Application of machine learning and artificial inteligence in plasma spectroscopy

[I. Traparić ]

95 lista

Fizika - Fizika jonizovanog gasa i plazme / Physics - Physics of Ionised Gases and Plasma

U ovoj doktorskoj disertaciji ispitana je mogućnost primene mašinskog učenja i veštačke inteligencije za rešavanje nekih problema u oblasti optičke i ekstremne ultraljučičaste emisione spektroskopije plazme. Model slučajne šume, obučen je za modelovanje Stark-ove poluširine emitovane spektralne linije jer se za tu namenu pokazao kao najbolji od svih razmatranih modela. Napravljeni model uspešno je predvideo kvantnu prirodu atomske emisije i regularnosti duž izoelektronskog niza i spektralnih serija neutralnog atoma litijuma. Pokazano je da se ovaj model može koristiti za predviđanje novih Stark-ovih poluširina sa tačnošću od oko 20 %. Zatim, pokazano je da se u slučaju kvantitativne analize korištenjem spektroskopije laserski indukovanog proboja za obučavanje modela mašinskog učenja mogu iskoristiti baze koje nisu snimljene na vlastitoj aparaturi ponavljanjem istih eksperimentalnih uslova. Ovo je značajan rezultat jer dosta laboratorija nema sredstava da priušti dovoljan broj standardnih uzoraka za potrebe ovakvih analiza, i pruža mogućnost standardizacije snimanja ovih baza. Konačno, primenom varijacionog autoenkodera modelovan je UTA (eng. Unresolved Transition Array) struktura emitovanog spektra volframa u ekstremnoj ultraljubičastoj oblasti spektra za tipične uslove u jezgru plazme heliotrona LHD u Japanu. Model je pokazao zadovoljavajuće poklapanje sa eksperimentalnim merenjima, i tačno predvideo temperaturne zavisnosti oblika emitovanog spektra. Model je pokazao i dodatni napredak u modelovanju, omogućava brzo predviđanje oblika emitovanog spektra za zadate uslove plazme u svega nekoliko stotina milisekundi, za razliku od trenutnih kolizono - radijativnih modela kojima za istu svrhu treba po nekoliko nedelja za račun.

In this doctoral thesis the possibility of application of machine learning and artificial intelligence for solving of some particular problems connected with plasma optical and extreme ultraviolet emission spectroscopy was investigated. Random Forest model was trained for prediction of spectral line width due to Stark effect as it showed best performance of all considered models. Created model successfully predicted quantum nature of atomic emission and regularities along isoelectric sequence and spectral series of neutral lithium atom. It was also shown that the model can be used for prediction of new spectral line widths with the uncertainty of around 20 %. Furthermore, it was shown that databases that weren’t recorded inside one’s own laboratory can be used for training of machine learning algorithms for the quantitative LIBS analysis by repeating the same experimental conditions. This is significant result, as many laboratories don’t have the funds for enough standard samples for this kind of analysis, while at the same time this result provides the opportunity for standardisation of database recording process. Finally, using the variational autoencoder the UTA (Unresolved Transition Array) structure of tungsten spectra in the extreme ultraviolet part of the spectrum for typical plasma parameters inside the plasma core of heliotron LHD was modeled. Model showed good agremeent with experimental measurements, while also correctly modeling the temperature dependence of the shape of the emitted spectra. This model provides the opportunity for quick and accurate generation of new spectra for given plasma parameters in several hundreds of miliseconds, contrary to the developed collisional - radiative models for the same purpose, which need few weeks of computational time for same calculations.

mašinsko učenje, veštačka inteligencija, fizika plazme, spektroskopija plazme, fuzija, LIBS

machine learning, artificial inteligence, plasma physics, plasma spectroscopy, fusion, LIBS

533.9:004.8(043.3)

Tekst