Bimetallic nanoparticles combine two metals to access a structural and electronic landscape unattainable by either component alone. The Bismuth–Platinum (Bi–Pt) system illustrates this versatility: it can form disordered alloys, core–shell architectures, Janus morphologies, or ordered intermetallic phases, with applications ranging from catalysis to temperature measurement in nuclear reactors, and extending to nanomedicine, where ligand-functionalised Bi–Pt nanoparticles are attracting growing interest. As this configurational space, further multiplied by the presence of surface ligands, is far too vast for exhaustive DFT sampling, this thesis develops machine-learning interatomic potentials (ML-IAPs) to explore it, using DFT (PBE functional with corrected dispersion effects) as the reference method. Since the quality of an ML-IAP is dictated by that of its training data, particular attention is devoted to dataset construction, curation, and selection throughout this work.

A first potential, ChIMES, is built by active learning, the training set being iteratively enriched with the configurations where the potential is least reliable, and coupled with particle swarm optimization to explore the energy surface of the bare bimetallic core. This exploration reveals 34 genuine Bi20Pt20​ minima, whose classification by unsupervised learning (principal component analysis and clustering on structural descriptors) identifies the dominant morphological families and confirms the surface segregation of bismuth, driven by charge transfer from Bi to Pt. It culminates in the discovery of a Bi18Pt24​ nanocage absent from any training set, offering a unique system to determine the potential's ability to extrapolate beyond its training data.

Extending to a ligand-functionalised system representative of these applications, the MACE MH-1 foundation model is evaluated as-is on the free ligand and the PEG–bimetallic interface, where it fails to faithfully reproduce these structures. Fine-tuning MH-1 on the Bi–Pt system then requires building a dataset that is both compact and representative: a subset of 3000 configurations is extracted by an unsupervised, diversity-weighted selection that maximizes coverage of configurational space rather than sheer data volume. The potential trained on this curated set reproduces the DFT-relaxed structures of the ligand and interface far more faithfully than its own unrefined starting point, delivering a validated production potential for the ligand–bimetallic interaction.