Trustworthiness of laser-induced breakdown spectroscopy predictions via simulation-based synthetic data augmentation and multitask learning - CEA - Commissariat à l’énergie atomique et aux énergies alternatives Accéder directement au contenu
Article Dans Une Revue EPJ Web of Conferences Année : 2023

Trustworthiness of laser-induced breakdown spectroscopy predictions via simulation-based synthetic data augmentation and multitask learning

Résumé

Laser-induced breakdown spectroscopy is a versatile technique that can be used to quickly measure the concentration of elements in ambient air. We tackle the issues of performance and trustworthiness of the statistical model used for predictions. We propose a method for improving the performance and trustworthiness of statistical models for LIBS. Our method uses deep convolutional multitask learning architectures to predict the concentration of the analyte and additional information as auxiliary outputs. We also introduce a simulation-based data augmentation process to synthesize more training samples. The secondary predictions from the model are used to characterize, quantify and validate its trustworthiness, taking advantage of the mutual dependencies of the weights of the neural networks. As a consequence, these output can be used to successfully detect anomalies, such as changes in the experimental conditions, and out-of-distribution samples. Results on different types of materials show that the proposed method improves the robustness and trueness of the predictions.
Fichier principal
Vignette du fichier
epjconf_animma2023_01005.pdf (1.49 Mo) Télécharger le fichier
Origine : Fichiers éditeurs autorisés sur une archive ouverte

Dates et versions

cea-04351411 , version 1 (09-04-2024)

Licence

Paternité

Identifiants

Citer

Riccardo Finotello, Daniel L'Hermite, Céline Quéré, Benjamin Rouge, Mohamed Tamaazousti, et al.. Trustworthiness of laser-induced breakdown spectroscopy predictions via simulation-based synthetic data augmentation and multitask learning. EPJ Web of Conferences, 2023, 288, pp.01005. ⟨10.1051/epjconf/202328801005⟩. ⟨cea-04351411⟩
27 Consultations
5 Téléchargements

Altmetric

Partager

Gmail Facebook X LinkedIn More