University of Thessaly (UTH)
Development of explainable machine learning methods for physical systems and fusion reactors
Abstract
dc:descriptionThe development of explainable models for complex systems in the physical sciences remains a persistent challenge in computational science, particularly when the underlying phenomena span multiple scales of time and length. Although machine learning methods have improved predictive capabilities, black-box models offer limited physical insight into the mechanisms they capture. This thesis addresses this issue by establishing an explainable machine learning framework, centered on symbolic regression, that generates transparent, closed-form expressions directly from data. The investigation follows a deliberate bottom-up trajectory, addressing three problems of increasing physical complexity: the estimation of transport properties in fluids, the characterization of gas-surface interactions via scattering kernels, and the analysis of neutral gas pressure in the Wendelstein 7-X fusion reactor.Following this route, the investigation begins at the atomistic scale, where transport properties such as the shear viscosity, thermal conductivity, and self-diffusion coefficient are correlated with fluid temperature and density. The viscosity and thermal conductivity data originate from simulations covering the complete Lennard-Jones fluid phase space, spanning from dilute gases to dense liquids. The diffusion dataset encompasses both bulk and nanoconfined environments across nine real molecular fluids, incorporating channel height as an additional parameter in the corresponding analytical expressions. Across all investigated cases, the symbolic regression scheme recovers expressions that combine high predictive accuracy with physical interpretability, demonstrating its ability to capture governing physical laws purely from data.Building on this, the methodology is extended to the gas-surface interface, where molecular dynamics simulations of rarefied argon and helium atoms striking solid walls are used to investigate momentum and energy transfer across a temperature range of 300 to 500 K, under both isothermal and temperature-gradient conditions. For the isothermal argon-gold system, a symbolic regression workflow is introduced to bridge the gap between conventional parametric kernels and non-parametric, data-driven ones. Collision histograms are used to construct a three-dimensional probability surface, which is then approximated via symbolic regression. The resulting symbolic joint probability density function can be further decomposed to sample impinging and reflecting velocity components, by deriving the marginal and conditional distributions through simple analytical or numerical integration. This velocity sampling scheme can serve as a boundary condition for higher-scale kinetic codes.Finally, at the experimental scale, the machine learning scheme is employed to investigate the neutral gas pressure of the Wendelstein 7-X stellarator. In magnetically confined fusion reactors, perfect confinement cannot be guaranteed, and as a result, imperfect confinement leads to unavoidable plasma-surface interactions at the vessel walls. Among other physical processes, incident ions become neutralized and can move freely across magnetic field lines, affecting plasma purity and disrupting particle exhaust efficiency. Using operational data from the OP1.2b experimental campaign for the standard, high-iota, and high-mirror magnetic configurations at the AEH, AEP, and AEI ports, sub-divertor neutral gas pressure is correlated with plasma parameters such as the line-integrated electron density, heating power, radiated power, toroidal plasma current, coil control current, and core and edge ion temperatures. Statistical and feature selection analyses shape the resulting derived expressions, which ultimately achieve strong predictive accuracy and physical correspondence using only two or three variables, depending on the magnetic configuration. An extra-trees regression model was used as an accuracy benchmark and scored higher metrics than the symbolic expressions; however, its black-box nature limits reliable extrapolation beyond the training conditions, unlike the simple, transparent symbolic expressions. Taken together, this work shows that explainable machine learning offers a physically grounded path toward modeling particle exhaust in next-generation fusion devices.
Degree
thesis:*- Grantor dc:publisher
- University of Thessaly (UTH)
- Year dc:date
- 2026
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Angelis, Dimitrios
- Αγγελής, Δημήτριος
Subjects
dc:subject × 24- Επεξηγήσιμη τεχνητή νοημοσύνη
- Μηχανική μάθηση
- Συμβολική παλινδρόμηση
- Μοριακή δυναμική
- Θερμοπυρηνική σύντηξη
- Ιξώδες
- Θερμική αγωγιμότητα
- Συντελεστής διάχυσης
- Δυναμική αραιοποιημένων αερίων
- Explainable artificial intelligence
- Machine learning
- Symbolic regression
- Molecular dynamics
- Thermonuclear fusion
- Viscosity
- Thermal conductivity
- Diffusion coefficients
- Rarefied gas dynamics
- Φυσική
- Φυσικές Επιστήμες
- Φυσική ρευστών και πλάσματος
- Physical Sciences
- Natural Sciences
- Physics, Fluids and Plasmas
Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/62531