Machine Learning Accelerated Global Search for Adsorption Geometries of Merocyanine Molecule on Hexagonal Boron Nitride

The adsorption of the merocyanine dye HB238 on hexagonal boron nitride (hBN) was investigated using a machine learning (ML) assisted global search strategy. A series of MACE machine learning interatomic potentials with higher-order equivariant message passing were finetuned on density functional theory (DFT) reference data for single and dimer adsorbate configurations, providing accurate surrogate models for the potential energy surface. The Bayesian Optimisation Structure Search (BOSS) was used to search over translational and rotational degrees of freedom of the adsorbed molecules, followed by full PBE/D3 optimisation of the most promising structures. The ML-accelerated search revealed that HB238 prefers to adsorb in face-on orientation on hBN surface with the sulfur atoms located near hollow sites; however, the molecule exhibits no strong site selectivity, giving rise to a broad ensemble of configurations within energies 0.1 eV above the global minimum. When two HB238 molecules are adsorbed, they align parallel to each other and lie flat on the surface. Overall, our results demonstrate that combining finetuned ML potentials with Bayesian optimisation enables an efficient and accurate exploration of complex adsorption landscapes and provides fundamental insights into the physisorption of dipolar dyes on 2D insulators. This combined <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"> <mml:semantics><mml:mrow><mml:mtext>MACE</mml:mtext> <mml:mo>×</mml:mo> <mml:mtext>BOSS</mml:mtext></mml:mrow> <mml:annotation>$$ \mathrm{MACE}\times \mathrm{BOSS} $$</mml:annotation></mml:semantics> </mml:math> approach can be easily extended to investigate organic molecular aggregates on 2D surfaces.