A review.” Energy Reports 6 (2020): 1181-1199.
• Ahmad, Muhammad Waseem, Monjur Mourshed, and Yacine Rezgui. “Trees vs Neurons: Comparison between random forest and ANN for high-resolution prediction of building energy consumption.” Energy and buildings 147 (2017): 77-89.

Prerequisites

• A degree (and strong background) in data science and computer science (esp. software engineering)
• skills on numerical analysis, scientific computing and simulation
• interests in programming and modeling languages, and supporting envirionments
• interests in machine learning
• professional proficiency in english
• skills for presenting and writting
• autonomly, rigor and hard worker

How to apply

Send your CV, motivation letter, and grades of your bachelor and master with the diplomas to Benoit Combemale, Mathieu Acher and Quentin Perez