Date

Thursday 25 Jun 2026

Time

15:15 - 15:45

Location

BW017

Supervisor

Francesco Buda

Jury

Anthe Janssen

Machine learning interatomic potentials (MLIPs) have emerged as a powerful and rapidly evolving approach in materials modelling, offering a data-driven route to bridge the long-standing gap between quantum mechanical accuracy and the efficiency required for large-scale atomistic simulations. By learning potential energy surfaces from high-fidelity electronic structure calculations, MLIPs enable near-first-principles accuracy at a fraction of the computational cost, significantly extending the ac-cessible length and time scales compared to density functional theory (DFT), while overcoming many of the limitations associated with classical empirical force fields. The development of MLIPs is closely tied to advances in machine learning architectures and the increasing availability of high-quality train-ing data. These methods are capable of representing complex atomic environments in crystalline, dis-ordered, and reactive systems, making them broadly applicable across materials science, chemistry, and condensed matter physics. Their success has been driven by flexible representations of atomic structure, ranging from carefully designed descriptors to modern equivariant neural network archi-tectures that embed fundamental physical symmetries. Despite their capabilities, MLIPs remain con-strained by their dependence on training data and their limited ability to reliably extrapolate beyond sampled configurations. Ongoing research is therefore focused on improving robustness and general-ization through active learning strategies, uncertainty quantification, and the incorporation of physical inductive biases such as symmetry constraints and long-range interactions. Overall, MLIPs represent a significant shift in computational materials modelling, enabling accurate and scalable simulations of complex systems that were previously intractable. Their continued development is expected to further integrate atomistic modelling with multiscale and data-driven discovery frameworks.