2026-09-18T22:44:21Zhttps://iris.unitn.it/oai/request

oai:iris.unitn.it:11572/4839102026-05-14T07:50:45Zcom_11572_237821com_11572_101871col_11572_237822ou_ou00011

A scalable machine learning approach to thermal and non-thermal order-disorder phase transitions with ab initio accuracy Corradini, Andrea Corradini, Andrea Calandra Buonaura, Matteo Marini, Giovanni Machine learning interatomic potentials, liquid-liquid phase transitions, non-thermal melting, molecular dynamics, ultrafast photoexcitation, silicon, tellurium The study of out-of-equilibrium systems offers a gateway to transformative technological appli- cations and emerging physical phenomena that are inaccessible via standard adiabatic pathways. However, modeling these states is formidably challenging, as it requires describing non-trivial physical processes across vast temporal and spatial scales. This thesis addresses the fundamen- tal accuracy versus efficiency trade-off inherent in the atomistic modeling of these phenomena by developing and deploying rigorous methodological frameworks based on high-fidelity machine learning interatomic potentials. These tools are utilized to investigate three distinct out-of- equilibrium regimes: • Ultrafast non-thermal melting in silicon: a novel framework based on constrained density functional perturbation theory and machine learning interatomic potentials is developed to accurately model the effects of laser-induced photoexcitation and investigate the role of phonon softenings in the non-thermal transition. • Structural and thermodynamic anomalies in undercooled liquid tellurium: a general-purpose machine learning interatomic potential is optimized and deployed to probe the complex chemistry of liquid tellurium, identifying numerous structural and thermodynamic anoma- lies and exploring the potential existence of a liquid-liquid phase transition analogous to that claimed for water; • Vibrational physics of confined carbyne: an accurate machine learning interatomic po- tential is developed for confined carbyne and employed to reproduce its resonant Raman spectra, accounting for high-order phonon-phonon scattering processes via the stochastic self-consistent harmonic approximation. Collectively, this research demonstrates that properly trained machine learning interatomic po- tentials can effectively bridge the accuracy versus efficiency tradeoff and show enhanced predictive capabilities when compared with experimental observations. By enabling the simulation of com- plex metastable and photoexcited states with quantum-chemical accuracy, this thesis provides a robust protocol for exploring the complex and fascinating physics of out-of-equilibrium systems. 2026-04-20 info:eu-repo/semantics/doctoralThesis https://hdl.handle.net/11572/483910 eng firstpage:1 lastpage:224 numberofpages:224 info:eu-repo/semantics/openAccess Università degli studi di Trento place:TRENTO license:Creative commons license uri:http://creativecommons.org/licenses/by/4.0/