Electronic Structure and machine learning protocols for pre-screening of near-room temperature spin-crossover materials

Abstract

[eng] Spin-crossover (SCO) is a phenomenon observed in certain transition metal complexes, particularly d4–d7 metals, where a reversible transition between high-spin (HS) and low-spin (LS) electronic states occurs in response to external stimuli like temperature, pressure, or light. This switch alters magnetic, optical, and structural properties, making SCO materials attractive for applications in molecular electronics, data storage, sensors, and smart devices. Among transition metal complexes, FeIII SCO complexes are widely studied because of their distinct electronic configurations and stability. The transition between HS and LS states is governed by the metal-ligand interaction, specifically the ligand field strength. A stronger ligand field stabilizes the LS state, while a weaker one favours the HS state, influencing the system's magnetic properties. These transitions, characterised by changes in magnetic moment and colour, make FeIII complexes a key focus for material design. This thesis investigates the SCO behaviour of FeIII complexes through computational and machine learning (ML) techniques, with a focus on ligand functionalization, benchmarking of density functional theory (DFT) methods, and studying dinuclear and polynuclear systems. The research begins with a systematic benchmark analysis of different DFT functionals to determine the best-suited computational approaches for predicting the spin state energetics and transition temperatures of FeIII complexes. The results show that while certain functionals provide accurate predictions of SCO properties, the accuracy depends heavily on the specific characteristics of the FeIII systems being studied. A key contribution of the research is the exploration of ligand design and its impact on SCO behaviour. By altering ligand substituents, the electronic environment around the metal ion can be fine-tuned, providing control over the transition temperature (T(1/2) and other SCO properties. For instance, electron-donating groups on the ligand tend to lower T(1/2) , while electron-withdrawing groups increase it. These ligand-induced modifications are particularly important in FeIII complexes, as both electronic and steric factors play critical roles in governing the spin state transition. The study demonstrates how strategic ligand design can be used to tailor SCO properties for specific applications. In addition to mononuclear FeIII complexes, the thesis examines dinuclear systems, where the presence of two metal centres introduces additional complexity. In these systems, the interaction between metal centres results in cooperative SCO behaviour, such as two-step transitions or the stabilization of intermediate spin states. The research highlights the need for more sophisticated computational models to accurately capture these complex behaviours in dinuclear and polynuclear systems. The findings contribute to the growing understanding of how intermetallic interactions can be leveraged to design SCO materials with specific magnetic properties, which are critical for potential applications in sensors and molecular electronics. Machine learning (ML) models, particularly Kernel Ridge Regression (KRR) and Gaussian Processes (GP), are introduced as complementary tools to traditional computational methods. These ML models are trained on datasets generated from DFT calculations and are used to predict SCO properties such as transition temperatures and spin state energetics. The ML models offer a scalable and efficient approach to studying larger and more complex systems, significantly reducing computational costs while maintaining high accuracy. Feature importance analysis reveals key molecular descriptors that drive SCO behaviour, providing valuable insights into which molecular modifications are likely to result in desirable SCO properties. This approach accelerates the discovery and design of new SCO materials. The integration of machine learning, ligand design, and advanced computational methods in this thesis presents a comprehensive framework for understanding and predicting SCO behaviour in FeIII complexes. The combination of these approaches enables the development of customizable materials for a range of technological applications, including molecular switches, sensors, and memory devices.

[cat] El fenomen de spin-crossover (SCO) es produeix en certs complexos de metalls de transició, particularment en els metalls d4–d7, on hi ha una transició reversible entre els estats de spin alt (HS) i spin baix (LS) en resposta a estímuls externs com la temperatura, la pressió o la llum. Aquest canvi afecta les propietats magnètiques, òptiques i estructurals, fent que els materials SCO siguin útils per a aplicacions en electrònica molecular, emmagatzematge de dades, sensors i dispositius intel·ligents. Entre els complexos de metalls de transició, els complexos de FeIII amb SCO són particularment estudiats per la seva estabilitat i configuracions electròniques úniques. Aquesta tesi investiga el comportament SCO dels complexos de FeIII mitjançant tècniques computacionals i de machine learning (ML). Es centra en la funcionalització de lligands, l'avaluació de mètodes de la teoria del funcional de densitat (DFT) i l'estudi de sistemes dinuclears i polinuclears. Els resultats mostren que alguns funcionals de DFT poden predir amb precisió les propietats de SCO, encara que la seva precisió depèn de les característiques específiques dels sistemes de FeIII. També s'explora com el disseny de lligands influeix en el comportament SCO. En modificar els lligands, es pot ajustar la temperatura de transició (T1/2) i altres propietats. Els grups donadors d'electrons tendeixen a disminuir el T1/2, mentre que els grups acceptors el fan augmentar, la qual cosa és clau per al disseny de materials personalitzats. A més, s'estudien sistemes dinuclears, on la interacció entre centres metàl·lics provoca comportaments més complexos, com transicions en dues etapes. Finalment, s'utilitzen models de ML per predir les propietats SCO de manera més eficient, reduint notablement els costos computacionals.