Poster De Conférence Année : 2026

Towards frugal Bayesian optimization for environmental impact minimization of computer experiments

Résumé

Expensive-to-evaluate blackbox simulations play a key role for many engineering and industrial applications. In this context, surrogate models have been widely used to address a large range of applications, e.g., aircraft design, deep neural networks, coastal flooding prediction, agriculture forecasting, or seismic imaging. Most blackbox simulations are complex and computationally expensive. Typically, multidisciplinary design of aircraft leads to calling solvers, such as computational fluid dynamics or finite element, that can take days to be executed across clusters. As a result, there has been a growing interest in more efficient surrogate models, particularly in the context of Bayesian optimization based on Gaussian processes. This work follows the PhD thesis of Paul Saves (2020-2023) in which high-dimensional multidisciplinary design optimization methods were developed for aircraft eco-design. Contributions on dimension reduction and mixed-variable techniques for Gaussian processes were added to the open source software SMT (https://smt.readthedocs.io/en/latest/) and applied on both large sets of industrial or academic tests. While mixed-categorical and hierarchical variables enable more complex modeling, they also significantly increase computational overhead. To mitigate the resulting increase in execution time, parallel computing has emerged to be the most effective solution. Consequently, high performance computing is now of main interest, even more in the machine learning field. Largescale tasks can require energy-intensive exascale systems infrastructure, such as the european supercomputer Jupiter, whereas smaller-scale operations may be more efficiently executed on regional clusters or standard workstations. As such, we evaluate an energy consumption minimization problem by distributing workloads across a network of computers and their cores, using numerical simulations with various hyperparameter configurations. This problem can be formulated as a System Architecture Optimization problem subject to an environmental budget threshold. Additionnaly, we plan to develop an acquisition function to be explicitly aware of the remaining budget, in a “non-myopic” look-ahead approach. Existing works around hyperparameter optimization minimizing energy consumption of neural network’s training have already shown that energy can be saved while keeping the same accuracy. The eco-design approach will be further improved by integrating spatial and time dependencies into the network of available computers. The targeted application of this PhD will be the eco-design of High Altitude Long Endurance drone by including the architectural choice of the computational infrastructure in the overall process.

Fichier principal
Vignette du fichier
Poster_MASCOT_NUM_Gaston_PLAT.pdf (3.17 Mo) Télécharger le fichier
Origine Fichiers produits par l'(les) auteur(s)
Licence

Dates et versions

hal-05581361 , version 1 (07-04-2026)

Licence

Identifiants

  • HAL Id : hal-05581361 , version 1

Citer

Gaston Plat, Paul Saves, Nathalie Bartoli, Thierry Lefebvre, Joseph Morlier. Towards frugal Bayesian optimization for environmental impact minimization of computer experiments. CNRS. MASCOT-NUM 2026 Conference, Apr 2026, Rennes, France. Proceedings of the MASCOT-NUM 2026 Conference, 2026, pp.15-16, 2026. ⟨hal-05581361⟩
380 Consultations
31 Téléchargements

Partager

  • More