Molecular photoswitches reversibly interconvert between isomers with photophysical properties upon irradiation, enabling applications in responsive materials, energy storage, photopharmacology, and molecular machines. Their photoswitching performance depends on both the photochemical processes that generate a metastable state and the thermal pathways through which it returns to the stable form. Understanding these thermal pathways is essential for controlling metastable-state lifetimes and identifying competing reactions. However, locating transition states for thermal isomerisation using density functional theory (DFT) is computationally demanding for high-throughput workflows, particularly for flexible molecules with multiple conformers and possible reaction mechanisms. Faster methods with comparative accuracy to DFT would improve the capacity to study novel photoswitch designs and substitution effects. In this project, we aim to accelerate DFT transition-state workflows for novel photoswitch architectures investigated in the group, including phosphaindorubins, diarylethenes, azoarenes, and hemiphosphoindigos using Universal Models for Atoms (UMA). UMA is a pretrained machine-learning interatomic potential that predicts ground-state energies and forces at substantially lower computational cost than DFT, with near-DFT accuracy demonstrated across molecular benchmarks. Its rapid evaluation of molecular structures offers an opportunity to explore reaction pathways and generate promising transition-state geometries before undertaking more expensive electronic-structure calculations. We will use UMA geometry optimisations and conformer searches to identify relevant reactant and product structures, followed by relaxed coordinate scans and reaction-path methods to investigate possible thermal interconversion mechanisms. Preliminary transition-state searches on the UMA potential energy surface will provide starting geometries for DFT refinement. Candidate transition states will subsequently be characterised by vibrational analysis and intrinsic reaction-coordinate calculations to establish their connection to the intended minima. Representative pathways will be compared with direct DFT searches to assess the reliability of the approach and the reduction in computational cost. The lower cost of UMA will allow us to investigate multiple competing pathways systematically. Ground-state molecular dynamics will provide additional sampling of accessible geometries and help identify conformational changes associated with thermal isomerisation. We will use the resulting mechanisms and DFT-validated barriers to understand how structural modifications influence metastable-state lifetimes, guiding the design of photoswitches with thermal stability suited to their intended applications. A complementary objective is to improve photoswitch performance by increasing the separation between the UV–Vis absorption bands of the stable and metastable isomers, favouring higher populations of the desired form at the photostationary state. UMA conformer searches will identify low-energy structures for prediction of absorption spectra. This workflow will be combined with a genetic algorithm to explore substituent combinations and architectural modifications with large chemical spaces. Together, these approaches will enable broader computational exploration of photoswitch structure and reactivity, concentrating expensive DFT calculations on the most informative geometries and providing experimentally testable predictions for synthesis and characterisation.