Description of the thesis topic

The proposed thesis is part of the ANR collaborative project ATHENA (Intelligent Automation of THErmodynamic Cycles for a New Energy Approach).
General context of ATHENA: Thermodynamic cycles (generator cycles, refrigeration cycles, heat pumps) are essential for the energy transition because they open up prospects for improving energy efficiency and decentralization (adaptation to local energy sources). There is no universally optimal cycle; optimality depends on precise specifications defining the objectives to be achieved and the constraints of the problem.
Problem: Traditional heuristic approaches of experts do not allow the exploration of non-intuitive process configurations, which are often more efficient than conventional structures. Superstructure optimization, which uses a predefined architecture with numerous unit operations and possible paths, offers an alternative for identifying non-intuitive process structures. However, this method is limited by the inductive bias of the predefined superstructure. Innovation: Increased computing power and advances in data science have popularized new algorithms for automatic process generation. Generative approaches, using deep learning algorithms, can generate new process structures, surpassing conventional optimization techniques.
Objectives of the ATHENA project: To develop powerful tools for the synthesis of innovative thermodynamic cycles. The project combines a methodological study and a practical application to harness the potential of deep learning in process synthesis. In particular, we aim to develop process knowledge bases to support generative methods and, secondly, to develop a tool for determining the optimal structure of any type of thermodynamic cycle by optimizing a superstructure automatically generated using a deep learning approach. The project brings together the LRGP (Reaction and Engineering Laboratory, CNRS-Université de Lorraine), the LPSM (Probability, Statistics and Modeling Laboratory, CNRS-Université de Lorraine), EDF (Electricité de France), and Fives-Prosim.
This doctoral program focuses on generative models for energy cycles. Its main objective is to develop systematic tools capable of proposing process architectures relevant to given objectives (e.g., energy, environmental, economic) by using generative models to explore the search space in a broad and impartial manner.

The work can draw on algorithms from the recent work of the process design management team. Several generative algorithms have been produced: using evolutionary approaches that build processes by mutation operators [1], natural language processing techniques with recurrent short-term memory (LSTM) neural networks [2], and variational autoencoders (VAE) for representation learning [3]. However, these approaches remain limited:
- Restricted application domains (reaction-separation-recycling problem [1]), membrane separation cascades [4], and supercritical CO2 cycles [2, 4, 5];
- Lack of common representation for process flowsheets (graph incidence matrix, custom-made dedicated language, SFILES 2.0 standard)
- Various custom-made process simulation environments, without the use of simulation software familiar to process engineers. In this thesis, we aim to:
- Propose generative models for other types of cycles, based on existing models. To do this, we could use transfer learning to transpose the recurrent neural network (RNN) model available for supercritical CO2 power cycles to other cycles. Since thermodynamic conditions vary greatly depending on the fluid and temperature domain, resulting in a diverse cycle architecture (Brayton, Rankine, etc.), transfer learning may not be sufficient, and model retraining will be considered.
- Explore process generation with various objectives (model tuning), such as energy indicators (e.g., efficiency, exergy analysis) or parameters affecting investment costs (e.g., heat exchanger pinch, fluid flow rates), in order to build a rich and diverse process database. The idea is to identify as many potential solutions as possible, which will feed into a superstructure optimization (in another work package) that will then identify the best Pareto solutions in the given context (constraints, objectives).
- Use graph generative model development to compare other learning model architectures and select the one that offers the best convergence in the targeted generation. Other graph generative models from fields other than processes (e.g., molecular design) can be tested, provided that a simulator capable of calculating aptitudes is available. In particular, VAE allows a representation similar to molecule search. At this stage, there is no exploration in the latent space, and we propose to perform exploration/optimization in the latent space to produce new processes with the targeted properties. - It is proposed to extend this work to compare several approaches (generative and conventional) on identical synthesis problems, in terms of optimality, cost, and reproducibility, paying particular attention to the appropriation of methods, tools, and results by process engineers.
This research will be conducted in collaboration with others working within the ATHENA project. Generative models will be guided by a fit assessment using communication developed by other partners between the generation models (usually in the form of a graph or a string) and the process simulator, and accelerated by the use of metamodels. The search space could also be statistically preconditioned using advanced Monte Carlo techniques developed by other partners to identify the boundaries of the feasibility space and target the exploration effort. It is also planned to generate knowledge throughout the exploration of process flowsheets (good or bad), which will be published openly in the form of process flowsheet databases.

Skills: Machine Learning/Deep Learning skills are essential, as well as programming proficiency, as well as some knowledge of energy or process engineering.

Application Documents: In addition to a CV and a personal cover letter, transcripts from the last few years of study are required. Letters of recommendation may be provided.

References :

[1] T. Neveux. “Ab-initio process synthesis using evolutionary programming”. In: Chemical Engineering Science 185 (2018), pp. 209–221. doi: 10.1016/j.ces.2018.04.015.
[2] T. Nabil, M. Noaman, and T. Morosuk. “Data-Driven Structural Synthesis of Supercritical CO2 Power Cycles”. In: Frontiers in Chemical Engineering 5 (2023), p. 1144115. doi: 10.3389/fceng.2023.1144115.
[3] A. Rocha Azevedo, J.- M. Commenge, V. Loubière, T. Neveux, R. Privat, and T. Nabil. “Representation Learning for Flowsheets: Generating Structures for Process Synthesis”. In: ESCAPE34- PSE24 Books of Abstracts. 2024, p. 01. doi: 10.3303/BOA2401.
[4] T. Neveux, B. Addis, C. Castel, V. Piccialli, and E. Favre. “A Comparison of Process Synthesis Approaches for Multistage Separation Processes by Gas Permeation”. In: Computer Aided Chemical Engineering. Vol. 51. Elsevier, 2022, pp. 685–690. doi: 10.1016/B978-0-323-95879-0.50115-6.
[5] A. Rocha Azevedo, T. Nabil, V. Loubière, R. Privat, T. Neveux, and J.- M. Commenge. “ Comparing generative process synthesis approaches with superstructure optimization for the conception of supercritical CO2 Brayton cycles”. In: Computers & Chemical Engineering 201 (2025), p. 109255. doi: 10.1016/j.compchemeng.2025.109255.

Work Context

The doctoral program will be conducted in Nancy at the LRGP (Reaction and Process Engineering Laboratory), a joint unit of the University of Lorraine and the CNRS, located on the ENSIC campus in Nancy (54). The successful candidate will join the joint EDF-LRGP research team 'MELUSINE', supervised by LRGP and EDF staff (thesis supervisor: Jean-Marc COMMENGE (LRGP); co-supervisors and supervisors: Romain PRIVAT (LRGP), Thibaut NEVEUX (EDF), and Tahar NABIL (EDF)). The work will be carried out in close collaboration with the ATHENA project partners.
The position is located in a sector covered by the Protection of Scientific and Technical Potential (PPST), and therefore, in accordance with regulations, requires that your arrival be authorized by the competent MESR authority.

The position is located in a sector under the protection of scientific and technical potential (PPST), and therefore requires, in accordance with the regulations, that your arrival is authorized by the competent authority of the MESR.

Constraints and risks

The work proposed will be exclusively digital and does not present any risks other than those associated with working on a screen.