GeoTS: A time series classification framework to identify geological formations
Shwetha Salimath, Centrale-Supelec
Lundi 23 juin à 11h30, CEA-Saclay, Orme-des merisiers, Bat 709, salle Rubin
Studying the lithography of the Earth's subsurface in geoscience involves analyzing different geological formations to model and characterize reservoirs. This process uses drilled well measurements to connect specific geological formations or tops. Reservoir modeling is essential in geothermal, mineral mining, oil and gas, and carbon storage. Traditional well correlation algorithms are time-consuming and costly, but Deep Learning (DL) models have shown promising results. This paper presents GeoTS, a Python library that employs advanced time series classification DL models for well correlation. It uses drilling trajectory depth and gamma-ray well logs as inputs, predicting the depths of formations' tops. Gamma-ray signatures around these depths are extracted, cleaned, and clustered using Dynamic Time Warping (DTW) and machine learning models like HDBSCAN and OPTICS. The implementation includes various deep learning architectures (FCN, InceptionTime, XceptionTime, XCM, LSTM-FCN) and new models (LSTM-2dCNN, LSTM-XCM). The results indicate faster computation and higher accuracy than industry benchmarks, making this the first open-source benchmark for the well correlation task, to our knowledge.
Towards a better understanding of the genetics of the human cortical folding
Antoine Dufournet, CEA, Neurospin
Mardi 8 avril à 11h30, CEA-Saclay, Neurospin, Bat 145, Amphiteatre salle 183
How self-supervised deep learning help us understand cortical folding variability?
Joël Chavas, CEA, Neurospin
Mardi 4 mars à 11h30, CEA-Saclay, Neurospin, Bat 145, Amphiteatre
The human brain is folded, and its folding is highly variable among individuals. A long-standing aim in studying cortical folding is to understand and quantify if cortical folding relates to cognitive or clinically relevant parameters, like early-life factors or diseases. We now have available more than 100.000 brain images: this quantity of data makes the field mature for the application of recent Deep learning algorithms. I will present our approach, based on self-supervised algorithms (more precisely on Barlow Twins algorithms), to build regional latent spaces that we can then analyze. Simple linear algorithms applied to these latent spaces permit to identify folding patterns related to early-life factors, like handedness and birth weight.
Superconducting magnet design through multi-physics optimisation
Damien Minenna, CEA/irfu/DACM
Mardi 11 février à 11h30, CEA-Saclay, Orme-des merisiers, Bat 709, salle Cassini
Designing superconducting magnets presents a challenge due to their multi-physics complexity, diverse analytical tools, and often imprecise specifications. To streamline this process, we introduce ALESIA, a novel optimisation and data management toolbox developed at CEA-IRFU. ALESIA leverages advanced algorithms, including nonlinear programming techniques, evolutionary algorithms, active learning strategies, and surrogate modelling, to accelerate the design process. By intelligently exploring the parameter space, ALESIA enables rapid convergence towards optimal solutions while minimising computational cost. ALESIA’s flexible architecture allows integration with any physics simulation software, encompassing magnetic field calculations (OPERA), and mechanical analysis (CAST3M), but its applicability can be broadening beyond magnet design. Crucially, ALESIA’s automated optimisation loop simultaneously considers all stages – magnetism, conductor properties, mechanics, and quench behaviour – ensuring holistic and robust design solutions.
