|Tipo di tesi||Tesi di laurea magistrale|
|Titolo||Tecniche di Machine Learning applicate alla risonanza elettronica di spin|
|Titolo in inglese||Machine Learning-assisted on-chip electron spin resonance spectroscopy|
|Struttura||Dipartimento di Scienze Fisiche, Informatiche e Matematiche|
|Corso di studi||PHYSICS – FISICA (D.M.270/04)|
|Data inizio appello||2022-03-25|
|Disponibilità||Accessibile via web (tutti i file della tesi sono accessibili)|
Recenti studi hanno descritto l’efficacia degli ensemble di spin molecolari nel contesto delle memorie quantistiche. Il memory time di questi dispositivi è ottimizzato attraverso sequenze di impulsi alle microonde, tra le quali la sequenza di eco di Hahn.
Recent studies have shown the effectiveness of molecular spin ensembles in the context of quantum memories. The memory time of these devices is characterized through microwave pulse sequences, such as the Hahn’s echo sequence. Hahn's protocol is used for stimulating the spin ensemble through two microwave pulses (π/2 and π) of different duration. At the end of the Hahn sequence, a signal called echo is emitted, as a result of the refocusing of the spins. Getting a good echo signal requires optimizing the duration of both pulses. In laboratory experiments, the determination of the durations of the π/2 and π pulses takes place through a calibration process aimed at maximizing the intensity of the echo, which is typically carried out in two steps. In the first one, the intensity of the echo is maximized, while keeping the duration of the second pulse fixed and varying only the duration of the first. In the second step, the amplitude of the echo is maximized by setting the duration of the first pulse to the value that produced the maximization in the first step, while the duration of the second pulse is varied. The protocol described above can require a considerable amount of time and the calibration process must be repeated every time the conditions in which the experiment is carried out are changed. The goal of this thesis project is to find methods aimed at reducing the calibration times of the microwave pulses used in electron spin resonance experiments and to minimize operator intervention as much as possible. In particular, the focus is on the implementation of machine learning algorithms to this aim. To characterize the echo signal, the parameters considered are the maximum amplitude of the signal and its symmetry, calculated from the difference between the right and left integrals of the signal. These parameters were extracted from raw experimental measurements. The first approach considered is the clustering process using the K-means algorithm. This is a machine learning algorithm that uses the calculation of distances between points distributed in a space of parameters to divide the data into the number of classes desired by the user (here, one for the noise and one for the echo signals). By mapping the behavior of the algorithm, it was possible to study the border separating noise from echo signal. More refined classification results were obtained by exploiting a neural network approach. The network was first trained using experimental data and then validated using a new data set. With this second approach, the boundary between the classes turned out to have a more complex and reliable shape, allowing a better resolution of the boundary between noise and echo. Due to the promising results, the neural network approach was selected for further tests, in particular for the recognition of the output signal produced by "storage and retrieval" protocol, the one which allows the use of spin as memory for information. It has been shown that the method is able to recognize the input binary sequence sent to the spins, by the only analysis of the output signal.