A thesis transferring neutrino ´s experience to the industrial sector
Last November 27th, Sebastian Piña Otey has successfully defended his doctoral thesis entitled “Deep learning and Bayesian techniques applied to Big Data in the industry and neutrino oscillations”. Although this doctorate thesis is physics centered, there is also an industrial component developed with AIA within the industrial doctorate program framework sponsored by the Catalonian Government called: “Doctorats Industrials de la Generalitat. Sebastian ´s thesis, codirected by Vicens Gaitan, AIA´s representative, along with Thorsten Lux, neutrino´s team main researcher from the Institut de Física d’Altes Energies (IFAE), has developed a project that connects neutrino oscillations (mentioned in a previous blog post) with cutting-edge Deep Learning technology.
The thesis has focused on a new approach for data analysis and extraction of relevant information for the T2K neutrino experiment (Tokai to Kamioka) in Japan. In particle physics experiments, and more particularly, neutrino experiments, data analysis has three main components: detectors´ direct data processing, theoretical model simulation and model parameters determination. These three key aspects are essential for a rigorous and correct scientifically based results´ delivery. Sebastian ´s research had an impact on all three aspects by introducing different proofs of concepts of new technologies.
Firstly, he has successfully implemented an algorithm of Graph Neural Networks to help detecting not physic ambiguous signs during the simulation of a T2K future detector. As a result, the effect on the neutrino can be rebuilt correctly, increasing the performance as compared to traditional techniques. (The paper has been submitted to the Physical Review D and is currently under evaluation. Pre-publication can be download at https://www.tdx.cat/ )
Secondly, Sebastian has combined the normalizing flows technology, invertible neural networks that allow to evaluate and generate data from a probability distribution based on traditional statistical techniques, rejection sampling under the algorithm of Exhaustive Neural Importance Sampling (ENIS). More precisely, along with his supervisors, they have applied ENIS to generate data from a typical neutrino event in T2K, the interaction called charged-current quasi-elastic (CCQE). This paper was published on July 16th in the Physical Review D
Finally, as mentioned in our previous blog post, the normalizing flows were used to determine the parameters of the two neutrino flavors oscillation model. This paper was also published in the Physical Review D on June 2nd
However, the experience gained by the somewhat exotic application of the neutrino oscillations was also transferred to the industrial field. Sebastian and Vicens worked with Red Eléctrica Española through Elewit and could show the impact and value added of these techniques on potential projects of our clients. Thus, Sebastian has successfully achieved our first industrial doctorate, potentiating AIA´s relation with the IFAE and showing the contribution of the industrial sector to pure research.